Styaz P M A I A S V 1 P M T
Styaz P M A I A S V 1 P M T
Predictive modeling involves the use of data to forecast future events. It relies on
capturing relationships between explanatory variables and the predicted variables
from past occurrences and exploiting these relationships to predict future outcomes.
Forecasting future financial events is a core actuarial skill actuaries routinely apply
predictive modeling techniques in insurance and other risk management applications.
This book is for actuaries and other financial analysts who are developing their
expertise in statistics and wish to become familiar with concrete examples of predictive
modeling. The book also addresses the needs of more seasoned practicing analysts
who would like an overview of advanced statistical topics that are particularly relevant
in actuarial practice.
Predictive Modeling Applications in Actuarial Science emphasizes life-long learning by developing tools in an insurance context, providing the relevant actuarial
applications, and introducing advanced statistical techniques that can be used by
analysts to gain a competitive advantage in situations with complex data.
Edward W. Frees is the Hickman-Larson Professor of Actuarial Science at the
Wisconsin School of Business, University of Wisconsin-Madison.
Richard A. Derrig is the president of Opal Consulting LLC and a visiting professor
of Risk, Insurance, and Healthcare Management at Fox School of Business, Temple
University.
Glenn Meyers has recently retired as vice president and chief actuary at ISO Innovative
Analytics.
INT E R NAT I O NA L S E R I E S O N AC T UA R I A L S C I E N C E
Editorial Board
Christopher Daykin (Independent Consultant and Actuary)
Angus Macdonald (Heriot-Watt University)
The International Series on Actuarial Science, published by Cambridge University Press
in conjunction with the Institute and Faculty of Actuaries, contains textbooks for students
taking courses in or related to actuarial science, as well as more advanced works designed for
continuing professional development or for describing and synthesizing research. The series
is a vehicle for publishing books that reflect changes and developments in the curriculum, that
encourage the introduction of courses on actuarial science in universities, and that show how
actuarial science can be used in all areas where there is long-term financial risk.
A complete list of books in the series can be found at www.cambridge.org/statistics. Recent
titles include the following:
Computation and Modelling in Insurance and Finance
Erik Blviken
Solutions Manual for Actuarial Mathematics for Life Contingent Risks (2nd Edition)
David C.M. Dickson, Mary R. Hardy, & Howard R. Waters
Actuarial Mathematics for Life Contingent Risks (2nd Edition)
David C.M. Dickson, Mary R. Hardy, & Howard R. Waters
Risk Modelling in General Insurance
Roger J. Gray & Susan M. Pitts
Financial Enterprise Risk Management
Paul Sweeting
Regression Modeling with Actuarial and Financial Applications
Edward W. Frees
Nonlife Actuarial Models
Yiu-Kuen Tse
Generalized Linear Models for Insurance Data
Piet De Jong & Gillian Z. Heller
PREDICTIVE MODELING
A P P L I C AT I O N S I N AC T UA R I A L
SCIENCE
Volume I: Predictive Modeling Techniques
Edited by
E DWA R D W. F R E E S
University of Wisconsin, Madison
RICHARD A. DERRIG
Opal Consulting LLC, Providence, Rhode Island
GLENN MEYERS
ISO Innovative Analytics, Jersey City, New Jersey
Contents
Contributor List
Acknowledgments
1
page xiii
xix
1
1.1 Introduction
1.2 Predictive Modeling and Insurance Company Operations
1.3 A Short History of Predictive Modeling in Actuarial Science
1.4 Goals of the Series
References
1
3
5
8
9
13
2.1 Introduction
2.2 Linear Model Theory with Examples
2.3 Case Study
2.4 Conclusion
2.5 Exercises
References
13
15
45
59
60
63
65
Montserrat Guillen
3.1
3.2
3.3
3.4
65
66
67
78
vii
viii
Contents
79
81
85
86
87
Jean-Philippe Boucher
4.1 Introduction
4.2 Poisson Distribution
4.3 Poisson Regression
4.4 Heterogeneity in the Distribution
4.5 Zero-Inflated Distribution
4.6 Conclusion
4.7 Further Reading
References
87
87
89
92
102
105
105
105
107
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
5.9
107
110
115
118
121
122
129
133
138
133
135
136
Edward W. Frees
6.1
6.2
6.3
6.4
6.5
138
140
148
152
160
Contents
ix
161
162
164
167
Edward W. Frees
7.1 Introduction
7.2 Linear Models
7.3 Nonlinear Models
7.4 Additional Considerations
7.5 Further Reading
References
167
172
176
180
181
181
182
182
192
201
213
215
217
9.1 Introduction
9.2 Credibility and the LMM Framework
9.3 Numerical Examples
9.4 Theory versus Practice
9.5 Further Reading
9.6 Appendix
References
10 Fat-Tailed Regression Models
217
220
224
227
232
233
234
236
Peng Shi
10.1 Introduction
10.2 Transformation
10.3 GLM
236
238
241
Contents
243
250
255
256
258
260
11.1 Introduction
11.2 Exploratory Analysis of Spatial Data
11.3 Spatial Autoregression
11.4 Average Claim Size Modeling
11.5 Hierarchical Model for Total Loss
11.6 Discussion and Conclusion
References
12 Unsupervised Learning
260
262
265
269
273
278
278
280
Louise Francis
12.1 Introduction
12.2 Datasets
12.3 Factor and Principal Components Analysis
12.4 Cluster Analysis
12.5 Exercises
References
280
283
285
294
309
310
315
Brian Hartman
315
316
316
319
326
330
330
331
334
14.1 Introduction
334
Contents
xi
335
338
354
358
362
365
367
367
370
388
395
396
398
16.1
16.2
16.3
16.4
16.5
16.6
Introduction
Model Families for Multilevel Non-Gaussian Data
Generalized Linear Mixed Models
Nonlinear Mixed Models
Bayesian Approach to (L,GL,NL)MMs
Example: Poisson Regression for Workers
Compensation Insurance Frequencies
References
IV
398
399
399
411
411
412
423
Longitudinal Modeling
427
Piet de Jong
427
431
434
443
448
448
449
Greg Taylor
449
453
xii
Contents
455
467
479
481
Jim Robinson
481
482
482
486
497
507
511
511
513
514
515
20.1
20.2
20.3
20.4
516
518
522
Index
539
534
536
Contributor List
xiv
Contributor List
Shuo-Li Chuang received her PhD from the Department of Information, Risk, and
Operations Managements at the Red McCombs School of Business, The University
of Texas at Austin. Her research area is mortality modeling and application of statistical modeling. She was a doctoral student in the school of statistics, University of
Minnesota at Twin Cities, and was involved in various academic projects applying
statistical modeling in economics, finance, education, and insurance.
Claudia Czado is an associate professor of Applied Mathematical Statistics at the
Zentrum Mathematik of the Technische Universitat in Munich. Her interests are in
regression modeling with space and time effects, as well as general dependence modeling using vine copulas. She has published about 100 research articles on statistical
models, methods, and computation, which have been applied to topics in insurance,
finance, medicine, and biology.
Enrique de Alba is professor emeritus at Instituto Tecnologico Autonomo de Mexico
(ITAM), where he was dean of the Division of Actuarial Science, Statistics and
Mathematics for more than 20 years. He is also an adjunct professor at the University
of Waterloo in Canada. He is currently co-editor of the North American Actuarial
Journal and vice president of the National Statistical Institute of Mexico (INEGI).
Curtis Gary Dean is the Lincoln Financial Distinguished Professor of Actuarial
Science at Ball State University. He is a Fellow of the Casualty Actuarial Society
and member of the American Academy of Actuaries. He worked for many years at
American States Insurance and later at SAFECO and Travelers.
Piet de Jong is Professor of Actuarial Studies at Macquarie University in Sydney,
Australia. He has held a variety of academic positions throughout the world and has
been involved in many quantitative consulting projects. His research interests focus
on quantitative tools and techniques, particularly in relation to issues in actuarial
science, finance, and economics, including time series analysis and generalized linear
modeling.
Richard A. Derrig is president of OPAL Consulting LLC, established in February
2004 to provide research and regulatory support to the property-casualty insurance
industry. Before forming OPAL, Dr. Derrig held various positions at the Automobile
Insurers Bureau (AIB) of Massachusetts and at the Insurance Fraud Bureau (IFB) of
Massachusetts over a 27-year period, retiring in January 2004 as senior vice president
at AIB and vice president of research at IFB.
Harald Dornheim, ASA, CERA, is an actuarial and risk management consultant at
KPMG in Zurich. He completed his PhD in mathematical statistics at the University
of Wisconsin-Milwaukee on the topic of robust efficient fitting of mixed linear models
Contributor List
xv
for prediction and risk pricing in insurance. He is a Fellow of the German Actuarial
Association and of the Swiss Actuarial Association.
Louise Francis, FCAS, MAAA, is the consulting principal and founder of Francis
Analytics and Actuarial Data Mining, Inc., where she leads reserving, pricing, predictive modeling, simulation, and related actuarial projects and engagements. She
is a former Vice President of Research for the Casualty Actuarial Society and has
been involved in a number of CAS initiatives, including estimating reserve variability,
improving data quality, and reviewing papers for its journal Variance. She presents
frequently on data-mining-related topics and is a five-time winner of the CASs Data
Management and Information call paper program.
Edward W. (Jed) Frees is the Hickman-Larson Professor of Actuarial Science at the
University of Wisconsin-Madison. He is a Fellow of both the Society of Actuaries
and the American Statistical Association. He has published extensively (a four-time
winner of the Halmstad Prize for best paper published in the actuarial literature) and
has written three books; his most recent is Regression Modeling with Actuarial and
Financial Applications (Cambridge University Press, 2009).
Montserrat Guillen is a full professor of econometrics at the University of Barcelona
and director of the research group Riskcenter. She is member of the Reial Acad`emia
de Doctors, and in 2011 she served as president of the European Group of Risk and
Insurance Economists. Her books include Quantitative Models for Operational Risk
(Chapman and Hall, 2012).
James Guszcza is the U.S. Predictive Analytics Lead for Deloitte Consultings Actuarial, Risk, and Advanced Analytics practice. He is a Fellow of the Casualty Actuarial
Society and a past faculty member in the Department of Actuarial Science, Risk
Management, and Insurance at the University of Wisconsin-Madison.
Brian Hartman is an assistant professor of Actuarial Science at the University of
Connecticut and an associate of the Society of Actuaries. He has worked as an
actuary for Mercer (retirement consulting) and The Hartford (small commercial auto
insurance). His current research interests include Bayesian methods, regime-switching
models, predictive modeling, and risk management.
Bruce Jones is a professor in the Department of Statistical and Actuarial Sciences at
the University of Western Ontario, where he has been a faculty member since 1996.
His research addresses a variety of questions related to modeling for life and health
insurance. He is a Fellow of the Society of Actuaries and of the Canadian Institute of
Actuaries.
xvi
Contributor List
Glenn Meyers, FCAS, MAAA, CERA, PhD, recently retired after a 37-year actuarial career that spanned both industry and academic employment. For his last 23
years of working he was employed by ISO as a research actuary. He has received
numerous awards for his publications from the Casualty Actuarial Society, including
being the first recipient of the Michaelbacher Significant Achievement Award, which
recognizes a person or persons who have significantly and fundamentally advanced
casualty actuarial science.
Luis E. Nieto-Barajas is a full-time professor at Instituto Tecnologico Autonomo
de Mexico. He has published numerous research articles on Bayesian and Baeysian
nonparametric statistics, with applications ranging from survival analysis to claims
reserving. He has served as consultant for the federal presidential elections in Mexico.
Utai Pitaktong is the Senior Modeling Analyst for Actuarial Department at USAA.
His former research positions were in the Actuarial Department of State Farm Insurance and in the Predictive Modeling Research Unit of FICO. One of his publications
that he co-authored won the 2004 Robert Mehr Award from the American Risk
and Insurance Association. He received his PhD in Management Science from the
McCombs School of Business, The University of Texas at Austin, and his M.S. in
Applied Mathematical Science from Rice University.
Ponmalar Ratnam is a PhD candidate in the Department of Mathematical Sciences
at the University of Wisconsin-Milwaukee working on the topic of robust statistical
methods for claim frequency models. She has more than 10 years of actuarial modeling
experience in the insurance industry and is currently employed as a data scientist at
Systech in Los Angeles.
Jim Robinson, PhD, FSA, is the Director of the Center for Health Systems Research
& Analysis at the University of Wisconsin-Madison.
Marjorie Rosenberg is a professor of Actuarial Science, Risk Management, and
Insurance in the School of Business at the University of Wisconsin-Madison with a
joint appointment in Biostatistics and Medical Informatics in the School of Medicine
and Public Health. Her research interests are in the application of statistical methods
to health care and applying her actuarial expertise to cost and policy issues in health
care.
Peng Shi is an assistant professor of Actuarial Science, Risk Management, and
Insurance at the University of Wisconsin-Madison. He is an associate of the Society
of Actuaries. His research interests are predictive modeling, multivariate regression
and dependence models, longitudinal data, and asymmetric information in insurance.
Contributor List
xvii
Greg Taylor holds an honorary professorial position in Risk and Actuarial Studies
at the University of New South Wales. He previously spent 44 years in commercial
actuarial practice and 8 years as an actuarial academic. Taylor has published two
books on loss reserving and numerous articles in mathematics, statistics, and actuarial
science. He is an Officer of the Order of Australia and holds a Gold Medal from the
Australian Actuaries Institute and a Silver Medal from the United Kingdom Institute
and Faculty of Actuaries.
Weijia Wu
Yanwei (Wayne) Zhang is a PhD student of marketing at the University of Southern
California. He is a Fellow of the Casualty Actuarial Society. He has published in
statistical and actuarial journals, and his current research is on social media and
consumer decision making.
Acknowledgments
Many, many people contributed to this project. Our editorial and author team represents academia and industry, as well as seven different countries, reflecting the broad
interest in predictive modeling. The editors would like to especially commend Xiaoli
Jin, who served as project manager for this book while in the doctoral program at the
University of Wisconsin-Madison. Her efforts helped improve this book immensely.
We were also fortunate to receive extensive comments from many chapter reviewers.
We thank them for their contributions.
Funding for this project was provided by the Casualty Actuarial Society and the
Canadian Institute of Actuaries.
Reviewer Acknowledgment
Daniel Alai, University of New South Wales
John Baldan, ISO
Lee Bowron, Kerper and Bowron LLC
Doug Bujakowski, University of Wisconsin-Madison
Alan Chalk, AIG
Wai-Sum Chan, Chinese University of Hong Kong
Arthur Charpentier, University of Rennes
Dave Clark, Munich Reinsurance America, Inc.
Steven Craighead, Pacific Life
Marc-Andre Desrosiers, Intact Insurance Company
Mario DiCaro, Ultimate Risk Solutions
Robert Erhardt, Wake Forest University
Paul Ferrara, Homesite Group Incorporated
Luyang Fu, Cincinnati Financial Corporation
Wu-Chyuan Gau, Florida Blue
Mordechai Goldburd, ISO
xix
xx
Acknowledgments
1
Predictive Modeling in Actuarial Science
Edward W. Frees, Richard A. Derrig, and Glenn Meyers
Chapter Preview. Predictive modeling involves the use of data to forecast future events.
It relies on capturing relationships between explanatory variables and the predicted
variables from past occurrences and exploiting them to predict future outcomes. The
goal of this two-volume set is to build on the training of actuaries by developing
the fundamentals of predictive modeling and providing corresponding applications in
actuarial science, risk management, and insurance. This introduction sets the stage
for these volumes by describing the conditions that led to the need for predictive
modeling in the insurance industry. It then traces the evolution of predictive modeling
that led to the current statistical methodologies that prevail in actuarial science today.
1.1 Introduction
A classic definition of an actuary is one who determines the current financial impact of
future contingent events.1 Actuaries are typically employed by insurance companies
who job is to spread the cost of risk of these future contingent events.
The day-to-day work of an actuary has evolved over time. Initially, the work
involved tabulating outcomes for like events and calculating the average outcome.
For example, an actuary might be called on to estimate the cost of providing a death
benefit to each member of a group of 45-year-old men. As a second example, an actuary
might be called on to estimate the cost of damages that arise from an automobile
accident for a 45-year-old driver living in Chicago. This works well as long as there
are large enough groups to make reliable estimates of the average outcomes.
Insurance is a business where companies bid for contracts that provide future
benefits in return for money (i.e., premiums) now. The viability of an insurance
company depends on its ability to accurately estimate the cost of the future benefits it
promises to provide. At first glance, one might think that it is necessary to obtain data
1
from sufficiently large groups of like individuals. But when one begins the effort
of obtaining a sufficient volume of like data, one is almost immediately tempted to
consider using similar data. For example, in life insurance one might want to use
data from a select group of men with ages between 40 and 50 to estimate the cost of
death benefits promised to a 45-year-old man. Or better yet, one may want to use the
data from all the men in that age group to estimate the cost of death benefits for all
the men in that group. In the automobile insurance example, one may want to use the
combined experience of all adult men living in Chicago and Evanston (a suburb north
of Chicago) to estimate the cost of damages for each man living in either city arising
from an automobile accident.
Making use of similar data as opposed to like data raises a number of issues.
For example, one expects the future lifetime of a 25-year-old male to be longer that
of a 45-year-old male, and an estimate of future lifetime should take this difference
into account. In the case of automobile insurance, there is no a priori reason to expect
the damage from accidents to a person living in Chicago to be larger (or smaller) than
the damage to a person living in Evanston. However, the driving environment and the
need to drive are quite different in the two cities and it would not be prudent to make
an estimate assuming that the expected damage is the same in each city.
The process of estimating insurance costs is now called predictive modeling.2
In a very real sense, actuaries had been doing predictive modeling long before the
term became popular. It is interesting to see how the need for predictive modeling has
evolved in the United States.
In 1869, the U.S. Supreme Court ruled in Paul v. Virginia that issuing a policy
of insurance is not a transaction of commerce. This case had the effect of granting
antitrust immunity to the business of insurance. As a result, insurance rates were
controlled by cartels whose insurance rates were subject to regulation by the individual
states. To support this regulation, insurers were required to report detailed policy and
claim data to the regulators according to standards set by an approved statistical plan.
The Supreme Court changed the regulatory environment in 1944. In United States
v. Southeast Underwriters Association it ruled that federal antitrust law did apply
under the authority of the Commerce Clause in the U.S. Constitution. But by this
time, the states had a long-established tradition of regulating insurance companies.
So in response, the U.S. Congress passed the McCarran-Ferguson Act in 1945 that
grants insurance companies exemption from federal antitrust laws so long as they
are regulated by the states. However, the federal antitrust laws did apply in cases of
boycott, coercion, and intimidation.
The effect of the McCarran-Ferguson Act was to eliminate the cartels and free the
insurers to file competitive rates. However, state regulators still required the insurers
2
The name predictive modeling is also used in many enterprises other than insurance.
to compile and report detailed policy and claim data according to approved statistical
plans. Industry compilations of these data were available, and insurance companies
were able to use the same systems to organize their own data.
Under the cartels, there was no economic incentive for a refined risk classification
plan, so there were very few risk classifications. Over the next few decades, insurance
companies competed by using these data to identify the more profitable classes of
insurance. As time passed, predictive modeling led to more refined class plans.
As insurers began refining their class plans, computers were entering the insurance
company workplace. In the 60 s and 70 s, mainframe computers would generate thick
reports from which actuaries would copy numbers onto a worksheet, and using at
first mechanical and then electronic calculators, they would calculate insurance rates.
By the late 70 s some actuaries were given access to mainframe computers and the
use of statistical software packages such as SAS. By the late 80 s many actuaries had
personal computers with spreadsheet software on their desks.
As computers were introduced into the actuarial work environment, a variety of
data sources also became available. These data included credit reports, econometric
time series, geographic information systems, and census data. Combining these data
with the detailed statistical plan data enabled many insurers to continue refining their
class plans. The refining process continues to this day.
40
400
60
450
500
80
550
100
350
1050
Count
Top 10
300
20
50+
1990
1993
1996
1999
2002
2005
2008
Fig. 1.1. Growth in market share by large insurers. The bar chart shows the percentage, on the left-hand vertical axis, of personal automobile premiums written by the
top 10 insurers (in terms of premium volume), the next 40, and other insurers. The
right-hand vertical axis shows the decreasing number of insurer groups over time, as
indicated by the dashed line. Source: ISO.
classification systems and skimming the cream underwriting strategies. Figure 1.1
shows how the top American insurers have increased their market share over time.
It was this growth in market share that fueled the growth in predictive modeling in
insurance.
Actuaries learn and develop modeling tools to solve actuarial problems. With
these tools, actuaries are well equipped to make contributions to broader company
areas and initiatives. This broader scope, sometimes known as business analytics,
can include the following areas:
r Sales and Marketing these departments have long used analytics to predict customer
behavior and needs, anticipate customer reactions to promotions, and to reduce acquisition costs (direct mail, discount programs).
r Compensation Analysis predictive modeling tools can be used to incentivize and
reward appropriate employee/agent behavior.
r Productivity Analysis more general than the analysis of compensation, analytic tools
can be used to understand production by employees and other units of business, as well
as to seek to optimize that production.
r Financial Forecasting analytic tools have been traditionally used for predicting financial results of firms.
Predictive modeling in the insurance industry is not an exercise that a small group of
actuaries can do by themselves. It requires an insurer to make significant investments
in their information technology, marketing, underwriting, and actuarial functions.
Figure 1.2 gives a schematic representation of the investments an insurer has to make.
RETURN
Statistical Analysis
Predictive Modeling
INVESTMENT
Search for new data,
method, and new lift
Multidimensional customer
view, including profitability
Customized Third-Party
Scores
Linear Model-Based
Ratemaking
Disruptive innovation
Knowledge of drivers
in value creation
Mr. Bailey is the son of Arthur L. Bailey, who introduced Bayes theorem to the actuarial theory of credibility.
(4) It should produce a rate for each subgroup of risk which is close enough to the
experience so that the differences could reasonably be caused by chance.
Bailey and LeRoy then considered a number of models to estimate the relative
loss ratios, rij , for use class i and merit rating class j . The two most common
models they considered were the additive model rij = i + j and the multiplicative
model rij = i j . They then described some iterative numerical methods to estimate
the coefficients {i } and {j }. The method that most insurers eventually used was
described a short time later by Bailey (1963) in his paper titled Insurance Rates with
Minimum Bias. Here is his solution for the additive model:
(1) Calculate the initial estimates,
j =
nij rij
for each j . For the additive model, Bailey used the weights, nij , to represent the
exposure measured in car years.
(2) Given the balanced criterion (#1 above) for each i, we have that
nij (rij i j ) = 0.
j
This iterative method is generally referred to as the Bailey minimum bias additive
model. There is a similarly derived Bailey minimum bias multiplicative model. These
models can easily be generalized to more than two dimensions.
As mentioned in the description of the iterative method, the balance criterion
is met by the design of the method. The method meets the credibility criterion
because its choice of weights are equal to the exposure. To satisfy the minimum
departure or goodness of fit (in current language) criterion, Bailey and Simon
recommended testing multiple models to see which one fit the best. They proposed
using the chi-square statistic to test if the differences between actual and fitted were
due to chance. Although their technology and language were different from now,
they were definitely thinking like current actuarial modelers.
The Bailey minimum bias models were easy to program in FORTRAN and BASIC,
the predominant computer programming languages of the sixties and seventies, and
they were quickly adopted by many insurers. We did a quick poll at a recent Casualty
Actuarial Society seminar and found that some insurers are still using these models.
Meanwhile, statisticians in other fields were tackling similar problems and developing general statistical packages. SPSS released its first statistical package in 1968.
Minitab was first released in 1972. SAS released the first version of its statistical
package in 1976. The early versions of these packages included functions that could
fit multivariate models by the method of least squares.
Lets apply the least squares method to the previous example. Define
nij (rij i j )2 .
B=
i,j
Brown attributes his introduction to GLMs to Ben Zehnwirth and Piet DeJong while on a sabbatical leave in
Australia.
Zehnwirth (1994), and then a host of others. The connection between GLMs and Bailey minimum bias models reached its best expression in Mildenhall (1999). For any
given GLM model there is a corresponding set of weights {wij }, (given by Mildenhalls
Equation 7.3) for which
wij (rij ij ) = 0 for all i and
wij (rij ij ) = 0 for all j.
i
As an example one can set ij = i + j and there is a GLM model for which the
corresponding set of weights {wi,j } = {nij } yields the Bailey additive model. If we set
ij = i j there is a GLM model for which the corresponding set of weights {wij }
yields the Bailey multiplicative model. Other GLM models yield other minimum bias
models.
The strong connection between the Bailey minimum bias models and the GLM
models has helped make the actuarial community comfortable with GLMs. The actuarial community is rapidly discovering the practical advantages of using statistical
packages such as SAS and R. These advantages include the following:
r the ability to include continuous variables (such as a credit score) in the rating plan
r the ability to include interactions among the variables in the rating plan
r the diagnostic tools that allow one to evaluate and compare various models
As a result, the leading edge of actuarial scientists are now using statistical packages
for their predictive modeling tasks.
As described in Section 1.2, these tasks go beyond the classification ratemaking
applications discussed earlier. Loss reserving is another important task typically performed by actuaries. An early reference to loss reserving that reflects the statistical
way of thinking is found in Taylor (1986). Since then there have been several papers
describing statistical models in loss reserving. We feel that the papers by Mack (1994),
Barnett and Zehnwirth (2000), and England and Verrall (2002) represent the different
threads on this topic. These papers have led to a number of specialized loss reserving
statistical packages. Although some packages are proprietary, the R chainladder
package is freely available.
References
10
Part I
Predictive Modeling Foundations
2
Overview of Linear Models
Marjorie Rosenberg and James Guszcza
Chapter Preview. Linear modeling, also known as regression analysis, is a core tool
in statistical practice for data analysis, prediction, and decision support. Applied data
analysis requires judgment, domain knowledge, and the ability to analyze data. This
chapter provides a summary of the linear model and discusses model assumptions,
parameter estimation, variable selection, and model validation around a series of
examples. These examples are grounded in data to help relate the theory to practice.
All of these practical examples and exercises are completed using the open-source R
statistical computing package. Particular attention is paid to the role of exploratory
data analysis in the iterative process of criticizing, improving, and validating models
in a detailed case study. Linear models provide a foundation for many of the more
advanced statistical and machine-learning techniques that are explored in the later
chapters of this volume.
2.1 Introduction
Linear models are used to analyze relationships among various pieces of information
to arrive at insights or to make predictions. These models are referred to by many
terms, including linear regression, regression, multiple regression, and ordinary least
squares. In this chapter we adopt the term linear model.
Linear models provide a vehicle for quantifying relationships between an outcome
(also referred to as dependent or target) variable and one or more explanatory (also
referred to as independent or predictive) variables. Tradition perhaps accounts for
some of the widespread use of linear models, because they are among the oldest
data analysis techniques. As discussed in Stigler (1986), Carl Friedrich Gauss and
Pierre Simon Laplace developed the method of least squares for estimating planetary
orbits from observational data where the functional form of the relationship was
provided by physical theory. A second major conceptual advance occurred when the
13
14
Source code and data for the examples and exercises are available on the books website.
15
The calculation of BMI = 703 times weight in pounds divided by height in squared inches.
16
Example 2.1 (Basic Average Body Mass Index). We analyze a sample of data for
1,101 persons. BMI is the outcome variable of interest and is denoted by (BMI i , i =
1, . . . , 1,101). Our statistical model (with n = 1,101) is
BMI i = + i
i = 1, . . . n.
(2.1)
The BMI population average equals , an unknown quantity. Each individual error is
represented by i , where E [i ] = 0 and Var [i ] = y2 . The subscript y on y2 refers
to the outcome variable, in this case BMI, and is further discussed in Section 2.2.2.
Essentially a persons BMI is predicted by the mean in this model. If no other information is available, the best predictor of BMI, in the sense of minimizing squared errors,
is the sample mean BMI = n1 ni=1 BMI i . The residual, ei , is an estimate of i and
is equal to ei = BMI i BMI. In our sample of n = 1,101 observations, the sample
average BMI is BMI = 31.97, and the sample variance is equal to 55.42. Thus we use
31.97 as our prediction of a persons BMI and 55.42 as our estimate of y2 .
Example 2.2 (One-Variable Regression for Body Mass Index). We can improve
the basic model by incorporating one or more explanatory (or independent) variables to predict BMI. Suppose we construct a linear model that uses AGE to predict
BMI. Here the statistical model can be represented as
BMI i = 0 + 1 AGEi + i
i = 1, . . . , n.
(2.2)
Comparing this equation with Equation 2.1, the outcome variable is BMI, i is the error
term, and we have modified to vary by person (i = 0 + 1 AGEi ). The personlevel mean is a linear function of AGE, where 0 is the intercept and 1 is the slope.
Note in Example 2.1 that the slope is equal to zero with an intercept equal to 31.97.
The empirical relationship between BMI and AGE is suggested by the scatterplot in
Figure 2.1. In this plot, vertical and horizontal lines have been placed at the sample
average values of AGE and BMI, respectively.
The estimated linear model relating BMI to AGE contains an intercept term equal to
36.98 and a slope term on AGE equal to 0.085. The solid line in Figure 2.1 indicates
that as AGE increases by 1, then BMI decreases by 0.085. As a check, the correlation
between BMI and AGE, which measures the linear association between these two
variables, is 0.15. Note that the sign of the slope coefficient agrees with that of the
correlation statistic. We see in Example 2.5 how these statistics are related.
17
70
60
BMI
50
40
30
20
10
20
30
40
50
60
70
80
AGE
Fig. 2.1. Scatterplot of BMI vs. AGE. Dotted lines represent sample mean of BMI and
AGE. Solid line is the least squares line.
takes on a countable, possibly infinite, number of distinct values. Our variable AGE in
Example 2.2 could be considered either continuous or discrete.
In contrast, a categorical variable has a distinct number of levels that can be either
ordered or unordered. The estimation of the regression coefficients is not affected
by whether a categorical variable is viewed as ordered or unordered. However, the
regression coefficients are expected to maintain a certain relationship for ordered
categorical variables.
A binary (also referred to as dummy or indicator) variable is a special case of
a categorical variable, denoting group membership where the levels are 1 if the
observation is a member of the group and 0 otherwise.
Example 2.3 (Linear Model with Sex). Suppose we are interested in the relationship between body mass index (BMI) and sex for individuals who have been diagnosed
with diabetes. Using the MEPS data, we regress BMI on the two-level categorical variable SEX. The question of interest is whether ones sex has an effect on the magnitude
of BMI. Table 2.1 summarizes the results of three alternate, but equivalent, models
using panel 13 of the MEPS data.
The first model, displayed in Table 2.1(a), transforms the categorical variable for sex into a binary indicator variable for males. The estimating equation
= 32.7923 1.9135 MALE, with the estimated BMI for males equal to
is BMI
30.8788 (= 32.7923 1.9135(1)), whereas the estimated BMI for females is 32.7923
(= 32.7923 1.9135(0)). In this particular model, the reference category is females;
that is, the results for males are with reference to the base category of female. We see
that the BMI for males is lower than that for females by 1.9135.
18
32.7923
1.9135
30.8788
1.9135
30.8788
32.7923
BMI
70
60
50
40
30
20
10
Female
BMI
19
Male
Other
Black
70
60
50
40
30
20
10
F.Other
M.Other
F.Black
M.Black
Fig. 2.2. Boxplots of BMI vs. SEX; BMI vs. RACE; and BMI vs. SEX/RACE combination. Solid lines in the boxplots represent the median BMI by category.
Example 2.4 (Multiple Variable Regression for Body Mass Index). We extend
Examples 2.2 and 2.3 to include the binary variables MALE and BLACK, along with
the continuous variable AGE:
BMI i = 0 + 1 AGEi + 2 MALEi + 3 BLACK i + i
i = 1, . . . , n
(2.3)
The marginal relationships between BMI and male (vs. females) and BMI and BLACK
(vs. OTHER RACES) are shown in Figure 2.2.
We see that the median BMI for males is slightly lower than that for females,
whereas the median BMI for BLACKS is higher than that for OTHER RACES. When
looking at the distribution across the four distinct categories, the FEMALE BLACK
median BMI is the highest and the interquartile range is the widest.
The parameter 1 for AGE is now estimated as 0.083, similar to the estimate found
in Example 2.2. The interpretation of the coefficient is that the estimated BMI for an
individual in the population of interest for a given sex and race changes by 0.083
for each additional age.
The estimate for 2 is 1.60, indicating that a male of a particular age and race is
expected to have a BMI approximately 1.60 lower than a female of the same age and
race. Finally, the estimate for 3 is 2.46, indicating that being black increases ones
BMI for a given age and sex. Thus for a given age, a black female has the highest
estimated BMI, a conclusion that concurs with the boxplot.
The estimate of the intercept term 0 is 36.95. In general, the intercept of a linear
model is the predicted value of the outcome variable given a value of zero for each of
the models explanatory variables. In the current example, the intercept term represents
20
the BMI for a (diabetic) non-black, female, newborn baby (BLACK = 0, MALE = 0,
AGE = 0). Because the minimum age in our data is 18, the interpretation of the
intercept term is not useful. Techniques such as centering the data, as discussed in
Section 2.3.3, adjust the use of the variables and enable a useful interpretation of the
intercept term.
Part of the art of statistics involves deciding how to incorporate explanatory variables. AGE could be available with integer and fractional components or used as age at
last birthday. Using AGE in the form as received is certainly one way of including its
impact. The assumption then is that there is a linear association between AGE and the
outcome variable. If the regression coefficient for AGE is positive, then an increase in
AGE by one unit would increase the outcome variable by the amount of the regression
coefficient.
However, if the assumed linear relationship is not true over all ages, the outcome
variable would not be modeled appropriately. In prediction, using AGE as a linear
variable would not result in a good fit at certain ages. One possible modification is
to include AGE as a polynomial, such as quadratic or cubic, to represent the change
of the outcome variable with a change in AGE. Another modification is to categorize
AGE into different levels, such as young, adult, and elder. A further approach might be
to include AGE as a linear term for younger ages and to create a categorical variable
for older ages. Such decisions are routinely made in data analyses and exemplify how
data analysis is an art requiring sound judgment and domain expertise in addition to
the mastery of statistical concepts.
2.2.2 Notation and Assumptions
Using these examples as background, we formally define some key linear model concepts. In general, we have n observations. Our outcome variable is (yi , i = 1, . . . , n),
and for each observation i, there is a set of k explanatory, or independent, variables
{xi1 , xi2 , . . . , xik }, each potentially exhibiting some relationship with yi . The idea is
to postulate a model that approximates the relationship between the outcome variable
and the explanatory variables.
With linear models, we use an estimator other than the sample mean to better
estimate yi and to reduce the error. Using a set of unknown coefficients {0 , . . . , k }
and a random error or disturbance term i , the linear model is defined as
yi = 0 + 1 xi1 + 2 xi2 + + k xik + i
i = 1, . . . n.
(2.4)
21
From these assumptions, it follows that y is a vector of independent, continuous random variables with E[yi |xi , ] = 0 + 1 xi1 + 2 xi2 + + k xik and
Var[yi |xi , ] = 2 . The assumption of constant error variance is called homoskedasticity.3
Another way to view the linear model is via the expected value of yi . This
representation is that used with generalized linear models (GLMs), which model
the mean of the outcome variable. If the errors are normally distributed, then
yi N(0 + 1 xi , 2 ) for a one-variable model. The mean of the normal distribution
for a given xi lies on the regression line. The variance is the spread of the values from
the line, conditional on xi , and is the same 2 as defined in the earlier error model.
2.2.3 Estimation
One way to estimate these parameters is by minimizing the sum of squared errors.
Starting with Equation 2.4, we solve for i and define a function of the -coefficients
to minimize:
n
(yi (0 + 1 xi1 + 2 xi2 + + k xik ))2 .
SS(0 , . . . , k ) =
i=1
To minimize the sum of squared errors with respect to the -coefficients, we take
separate derivatives of SS(0 , . . . , k ) with respect to each -coefficient and set each
equation equal to 0. This results in a system of (k + 1) equations that determine the
estimates of the -coefficients. The vector b is the solution of this system.4
Example 2.5 (One-Variable Linear Model). A one-variable linear model is sometimes referred to as simple linear regression or basic linear regression. For a set of
data, the model is yi = 0 + 1 xi + i with two equations to minimize:
n
SS(0 , 1 )
(yi (0 + 1 xi )) = 0.
= 2
0
i=1
(2.5)
n
SS(0 , 1 )
= 2
xi (yi (0 + 1 xi )) = 0.
1
i=1
3
4
This word is derived from the Greek roots homo, meaning same, and skedannoumi, meaning to scatter. Therefore
homoskedasticity can informally be translated as uniform scatter.
If the outcome data, y, are normally distributed conditional on the explanatory variables, then b is also normally
distributed. In this case, one can derive that the least squares estimator is equivalent to the maximum likelihood
estimator.
22
Solving for each of the -coefficients (relabeled as b0 , b1 for the least squares
estimators) yields
b0 = y b1 x.
(2.6)
n
(yi y)(xi x)
n
.
b1 = i=1
2
i=1 (xi x)
n
2
i=1 (xi x)
, the standard deviation of
Defining the standard deviation of x as sx =
n1
n
n
2
(y
y)
i
i=1
i=1 (yi y)(xi x)
y as sy =
,
and
the
sample
correlation
of
(x,
y)
as
r
=
xy
n1
(n1)sx sy
yields
sy
.
(2.7)
b1 = rxy
sx
This one-variable linear model shows the relationship between regression and correlation. The estimated slope coefficient b1 is proportional to the sample correlation
between x and y.
yi as the fitted value or predicted value. As
We define
yi = b0 + b1 xi , labeling
shown in Section 2.2.6, (b0 , b1 ) are unbiased estimators of (0 , 1 ). This property
implies that E [
yi ] = 0 + 1 xi or equivalently that the expected value of
yi lies on
yi , i =
a line determined by xi and . The residuals, defined as the set of (ei = yi
1, . . . , n), are our estimates of the error terms (i , i = 1, . . . , n). Interestingly, as a
consequence of the estimation process, Equation 2.5 implies that the sum, or average,
of the residuals is always equal to 0.
Example 2.6 (Output of One-Variable Linear Model). The estimate of the intercept from the regression of BMI on AGE is b0 = 36.98, and the estimate of the slope
is b1 = 0.085. The remainder of the output is discussed later in this section.
Regression of BMI on AGE
Estimate Std. Error t value Pr(>|t|)
(Intercept) 36.98262
1.01031
36.605
< 2e-16 ***
AGE
-0.08455
0.01661
-5.091
4.2e-07 ***
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Residual standard error: 7.361 on 1099 degrees of freedom
Multiple R-squared: 0.02304, Adjusted R-squared: 0.02215
F-statistic: 25.91 on 1 and 1099 DF, p-value: 4.196e-07
23
BMI
AGE
Mean
Std. Dev.
31.97
59.34
7.44
13.36
Using the correlation between BMI and AGE as calculated earlier (0.1518), along
with the mean and standard deviation of BMI and AGE shown in Table 2.2, we can
re-create the statistics in the regression output.
With Equation 2.7 we estimate the slope term as b1 = 0.1518 7.44/13.36 =
0.0845. Then using Equation 2.6, we find b0 = 31.97 (0.0845 59.34) = 36.98.
Example 2.7 (Output of Null Linear Model). The mean and standard deviation of
BMI shown in Table 2.2 are also outputs of the model shown in Example 2.1, known
Note that the coefficient estimate for the intercept is same as the mean of BMI
and that the residual standard error is the same as the sample standard deviation of
BMI. In this case, the standard error of the estimate (0.2244) is the quotient of the
residual standard error (7.444) and the square root of the degrees of freedom (1100),
or the usual formula to calculate the standard error of the estimate of the mean.
Section 2.2.4 discusses how to determine the degrees of freedom, and Section 2.2.6
illustrates the calculation of the standard error of the estimate when adding explanatory
variables.
Example 2.8 (Output of Multiple Variable Linear Model). The format of output
from the regression of BMI on AGE, MALE, and BLACK is identical to that with
Example 2.6, with the addition of two lines for the two additional variables.
24
We present the details of this output for completeness with further discussion of the
output later in this section.
unexplained
(2.8)
explained
Squaring both sides of Equation 2.8, summing over i, and performing a little algebra
yields
n
i=1
(yi y)2 =
Total SS
n
i=1
(yi
yi )2 +
Error SS
n
i=1
(
yi y)2 .
(2.9)
Regression SS
In this way, the linear model enables us to decompose the total sum of squares
(Total SS) into two portions: the error sum of squares (Error SS) and the regression
25
sum of squares (Regression SS). Note that the Total SS is a property of the outcome
2
i y)
variable and is not model dependent. Recall that ni=1 (yn1
is used to calculate the
marginal variance of the yi .
As the Regression SS increases, the Error SS necessarily decreases by the same
amount. Recall that the linear model coefficients are the parameter estimates that
minimize SS(0 , . . . , k ). From the above equation, we see that this is tantamount to
shifting some of the Total SS to Regression SS.
A common way to summarize this relationship is through the coefficient of determination:
R2 =
Regression SS
Error SS
=1
.
Total SS
Total SS
(2.10)
R 2 varies between [0, 1] and is a summary statistic that measures the goodness of
fit of a model. It can be shown that R 2 is equal to the square of the Pearson correlation
coefficient between the outcome variable y and the fitted values
y. The positive square
root of the coefficient of determination is denoted by R and is called the multiple
correlation coefficient.
An analysis of variance or ANOVA table is a useful way to summarize some
components of a linear model. For a linear model with n observations, k explanatory
variables, and an intercept term, the ANOVA table is defined as follows:
Analysis of Variance Table for Linear Model
With k Explanatory Variables + Intercept
Source
Sum of Squares
Regression
Error
Regression SS
Error SS
Total
Total SS
Deg. of Freedom
Mean Square
k
n (k + 1)
Regression MS
MSE
n1
sy2
The total degrees of freedom is n 1. The sum of the degrees of freedom associated
with the Regression SS and Error SS must equal n 1. The Regression SS has k
degrees of freedom corresponding to the k explanatory variables. The degrees of
freedom associated with the Error SS must equal n (k + 1) by subtraction. This
can be seen directly by beginning with the n observations and subtracting the (k + 1)
estimated regression parameters, corresponding to the k explanatory variables plus
the intercept term.
The mean square column of the table is found by dividing the sum of squares column
by the degrees of freedom. Dividing the Total SS (the left-side of Equation 2.9) by
26
Table 2.3. ANOVA Table for Regression of BMI on AGE, MALE, and BLACK
Source
Sum of Squares
Deg. of Freedom
Mean Square
Regression
Error
3,623
57,336
3
1,097
1,208.67
52.27
Total
60,959
1,100
55.42
SS
its degrees of freedom (n 1) results in the sample variance of y or sy2 = Total
,
n1
2
which is an unbiased estimator of y . Dividing the Error SS by n (k + 1) results
in the mean squared error (MSE), denoted s 2 . The MSE is an unbiased estimator of
2 , the variance of the linear model error term, and is viewed as a measure of the
improvement offered by the model over simply using the sample mean using sy2 .
The 2 parameter is that introduced at the beginning of Section 2.2.2 when defining
the Var [i ].
27
the basic types of statistics used to choose which variables to include in a linear
model. We discuss how to assess the statistical significance of a variable or set of
variables. In addition, we investigate whether the explanatory variables are collinear
with one another, which could affect their statistical significance. We use the output
from the three models in Examples 2.6, 2.7, and 2.8 to decide whether the variables
are statistically significant.
2.2.5.1 Variable Significance
One way of determining whether to include a particular variable xj is with a t-test.
We consider the null hypothesis (H0 ) that j = 0 against the alternative hypothesis
(H1 ) that j = 0.5 In a one-variable linear model as in Example 2.6, if 1 = 0 then
the regression line is a horizontal line with no slope, or as we have called the null
model. If 1 = 0, then there would be some positive or negative relationship between
the outcome and explanatory variable. This kind of test is called a two-tailed test. We
b
calculate the t-statistic as t = se(bj j ) , where se(bj ) is the standard error of the regression
estimator and is covered formally in Section 2.2.6. The sampling distribution of the
t-statistic is a t-distribution with n (k + 1) degrees of freedom, assuming that the
null hypothesis is true.
We start by selecting a significance level . Formally, = P r [Rejecting H0 | H0 ]
or the probability of rejecting H0 given that H0 is in fact true. The choice of significance
level of 0.05, although conventional, is arbitrary. Many other choices of , such as
1%, 5%, or 10%, are possible and are discipline and contextually dependent.
With a two-sided test, we reject H0 if | t | exceeds the (1 /2)th percentile of a
t-distribution with n (k + 1) degrees of freedom. Otherwise, we fail to reject H0 .
With = 0.05, a rule of thumb is that one rejects H0 (and concludes that the variable
in question is statistically significant) if | t | > 2.
As an alternative to examining the t-statistic, the p-value can be used to decide
whether to reject H0 . Intuitively, a p-value is the probability, under the assumption
that the null hypothesis is true, of observing a t-statistic at least as extreme as the
t-statistic actually observed. More formally, for a two-tailed test, the p-value =
2 P r [T > | t | | H0 ], where T is a t-distributed random variable with n (k + 1)
degrees of freedom. We reject H0 if the p-value < . See Exercise 2.5 for more
intuition for the hypothesis testing approach and its limitations.
Example 2.10 (Significance of Age in a Linear Model). Example 2.6 displays the
output of a regression with AGE as the single explanatory variable. The t-statistic is
equal to 5.091 and the p-value is < 0.001. We therefore reject the null hypothesis
5
28
that the coefficient is equal to zero and say that there is a statistically significant
relationship between BMI and AGE.
The F -statistic, located at the bottom of the regression output, tests whether all of
the slope parameters are equal to zero against the alternative hypothesis that at least
one parameter is not equal to zero.
With one slope coefficient in this model, the F -test and the t-statistic produce
identical results. Mathematically, the square of the t-statistic is the F -statistic. In
this case, the p-value of the t-statistic for a two-tailed test equals the p-value of the
corresponding F -statistic.
Instead of analyzing a single variable, we often wish to compare two models for
which the variables of one model are a subset of the variables of the other. Such
models are called nested models.
Example 2.11 (Nested Linear Model Test of RACE). Suppose we modify our variable BLACK and expand it to RACE, with categories WHITE, BLACK, and OTHER.
We might ask whether RACE with its multiple categories is statistically significant to
include in the model. We refer to the two models as the reduced (without RACE) and
full (with RACE) models. With three levels for the variable RACE, we essentially add
two binary variables to the model as shown in Equation 2.11:
BMI i = 0 + 1 AGEi + 2 MALEi + 3 BLACK i + 4 OTHERi + i
i = 1, . . . , n. (2.11)
Here we have designated the race category WHITE as the reference category
with the inclusion of the binary variables BLACK and OTHER. The ANOVA decomposition concepts previously introduced can be used to address whether the data
support the full model over the reduced model. Table 2.4 displays a portion of the
ANOVA table for this full model, as well as the residuals portion of the ANOVA table
for the reduced model.
We consider the null hypothesis (H0 ) that {3 , 4 } in Equation 2.11 are both equal
to zero against the alternative hypothesis (H1 ) that at least one of {3 , 4 } is not equal
to zero. Thus the reduced model is the one in which only MALE and AGE are retained.
The data collected are used to test whether there is enough evidence to reject the null
hypothesis in favor of the alternative hypothesis.
In this case, we calculate an F -statistic that considers information from the full and
the reduced models and use the information from Table 2.4:
F =
(2.12)
29
RACE
Residuals
Df
Sum Sq
Mean Sq
F value
Pr(> F )
2
1096
2226.99
56351.73
1113.49
51.42
21.66
< 0.0001
F value
Pr(> F )
Residuals
Df
Sum Sq
1098
58578.72
Mean Sq
53.35
The q (or 2 in this case) in the denominator refers to the number of coefficients that we
are testing as equal to zero, whereas the 51.42 is the sf2 ull , or the residual variance from
the full regression. Not surprisingly, the F -statistic that we calculated is identical to
that shown in the full model ANOVA output in Table 2.4. The corresponding p-value
indicates that at least one of the two coefficients (shown by degrees of freedom) for
RACE is significantly different from zero.
While in Example 2.11 we focused on inclusion of one variable RACE, multiple
variables can be tested for inclusion in the model with the F -test. In general, if our
reduced model has k slope parameters and we are testing whether q additional slope
parameters are useful, we calculate an F -statistic that considers information from the
full and the reduced models:
Error SSreduced Error SSfull
=
F =
2
q sfull
2
2
Rfull
Rreduced
q
2
1Rfull
n(k+q+1)
(2.13)
The first equation is the formula that we used in Example 2.11. The second equation is a mathematically equivalent version. The advantage of the second version is
that it can calculate the F -statistic from the usual regression output of the full and
reduced models. The disadvantage is that we need to assess the significance of that
value.
The sampling distribution of the F -statistic under the null hypothesis is a F distribution with q degrees of freedom in the numerator and n (k + q + 1) degrees
of freedom in the denominator, because the statistic is a ratio of two 2 random
variables divided by their degrees of freedom. We compare the F -statistic to the
(1 )th percentile of the F -distribution, because the support of the F -distribution
is the positive real line. Alternatively, the p-value of the F -statistic is calculated and
compared against to determine whether or not to reject H0 .
30
Example 2.12 (Null Linear Model). We can view Example 2.10 in this notation of
full and reduced models. If the coefficient on AGE is equal to zero, our reduced model
contains only an intercept term or what we have called the null model. Because
the intercept is the only parameter in the model, the estimator
y is y for every
2
observation. Therefore, the Error SSreduced = Total SS and Rreduced = 0. Using the
R 2 from Example 2.6 and setting q = 1, k = 0, and n = 1,101, the second part of
Equation 2.13 becomes
F =
2
Rfull
2
1Rfull
n2
0.02304
= 25.92.
(1 0.02304)/1,099
(2.14)
which is an F -statistic and ties to the results at the bottom of the regression output in
Example 2.6.
2.2.5.2 Collinearity
We can assess the importance of including one or more candidate explanatory variables in the model with the t and F tests. It is also good practice to check for
collinearity among the explanatory variables themselves. Collinearity occurs when
one explanatory variable is a linear combination of other explanatory variables. High
correlation among the explanatory variables results in high variances of the corresponding regression parameter estimators. This relationship is discussed further at the
end of Section 2.2.6 in Example 2.14.
The variance inflation factor (VIF) is a useful statistic in measuring the degree of
collinearity present among the explanatory variables.6 In this case, we are seeking to
determine whether one explanatory variable can be explained by a linear combination
of the other explanatory variables.
The VIF for the j th explanatory variable is found by regressing this variable against
the other explanatory variables in the model. The VIF j = (1 Rj2 )1 , where Rj2 is
the coefficient of determination for this linear model.
There is not a consensus for a rule of thumb to use to assess whether collinearity is
an issue. Brien (2007) Values for VIF above 5 or even above 10 are used to suggest
high collinearity of the explanatory variables.
Example 2.13 (VIF for AGE, MALE, and BLACK Model). Using the model with
explanatory variables as in Example 2.9, we calculate the VIF as:
The term aliasing is sometimes used synonymously with collinearity. In the general regression literature, VIF is
used to identify collinearity among the explanatory variables.
31
VIF
AGE
MALE
BLACK
1.0002
1.0200
1.0202
Because the VIF for each of the explanatory variables is below the threshold of
either 5 or 10, then we conclude that these variables are not collinear.
2.2.6 Expected Value and Variance of Regression Estimators
We represent the linear model with multiple explanatory variables as y = X + .
Here y represents the outcome variable and is a column vector of length n. The
explanatory variables are represented by X, which is sometimes referred to as the
design matrix and is of dimension n (k + 1). The design matrix contains a column
of ones corresponding to an intercept term and followed by k explanatory variables.
The column vector of length k + 1 represents the unknown intercept and k slope
parameters. The vector contains the errors and is represented by a column vector of
length n.
We use results from linear algebra to solve the system of (k + 1) equations. Starting
with the sum of squared errors, = (y X) (y X), we take derivatives with
respect to each of the (j , j = 0, 1, . . . , k). The solution of the system of linear
equations for is b = (X X)1 X y,
and the
values for y are equal to
y = Xb.
1fitted
X y, are functions of the known explanaThe least squares estimators, b = X X
tory variables and the outcome data before they are realized. The least squares estimates result once the data are realized and are substituted in the equation for the
estimators. The estimators are random variables, whereas the estimates are specific
values determined by the data being modeled.
The expected value of the regression parameter estimators is
y
1
1
E [b] = E X X
X y = E X X
X (X + ) =
and shows that b is an unbiased estimator for . Using the matrix property that
Var [AX] = A Var [X] A , where A is a constant matrix and X is a random matrix,
the variance of the regression estimators is
1
1
1
1
Xy = XX
X Var [y|X] X X
X = 2 X X
Var [b] = Var X X
where Var [y|X] = 2 as discussed in Section 2.2.2. The Gauss-Markov theorem
states that, of all unbiased estimators of , the least squares estimator has minimum
32
variance (Stigler 1986, p. 148). Note that the resulting matrix for the Var [b] is a
(k + 1) (k + 1) symmetric matrix.
Technically, we need to estimate 2 and call it s 2 , or the residual variance.
1 Thus
what we can actually calculate is the estimated variance or V
ar [b] = s 2 X X . The
standard error of bj , se(bj ), is the square root of the (j + 1)st diagonal value in the
matrix.
Example 2.14 (Linear Model Output). Referring again to Examples 2.6 to 2.8, the
square of the residual standard error is our estimate of s 2 , either 54.18 or 52.27,
respectively. Note that, with the addition of explanatory variables, we have reduced
the residual variance s 2 from that of the marginal variance sy2 (i.e., the null model
equal to 55.41).
The column labeled Std. Error in the regression output represents the square root
of the diagonal elements of V
ar [b]. It can be shown that the standard error of bj is
proportional to the square root of the VIF:
s VIF j
,
se(bj ) =
sxj n 1
where s is the residual standard error and sxj is the standard deviation of the j th
explanatory variable. A higher VIF corresponds to a higher se(bj ), which leads to a
lower t-statistic and a higher p-value. Thus collinearity, which manifests itself in a
higher VIF, can result in distorted measures of the statistical importance of particular
variables.
Note too that the standard error is a function of the number of observations. If that
number is large, then the standard error for all regression estimates will be small,
resulting in larger t-statistics and smaller p-values. Thus the larger sample size will
lead to more statistically significant slope coefficients. However, greater statistical
significance does not imply that the variable is more substantively meaningful to the
analysis.
2.2.7 Model Adequacy: Ra2 and AIC
Section 2.2.5 discussed choosing variables individually or in groups for variable
selection in situations where one model is nested in another. Suppose instead that one
wishes to compare two non-nested models. One might be tempted to use R 2 to decide
between the two models. However, this statistic is ill suited for this purpose because
R 2 will never decrease when adding any explanatory variable, even if the new variable
has no relationship to the outcome variable.
33
R2
Ra2
AIC
MALE
AGE
AGE +
AGE +
AGE +
AGE +
0.01623
0.02304
0.03905
0.05944
0.07558
0.07629
0.01534
0.02215
0.03730
0.05686
0.07221
0.07207
7531.882
7524.238
7508.045
7486.433
7469.372
7470.524
MALE
MALE + BLACK
MALE + RACE
MALE + RACE + UNINSURED
=1
Error SS
n(k+1)
Total SS
n1
=1
s2
,
sy2
(2.15)
34
where log(L) is the log-likelihood function of the model. If the errors are normally
distributed, then Equation 2.15 reduces to
AIC = n log(2) + n log(s 2 ) + n + 3 + k.
The AIC incorporates a penalty for model complexity as measured by the number
of parameters. Therefore selecting the model with the lowest AIC is one way of
optimizing a trade-off between goodness of fit and model complexity.
Example 2.16 (Comparison of AIC). Table 2.6 includes a column for the AIC statistic. Note that the statistic decreases down the list until the last model with inclusion
of the UNINSURED variable. Thus in this case, the AIC and Ra2 measures reach the
same conclusion. This does not necessarily occur in all cases.
2.2.8 Model Validation
There is no unique stopping rule indicating when the modeling process is complete.
Economic considerations may suggest a trade-off between the costs of additional
modeling efforts against the likely marginal improvements in predictive accuracy and
analytical insights. In this section we present different ways of analyzing the model
results to assess the adequacy of the model. First we define influential points (high
leverage and outliers) and discuss how to identify them. We then discuss various types
of graphical plots that may be of value. We end the section with a discussion of two
metrics useful in model selection and model validation.
2.2.8.1 Analysis of Influential Points
An influential point is an observation with a disproportionate impact on the overall
results. Two types of influential points are high leverage and outliers.
Observations that are high leverage points are those that are extreme in value
relative to one or more explanatory variables. If we are predicting BMI for people
with diabetes, then as we show in Example 2.17, persons with high income might
disproportionately influence the result of the analysis. Intuitively, one can think of a
balance beam to represent the regression line or plane, with an observation far out in
the range of an explanatory variable tipping the balance beam one way or another, or
as having a large impact on the value of the regression coefficient estimates.
In Section 2.2.6, we introduced
y =
Xb for
1the fitted value of y and the estimator
X y. If we define H = X(X X)1 X ,
for the regression coefficients as b = X X
then we can rewrite the prediction equation as
y = X(X X)1 X y = Hy. From this
matrix multiplication we see that each predicted observation is written as a linear
combination of the elements of H:
yi = hi1 y1 + hi2 y2 + + hin yn
(2.16)
0.01
0.00
0.00
0.01
0.02
Leverage
0.02
0.03
0.03
0.04
0.04
Income
35
200
200
400
600
800 1000
Observation
Fig. 2.3. Leverage plot for models including INCOME or Log(INCOME) in addition
to AGE, MALE, and RACE. The dotted line represents three times the average leverage
for each model.
The diagonal elements of H, {hii , i = 1, . . . , n}, are called the leverage values and
. There is no
are restricted to lie between n1 hii 1. The average leverage is k+1
n
standard rule of thumb for a high leverage point, but authors have suggested two or
three times the average as cut-off values.
Example 2.17 (High Leverage). Suppose we believe that in addition to AGE, MALE,
and RACE, INCOME is associated with BMI. We include this variable in the model
and find that it is significant with a p-value = 0.039 and a negative coefficient. The
adjusted-R 2 is 0.075, an improvement on the model without this variable as seen in
Table 2.6. Thus conceptually, this relationship makes sense as INCOME increases, BMI
decreases. We create a plot that visually shows the leverage values, hii , i = 1, . . . , n
in Figure 2.3.
For the model including INCOME there are 14 observations whose leverage values
exceed three times the average. The observed value for BMI for these observations will
receive greater weight in their predicted value for BMI as inferred from Equation 2.16.
Investigating these 14 observations reveals that these individuals all have incomes
exceeding $190,000.
One strategy for dealing with high leverage points is to do nothing and disclose the
influences of these observations. Another solution is to consider certain variable
36
The standard error of the residual se(ei ) is equal to 1 hii , where hii is the
ei
leverage. The standardized residual is therefore estimated as s 1h
. A large residual
ii
leads to a large outlier. But a large leverage value also increases the standardized
residual. The rule of thumb for detecting outliers is similar to that for high leverage
points, where 2 or 3 may be used as a threshold.
Example 2.18 (Outliers). Using the same models as in Example 2.17, Figure 2.4
shows graphs of outliers similar to those of the leverage values.
Both graphs indicate a number of observations whose residual exceeds the threshold
of 2. In fact, there are 49 such outliers for the model using INCOME and 48 outliers
for the model using log(INCOME), and are a subset of the outliers of the model
containing the INCOME variable.
For the model including INCOME, only 3 of the 14 high leverage points are outliers, whereas for the model including log(INCOME), 2 of the 4 high leverage
points are outliers. Figure 2.7 in Example 2.21 shows this pictorially and extends this
example.
2.2.8.2 Residual Plots
Plots of residuals, as well as other validation summaries, help ensure that given all of
the information available, the model seems reasonable and that further changes are
unnecessary. As mentioned previously, residuals measure the leftover variation of the
outcome variable from a model containing certain explanatory variables. We create
plots of the residuals to show (i) whether the residuals resemble a normal distribution,
Log(Income)
Stnd. Residual
37
200
200
Fig. 2.4. Standardized residual plot for models including INCOME or Log(INCOME)
in addition to AGE, MALE, and RACE. The 2 levels for outliers for each model are
represented as a dotted line.
(ii) no relationship between the residuals and explanatory variables in the model, (iii)
no relationship between residuals and explanatory variables not in the model, and (iv)
no relationship between residuals and the fitted values.
When plotted against explanatory variables already in the model, the residuals
could display a pattern, such as a quadratic, that would suggest that a squared term
of the variable is appropriate to add to the model. A similar plot against the fitted
values would show a similar pattern, but would not identify which variable to alter.
Plotting the residuals against the fitted values helps check for homoskedasticity of the
residuals. If there is a pattern where the spread of the residuals increases or decreases
with larger fitted values, then some adjustments must be made or other techniques
considered.
If no patterns exist, then plotting the residuals against explanatory variables not
in the model would be the next step. If a pattern exists, then additional iterations of
model building are required.
Example 2.19 (Residual Plots). Suppose we examine the standardized residuals
from the model of BMI on AGE + MALE + RACE + Log(INCOME) with respect
38
Stnd. Residual
White
Black
Other
20
30
40
50
AGE
60
70
80
26
28
Female
Male
30 32 34
Fitted Value
36
38
to AGE and another with respect to RACE and SEX. Example 2.4 examined a similar
relationship in Figure 2.2 where we saw that black females had a different median
BMI than other groups. We had grouped whites and non-blacks into the other category,
whereas now we separate them.
Figure 2.5 shows these residual plots, with the left-hand plot showing standardized
residuals versus AGE, and the right-hand plot showing standardized residuals plotted
against the fitted values with the plotting symbols differing by SEX and by RACE.
The AGE plot includes a smoothed nonparametric line, known as a LOESS curve, that
indicates that there exists a modest quadratic-type relationship between the residuals
and AGE.
The SEX and RACE plot is more complicated but quite informative. Data points for
males are colored black, whereas for females the dots are gray. Whites, blacks and
other races are denoted by triangles, boxes, and asterisks, respectively. Interestingly,
observations for females are a large proportion of the outliers. In the data, there are
approximately equal numbers of males and females for the white and other groups, but
twice as many black females as black males. The plot indicates that there might be an
interaction effect between SEX and RACE. Including such a term in the model amounts
to the creation of six categories, one for each SEX/RACE combination, similar to what
was presented in Figure 2.2.
39
40
Null Model
6
Sample Quantiles
2 1 0
1
2
Theoretical Quantiles
2 1 0
1
2
Theoretical Quantiles
Fig. 2.6. QQ-plots from null model for BMI versus BMI on AGE + AGE2 + MALE
+ RACE + MALE:RACE + Log(INCOME).
.
Di =
2
(k + 1) s
se(ei )
(k + 1) (1 hii )
In the first equality, Di summarizes the difference in fitted values from the original
model using all n observations versus those in the model after removing the ith
observation. The larger the difference, the larger is Di . The second equality shows a
different version of the statistic that separates the outlier component from the leverage
component; it is in a form that is easier to compute. Overall the statistic summarizes
the influence of the ith observation. Again, there is no standard rule of thumb for
4
, or Di > n4 , have appeared in
assessing Cooks distance, but Di > 1, Di > n(k+1)
the literature.
Example 2.21 (Cooks Distance). Using the same models as in Examples 2.17 and
2.18, Figure 2.7 shows a combination plot of Cooks distance, outliers, and high
Log(Income)
Stnd. Residual
26
28
30 32 34
Fitted Value
41
36
38
26
28
30 32 34
Fitted Value
36
38
Fig. 2.7. Combination plot of Cooks distance, outliers, and high leverage points for
models including INCOME or Log(INCOME) in addition to AGE, MALE, and RACE.
The dotted lines represent thresholds for outliers, the triangles represent the high
leverage points, the tiny dots represent the other observations, and the open circles
represent those observations with high Cooks distance.
leverage points, together with their fitted values. The approach is to plot standardized
residuals versus the fitted values. The triangles represent the high leverage observations, the tiny dots are the other observations, and the observations with high Cooks
distance have an open circle around them.
The plot is complicated but quite illustrative. First, points that have a high Cooks
distance are sometimes outliers, sometimes high leverage, sometimes both outlier and
high leverage, and sometimes neither outlier nor high leverage. Dots with circles are
those nonhigh leverage points that are above the threshold for a high Cooks distance.
For the left-hand plot using INCOME, we see many high leverage observations
with residuals close to zero with small fitted values. Examining the data reveals that
these observations have the highest values for INCOME. Thus these observations do
influence the estimation of the regression plane and result in small residual values.
Yet other observations with high leverage values do not have this same effect.
In reviewing both plots, the observations with large values of Cooks distance
generally seem to be the same observations that are outliers. However, there are
observations using INCOME or Log(INCOME) where the threshold for Cooks distance is exceeded but they are not outliers or high leverage points. These observations
42
bear some investigation because they may be ones whose characteristics give guidance
as to the next steps.
As the right-panel indicates, the use of Log(INCOME) helps reduce the number of
high leverage points and the influence on the regression parameter estimates. However,
many outliers warrant some investigation.
2.2.8.4 SSPE and PRESS
Ra2
and AIC statistics are good ways of assessing model adequacy with
Although the
the data used to estimate the model parameters, the gold standard of model validation
is out-of-sample testing. The idea is to first fit the model to a training (or referred to
as an in-sample) dataset. Once the estimates are found, they are applied to evaluate
the predictive accuracy of a model on a separate hold-out (or referred to as an out-ofsample) data set.
The sum of squared prediction errors (SSPE) is a statistic that considers the squared
residuals of the hold-out sample when applying the parameter estimates from the
training sample. Consider a dataset of n observations where n1 observations are used
in the training sample and n2 observations are used in the hold-out sample. The SSPE
is
SSPE =
n
1 +n2
(yi
yi )2 .
i=n1 +1
As with the AIC, the SSPE is not meaningful in isolation, but is useful for comparing
between models. The lower the SSPE, the better the prediction fit.
With small datasets, it may not be feasible to split a sample into the two pieces.
Sophisticated techniques called cross validation involve an iterative process using
one dataset. One such method is called predicted residual sum of squares (PRESS).
Similar to the notation introduced with Cooks distance, define
y(i) as the predicted
value for the ith observation after removing that observation to fit the model.
For
2each
y(i) and
of the n observations in the sample, calculate the squared residual yi
sum over the n observations. The PRESS statistic is
2
n
n
2
ei
yi
y(i) =
.
PRESS =
1 hii
i=1
i=1
As with Cooks distance, the PRESS statistic is a function of the residual and leverage.
A larger residual or larger leverage values lead to a larger PRESS statistic. Also as
with the SSPE, a lower PRESS is preferred.
Example 2.22 (Section 2.2 Examples Summary). In this section we have discussed
10 models to illustrate the concepts underlying the main linear models. Table 2.7
summarizes these models by listing their included variables, as well as indicating the
43
Explanatory Variables
Null
AGE
MALE
AGE + MALE +
AGE + MALE +
AGE + MALE
AGE + MALE +
AGE + MALE +
AGE + MALE +
AGE + AGE2 +
+ MALE:RACE
Examples
BLACK
RACE
RACE + UNINSURED
RACE + INCOME
RACE + log(INCOME)
MALE + RACE
+ log(INCOME)
examples in which these models are discussed. The overall summary measures are
included in Table 2.8, including the PRESS and SSPE results.
As discussed at the beginning of Section 2.2, the data consist of one year of
data from two different panels (13, 14) of MEPs data that represent different sets of
households. The PRESS statistics are calculated from panel 13 data that were used to
estimate the model parameters. The SSPE statistics are calculated from panel 14 data,
a separate sample of persons with diabetes. Panel 13 consists of 1,101 observations,
whereas panel 14 consists of 1,039 observations.
Table 2.8 shows the model name, as well as the summary measures that we have
discussed. The best model of a group of candidate models results in the lowest s 2 ,
AIC, PRESS, and SSPE statistics and the highest Ra2 statistic. Although this constraint
might not be met in most analyses, in this particular case, m9 with the explanatory
Table 2.8. Comparison of Summary Measures
Model
m0
m1
m2
m3
m4
m5
m6
m7
m8
m9
df
s2
R2
Ra2
AIC
PRESS
SSPE
1100
1099
1099
1097
1096
1098
1095
1095
1095
1092
55.42
54.19
54.57
52.27
51.42
53.35
51.42
51.26
51.19
50.02
0.00
0.02
0.02
0.06
0.08
0.04
0.08
0.08
0.08
0.10
0.00
0.02
0.01
0.06
0.07
0.04
0.07
0.07
0.08
0.10
7548
7524
7532
7486
7469
7508
7471
7467
7465
7443
61070
59765
60179
57751
56850
58884
56933
56775
56681
55497
51841
50137
51301
50707
49756
49587
49727
49795
49937
48580
44
variables AGE + AGE2 + MALE + RACE + MALE:RACE + log(INCOME) satisfies all constraints.
We are not advocating the use of this model to predict BMI for those with diabetes,
but of the models examined, this one is the best choice.
2.2.9 Prediction
In much of actuarial work, once a linear model is finalized, the next step is to
use the model to calculate predictions of some future outcome. As introduced in
Section 2.2.6, the linear model is expressed as y = X + , with the fitted values
expressed as E [y | X] = Xb =
y. For a given set of explanatory variables the fitted
value, or the average value of the outcome variable, is estimated as Xb.
To address the variability of this estimator, the literature presents two types of
intervals: the confidence interval and the prediction interval. The confidence interval addresses the variance from the regression line, whereas the prediction interval
addresses the variability of new observations.
We begin with Var [
z ], the variance of the fitted value for a particular observation
z represents a particular
with covariate vector of xz of dimension 1 (k + 1). Thus
value on the regression line. Recall that the design matrix, X, has dimension n
(k + 1) and the dimension of is (k + 1) 1:
Var [
z ] = Var [ xz b ]
= xz (Var [ b ]) xz
1
= 2 x z X X
xz .
1
[
xz .
Var
z ] = s 2 x z X X
1
for Var [ b ] as in Section 2.2.6 for the variance
The third line substitutes 2 X X
of the regression parameter estimator. Because 2 is not known, we substitute s 2 as
an estimator in the last line to indicate that we are estimating the variance.
Note that the dimension of the matrix V
ar [
z ] is a scalar of dimension 1 1. The
confidence
interval
for
the
fitted
value
of
this
observation
is expressed as
z t1/2
V
ar [
z ], where t1/2 is the t-value that results in an overall significance level of .
The prediction interval is developed similarly; however, the starting point is z =
xz + z , because we are indicating the variability for the value of the outcome
variable itself and not its average. Here the variability of the predicted value is
larger, because it includes an extra term for the variance
of the individual error term.
ar [
z ] or equivalently as
The prediction
interval is expressed as
z t1/2 s 2 + V
z t1/2 s
45
7
8
We continue to use the same MEPS data as in Section 2.2, but switch our analysis from studying BMI to analyzing
annual health care expenditures.
These data are a weighted sample from the United States to allow national estimates. In this case study we do not
take into account the weights of each observation, so the conclusions from this study are not representative of the
U.S. population. For our study we consider only those with expenditures greater than 0.
46
The initial dataset contained 2,164 observations. As is often the case in actuarial
work, several of the variables have missing values. Although a complete discussion
of the treatment of missing data is beyond the scope of this chapter, we mention a few
central concepts: for more detail, see Little and Rubin (2002) and Rubin (1976).
Broadly speaking, one can discard observations with missing data, elect not to use
variables with missing data, impute missing values, or employ a combination of all
three methods. If the outcome variable is missing for a particular observation, then
that observation would provide no benefit in the estimation phase. If observations
with missing data for explanatory variables are deleted from the dataset, the timing
of when these observations are deleted could influence the resulting predictions. If
the observations were deleted before the analysis started, then the dataset could be
reduced dramatically because some potential variables might have had many missing
data points. If those variables were determined not to be needed, then some observations may have been deleted prematurely and results could be biased. We deleted
24 observations that had missing values in either the COMORB.CNT or MENTHEALTH
variables. See Exercise 2.6 for a further exploration of the treatment of missing data.
Summary statistics and plots of the reduced sample of 2140 observations are displayed in Table 2.9 and Figures 2.8 to 2.10. The SMOKER variable, an important risk
factor influencing health expenditures, is problematic because it is missing in approximately 7% of the total cases. Furthermore, the median health care log-expenditures
of smokers and nonsmokers are not materially different, as seen in the boxplot in
Figure 2.10. Although not impossible, this finding is perhaps counterintuitive and
seemingly at odds with the literature considering the health complications associated
with smoking. The median expenditures for the missing data are lower than the medians for the smokers and nonsmokers, with a wider variance. These relationships are
true in the aggregate and are not conditional on other variables that could modify
the relationship. For the purpose of this case study we elect to discard the SMOKER
variable from consideration.
We see the mean of our outcome variable, EXPEND, is $11,162 with a standard
deviation of $18,331 and maximum of $247,828. The distribution is quite skewed, as
shown in Figure 2.8. Income is also a skewed variable. The mean of AGE is 60 with a
range of 18 to 85. Fifty-seven percent of the observations are FEMALE.
Note that the categorical variables RACE and MENTHEALTH are both treated as sets
of binary variables, where the sum of their means equals one. RACE is an unordered
categorical variable, whereas MENTHEALTH is ordered. Twenty-five percent of the
sample is BLACK, indicating the emphasis in the survey design of oversampling
minority populations. Eleven percent of the observations represent those who are
UNINSURED, and 18% have MEDICAID. Forty percent of the sample is EMPLOYED
as measured on their first interview during the calendar year. The proportion of
observations by MENTHEALTH category is highest in level 3 (good) and lowest in
47
Mean
Median
Std Dev
Min
Max
EXPEND
LOGEXPEND
INCOME
LOGINCOME
2,140
2,140
2,140
2,140
11,162
8.38
47,816
10.4
4804
8.48
34,518
10.45
18,331
1.52
43,511
0.91
4
1.39
1,000
6.91
247,828
12.42
359,704
12.79
AGE
BMI
2,140
2,140
59.78
31.77
60
30.7
13.72
7.26
18
9.4
85
75.1
COMORB.CNT
HIGHBP
ASTHMA
CANCER
CHOLEST
CORONARY
STROKE
EMPHYSEMA
2,140
2,140
2,140
2,140
2,140
2,140
2,140
2,140
2.16
0.76
0.14
0.15
0.74
0.20
0.12
0.05
2
1
0
0
1
0
0
0
1.21
0.42
0.35
0.36
0.44
0.40
0.32
0.22
0
0
0
0
0
0
0
0
7
1
1
1
1
1
1
1
SMOKER
MENTHEALTH.1
MENTHEALTH.2
MENTHEALTH.3
MENTHEALTH.4
1,981
2,140
2,140
2,140
2,140
0.16
0.25
0.24
0.36
0.15
0
0
0
0
0
0.36
0.43
0.43
0.48
0.36
0
0
0
0
0
1
1
1
1
1
MEDICAID
UNINSURED
2,140
2,140
0.18
0.11
0
0
0.39
0.31
0
0
1
1
EMPLOYED
HISPANIC
2,140
2,140
0.40
0.25
0
0
0.49
0.43
0
0
1
1
RACE.white
RACE.black
RACE.other
FEMALE
2,140
2,140
2,140
2,140
0.65
0.25
0.10
0.57
1
0
0
1
0.48
0.43
0.30
0.50
0
0
0
0
1
1
1
1
level 4 (fair or poor). Those with diabetes have from 0 to 7 of the specific comorbities,
with a mean of 2.16. Three-quarters of the sample have high blood pressure or high
cholesterol.
Figure 2.8 displays histograms of health care expenditures (EXPEND) as well as its
natural logarithm (LOGEXPEND). The former quantity is highly skewed. In contrast,
LOGEXPEND has a roughly symmetric, bell-shaped distribution. We model LOGEXPEND as a linear combination of explanatory variables.
In large-scale projects, the choice of which potential explanatory variables to collect
and analyze typically involves weighing the marginal cost of gathering and analyzing
48
0.00
0e+00
0.10
4e05
0.20
8e05
0.30
EXPEND
100000
250000
10
data elements against the likely benefits in the form of improved predictive accuracy.
In many contexts, it is not uncommon to confront databases with hundreds or even
thousands of candidate explanatory variables. According to the American Diabetes
Association website (American Diabetes Association 2013b), complications of having diabetes include weight issues, eye troubles, hearing difficulties, foot and skin
complications, heart troubles, risk of stroke, kidney disease, neuropathy, and mental
health issues. For this case study, we used our judgment to preselect a manageable
number of variables for further analysis.
A final step in the analysis design phase is to specify a model validation strategy.
The gold standard of model validation is to evaluate the predictive accuracy of a
model on a hold-out sample of data that were not used to train the model. We use
panel 13 (1,101 observations) to train candidate models and set aside panel 14 data
(1,039 observations) to validate the final selected model.
2.3.1.1 Exploratory Data Analysis
Once the analysis design is in place and the data have been prepared, the analysis
phase begins. It is best to visually explore data before launching into a formal modelbuilding process. The statistician John Tukey pioneered the discipline of graphical
exploratory data analysis (EDA) and commented that the greatest value of a picture
is when it forces us to notice what we never expected to see (Tukey, 1977, p. vi).
At a minimum, we want to inspect variable distributions, as well as the relationships
of the various explanatory variables with the outcome variable. For this case study,
much of this relevant information can be conveyed in two images.
Figure 2.9 displays a scatterplot matrix of the outcome variable LOGEXPEND and
the available continuous explanatory variables. The lower cells of this graphical matrix
plot one variable versus one other variable. Note that INCOME is included on both the
7 8 9
11
13
49
2 4 6 8 10
20 40 60 80
INCOME
LOGINCOME
BMI
0.06 0.08
20 40 60 80
0.17
10 30 50 70
7 8 9
AGE
0.39 0.28
0.38
COMORB.CNT
11
13
0.86
0 150000 350000
2 4 6 8 10
LOGEXPEND
0 150000 350000
10 30 50 70
raw and logarithmic scale. The upper cells of the matrix display the sample Pearson
correlation coefficient between variables. Finally, histograms of the variables are
included along the diagonal.
The plots of the relationship of the variables with LOGEXPEND are shown in
the last row, whereas the pairwise correlations with LOGEXPEND are displayed in
the last column. These relationships are all roughly linear, the strongest being with
age and comorbidity count. The scatterplot matrix also enables a check for pairwise
collinearity among the explanatory variables using the correlation statistic. The highest
explanatory variable correlation (between age and comorbidity count) is 0.39. All of
the correlation coefficients displayed are significant at the 0.01 level, meaning that all
correlations are significantly different from zero.
50
CHOLEST
CORONARY
24%
76%
86%
14%
85%
15%
26%
74%
80%
20%
STROKE
EMPHYSEMA
88%
12%
95%
5%
SMOKER
MENTHEALTH
99
UNINSURED
EMPLOYED
89%
11%
60%
40%
75%
25%
HISPANIC
RACE
MEDICAID
82%
18%
FEMALE
3 other
14
12
10
8
6
4
2
0
ASTHMA
2 black
14
12
10
8
6
4
2
0
HIGHBP
1 white
LOGEXPEND
14
12
10
8
6
4
2
0
43%
57%
Figure 2.10 displays boxplots summarizing the distribution of LOGEXPEND (on the
y-axis) by the levels for the various categorical variables. The outer edges of the boxes
denote the 25th and 75th percentiles of LOGEXPEND within the particular category.
The thick black line inside each box denotes the median value of LOGEXPEND within
each category. Under each boxplot is the percent of observations contained in that
level. These boxplots indicate that most of the categorical variables (with the possible
exceptions of FEMALE and SMOKER) are somewhat correlated with LOGEXPEND. The
SMOKER variable has three levels: smoker, nonsmoker, and missing (coded as 99).
2.3.1.2 Iterative Modeling Process
Given the small number of available explanatory variables, we begin by fitting a
model containing all of the explanatory variables other than comorbidity count. If
51
value Pr(>|t|)
9.966 < 2e-16 ***
0.599 0.549472
0.392 0.694879
2.605 0.009307 **
2.875 0.004119 **
-0.688 0.491442
-2.644 0.008322 **
-4.175 3.22e-05 ***
-3.028 0.002520 **
-6.862 1.14e-11 ***
0.593 0.553414
-0.189 0.850095
1.603 0.109203
3.379 0.000753 ***
1.509 0.131489
1.580 0.114432
5.400 8.19e-08 ***
2.161 0.030899 *
2.185 0.029099 *
1.938 0.052919 .
2.589 0.009745 **
0.01 * 0.05 . 0.1 1
all of the comorbidity indicators were included together with COMORB.CNT, perfect
multicollinearity would result, because COMORB.CNT is a linear combination of the
disease indicators. The correlations shown in the last column of Figure 2.9 show the
pairwise correlation of LOGEXPEND and motivate the inclusion of INCOME on a log
scale rather than the original scale. This decision has the added benefit of preventing
possible high leverage points later in the analysis. The resulting model, denoted CS1,
is displayed in Table 2.10.
The very low t-statistics for the model coefficients corresponding to LOGINCOME, FEMALE, and MEDICAID suggest that these variables have fairly low marginal
explanatory power in the presence of the other variables.9 Excluding these variables
yields Model CS2 displayed in Table 2.11.
9
Although the coefficient for FEMALE is not significant, it is usually prudent to keep this variable in the model for
communication with other stakeholders because it is an important variable of interest.
52
value Pr(>|t|)
20.315 < 2e-16 ***
2.553 0.010829 *
2.958 0.003167 **
-0.693 0.488714
-2.582 0.009967 **
-4.158 3.47e-05 ***
-3.299 0.001003 **
-7.393 2.87e-13 ***
-0.200 0.841696
1.604 0.108972
3.444 0.000595 ***
1.487 0.137385
1.604 0.109071
5.412 7.66e-08 ***
2.160 0.031015 *
2.199 0.028067 *
1.986 0.047302 *
2.602 0.009387 **
0.01 * 0.05 . 0.1 1
Because Models CS1 and CS2 are nested, an F -test can be performed to ascertain the joint significance of the three omitted variables: LOGINCOME, FEMALE, and
MEDICAID. The resulting F -statistic is 0.2258 (p-value = 0.8785); that is, does not
reject the null hypothesis that all coefficients are equal to zero. Excluding these three
variables decreases the AIC from 3767 to 3762. This is consistent evidence that the
three variables in question offer little explanatory or predictive power in the presence
of the remaining explanatory variables. Note that Model CS1 has a higher R 2 but a
lower Ra2 than Model CS2. This illustrates why it is unwise to decide between models
using R 2 .
Although both prior intuition and the EDA suggest a positive relationship between
EMPHYSEMA and LOGEXPEND, as well as between STROKE and LOGEXPEND, an
examination of the t-statistics from Model CS2 suggests that these comorbidity indicators are of marginal importance to the model. This prompts us to investigate the
effect of excluding these variables. The AIC increases from 3762 to 3763, and the
F -statistic is 2.54 (p-value = 0.08). These statistics are borderline as to whether to
exclude or include these variables. For purposes of this case study, we elect to retain
these variables in the model.
53
Stroke Only
Standardized Residuals
2 1
0
1
2
All Data
20
30
40
50 60
AGE
70
80
20
30
40
50 60
AGE
70
80
If we examine the signs of the coefficients, the directions are intuitive. As people
age or increase in BMI, their expenditures tend to increase. Relative to whites, blacks
and others have smaller expenditures. If one is employed, expenditures are less. If
one is uninsured, there is less spending. As ones mental health declines, expenditures
increase. Finally, the presence of a comorbidity increases expenditures.
54
on average 0.96% (= e0.0096 1) higher for each additional year of age. This interpretation follows from the fact that the model estimates expenditures on the log
scale.
However, it is possible that this rate of increase might differ for different subpopulations. The right-side of Figure 2.11 indicates that this is likely the case for STROKE.
The plot shows the standardized residuals versus AGE only for individuals diagnosed
with having had a stroke.10 The downward pattern in the residuals indicates that our
model fails to adequately capture the relationship between age and expenditures for
the population of stroke victims. This motivates us to add an interaction between AGE
and STROKE to the model. The resulting model is displayed in Table 2.12.
The AGE:STROKE interaction is significant, and lowers the AIC from 3762 to 3758.
The negative coefficient for this interaction term might seem counterintuitive, because
it suggests that expenditures of people who have not had a stroke increase more with
10
Individuals diagnosed with having had a stroke are all more than 40 years old, which is why the data points appear
bunched in the right-hand side of the plot.
55
STROKE=1
12
12
STROKE=0
11
Actual
Predicted CS2
Predicted CS3
10
LOGEXPEND
8
9
10
11
Actual
Predicted CS2
Predicted CS3
20
30
40
50 60
AGE
70
80
20
30
40
50 60
AGE
70
80
age than the expenditures of people who have had a stroke. When an interaction term
is present, all of the related model parameters must be interpreted jointly. In this case
study, this means that one needs to consider the coefficients for AGE and STROKE,
and the interaction term for both. Data visualization provides helpful insight for this
situation.
Figure 2.12 plots LOGEXPEND versus AGE separately for the subpopulations without and with a stroke. The thick black line in each plot is an empirical LOESS
curve of the data, whereas the gray lines are LOESS-smoothed predicted values for
Models CS2 and CS3, respectively. The scatterplot on the left suggests that for the
nonstroke subpopulation, Models CS2 and CS3 capture the relationship between AGE
and LOGEXPEND equally well, because the lines are indistinguishable.
The thick line on the right plot suggests that having a stroke is associated with
higher expenditures earlier in life that do not increase with age, but generally decrease
slightly. This is consistent with the coefficients of Model CS3. However, Model CS2
predicts an upward relationship that does not exist in the data. The coefficient for
AGE in Model CS3 shows that each additional year of age for otherwise identical
nonstroke victims is on average associated with 1.2% (= e0.012 1) higher health
care expenditures. This, together with the coefficient for the AGE:STROKE interaction,
implies that each additional year of age for otherwise identical stroke victims is on
average associated with 1.6% (= 1 e(0.0120.028) ) lower health care expenditures. It
is also noted from the right-hand graph that the level of expenditures is higher for
those having had a stroke versus those who have not, which concurs with our prior
knowledge.
56
As a caveat, what the included data do not indicate is the age at which the person had
a stroke or the time since they had a stroke. This extra information may be informative
for the analysis. Our purpose here is to demonstrate the use of graphical techniques
and ways to interpret regression output with interaction terms.
2.3.3 Interpreting Model Coefficients
The overall intercept term is equal to 6.818. This is the average LOGEXPEND for
unemployed, insured white, non-Hispanic individuals in excellent mental health and
with no comorbidities, and for whom AGE = 0 and BMI = 0. Because the last two
conditions are never met, the intercept term has no interpretation in isolation from the
other terms in the model.
When the AGE:STROKE interaction term is introduced, the coefficient for STROKE
jumps from 0.226 in Model CS2 to 2.103 in Model CS3. This seemingly large increase
in the STROKE coefficient cannot be analyzed in isolation from the AGE:STROKE
interaction term. Because there are no observations with AGE = 0, the STROKE coefficient has no meaning in isolation. In Model CS3, an amount on a log scale, ranging
from 0.504 (= 0.028 18) to 2.38 (= 0.028 85), is reflected in the fitted
equation for the interaction term AGE:STROKE depending on the age of the person.
More interpretable model parameters result if we center the continuous explanatory variables. Specifically, we replace AGE and BMI with AGE.C = AGE 60 and
BMI.C = BMI 32, where the values subtracted are the appropriate means of the
respective variables. All of the resulting model coefficients and goodness of fit statistics remain unchanged except for the intercept term, which increases from 6.818 to
8.095, and the STROKE coefficient, which decreases from 2.10 to 0.419. For otherwise
identical subpopulations of 60-year-old individuals, having had a stroke diagnosed
is associated with 52% higher (= e0.419 1) health care expenditures. Similarly, the
intercept term corresponds to the expenditures of individuals with average age and
BMI, and in the reference category for the other variables in the model.
2.3.4 Model Validation
A number of residual plots from Model CS3 are displayed in Figures 2.13 and 2.14.
The upper left image in Figure 2.13 is a QQ-plot indicating that the standardized
model residuals are, to a reasonable approximation, normally distributed. The upper
right image displays a scatterplot of the standardized residuals versus the predicted
values. This is the most commonly displayed residual plot. No relationship between
the residuals and predicted values is discernible, and the homoskedasticity assumption
does not appear to be violated. Further, horizontal lines are drawn at the empirical 2.5
and 97.5 percentiles of the standardized residual distribution. These lines are close to
57
Residuals versus Predicted
Normal QQ Plot
10
2
1
0
1
3
20
30
40
50
60
70
80
10
20
30
40
50
60
70
the corresponding theoretical standard normal percentile values of 1.96 and 1.96,
respectively.
The bottom images in Figure 2.13 display the standardized residuals plotted against
AGE and BMI, respectively. Again, no suspicious patterns are apparent. The boxplots
in Figure 2.14 indicate that the standardized residuals have essentially the same
distribution within each level of the various categorical explanatory variables. In
short, at a high level everything seems to be in order. Further checks for interactions
could be completed, but as discussed previously, the benefits of spending more time
must be weighed against the costs of further effort.
Once we are satisfied that our model adequately captures the salient features of
the training data, it is appropriate to evaluate its predictive accuracy on the hold-out
58
HIGHBP
ASTHMA
CANCER
CHOLEST
2
1
0
1
2
3
0
3
CORONARY
STROKE
EMPHYSEMA
MENTHEALTH
2
1
0
1
2
3
0
3
UNINSURED
EMPLOYED
HISPANIC
RACE
2
1
0
1
2
1
3 other
2 black
1 white
data. Recall that the hold-out data have not been used in the model estimation process.
Figure 2.15 displays two predictive validation analyses. On the left is a scatterplot
of predicted versus actual values of LOGEXPEND. The sample correlation of these
quantities is 0.49 (p-value < 0.0001), an improvement over the one-way correlations
observed in the EDA scatterplot.
On the right is a gains chart, a model validation tool that facilitates an assessment
of the economic significance of the predictive accuracy of the model. Here, the holdout data are sorted from highest to lowest predicted LOGEXPEND. Next, the data
are grouped into 10 equal-sized deciles. For example, decile 1 contains the 10% of
the observations with the highest predicted expenditures, and decile 10 contains the
10% of the observations with the lowest predicted expenditures. The total dollars
of expenditures are calculated within each decile. Finally, the cumulative percent of
these expenditures are plotted against the cumulative percent of individuals. The fourth
2.4 Conclusion
r = 0.49
Gains Chart
Cumulative % of EXPEND
0.0 0.2 0.4 0.6 0.8 1.0
Predicted LOGEXPEND
7
8
9
10
59
4
6
8
10
Actual LOGEXPEND
12
0.0
0.2
0.4
0.6
0.8
1.0
Cumulative % of Observations
point from the left of the gains chart (circled) indicates that the 30% of the individuals
with the highest predicted expenditures account for approximately 50% of the total
actual expenditures in the hold-out data. In contrast a random selection of 30% of
the individuals in the hold-out data would on average account for approximately 30%
of the expenditures. A gains chart conveys the degree to which using the model to
rank individuals in terms of predicted expenditures improves on randomly selecting
individuals.
Another model exercise, suggested in Exercise 2.3, is to reestimate the model
parameters on the hold-out data and apply that resulting model to the training data.
This enables a comparison of the two sets of model coefficients for consistency. It
provides further evidence that our model primarily reflects aspects of the processes
that generated the data, rather than accidental features of the data arising from random
variation. Therefore it is not unreasonable to expect the model to make comparably
good predictions when applied to future observations. Finally, applying this reestimated model to the training data to calculate the predicted values yields a correlation
coefficient of 0.52 with LOGEXPEND.
In short, although the model could be improved, no major problems have emerged
from our various validation checks, and it is reasonable to bring the analysis to a close.
2.4 Conclusion
This chapter has provided a practical overview of linear models. First, it outlined
fundamental assumptions and other statistical concepts. Second, these concepts were
woven into a series of examples to help reinforce the material. This chapter serves as
a foundation for the rest of the volume in that many of the subsequent chapters relate
to linear models in some way.
60
Chapter 5, Generalized Linear Models, relaxes two of the linear model assumptions. First, rather than assuming that the expected value of the outcome variable
equals a linear combination of explanatory variables, the GLM framework assumes
linearity on a monotonic scale (such as logarithmic, inverse, or log-odds) provided by
a link function. Second, the normality assumption for the dependent variable is relaxed
to one from an exponential family distribution. The Error SS concept is generalized to
the concept of deviance. Many of the variable selection, EDA, and model validation
concepts discussed in this chapter carry over to the broader GLM framework.
Chapters 8 and 16 on mixed models, also known as multilevel/hierarchical models
(Gelman and Hill 2008), can be viewed as regression models for which certain
model parameters are given probability submodels. Such models are often useful
when the data are naturally grouped along one or more dimensions. These models
incorporate parameter shrinkage, an important phenomenon that is closely related
to credibility weighting. Furthermore, Bayesian versions of (generalized) linear and
multilevel/hierarchical models can be specified by providing prior probability distributions for each of the model parameters.
Chapter 15 on generalized additive models generalize the GLM and linear modeling frameworks by replacement of a linear combination of predictors with a linear
combination of semi-parametric transformations of predictors.
Finally, a number of machine learning techniques can be viewed as extensions of
the linear model. For example the multivariate adaptive regression spline (MARS)
algorithm is a computationally intensive technique that searches through a highdimensional space of transformed variables and variable interactions. (Hastie, Tibshirani, and Friedman 2009). Other popular machine-learning techniques include classification and regression trees (CART), random forests, and support vector machines,
(Hastie et al. 2009). It can be helpful to use machine learning techniques in the
course of (generalized) linear modeling projects for at least two reasons. First, such
techniques can be used as heuristic tools for suggesting certain variables, variable
transformations, and variable interactions for further investigation. Second, highly
complex yet accurate black box models provided by machine-learning algorithms
can serve as baselines against which to compare the predictive accuracy of more
traditional models.
Many of the topics covered in later chapters in this volume, such as credibility,
survival models, and time series models, can be viewed within the linear model
framework.
2.5 Exercises
The first four exercises use the MEPS data (MEPS.Diabetes.csv) that were analyzed
in Sections 2.2 and 2.3 of this chapter.
2.5 Exercises
61
Exercise 2.1. This exercise uses the training data to explore aspects of correlation,
addition of variables, and interpretation of coefficients in a linear model.
(a) Regress LOGEXPEND on LOGINCOME. From this output, calculate the standard deviation of LOGEXPEND on LOGINCOME. Using the regression output and the sample
correlation shown in Figure 2.9, verify the formula introduced in Example 2.5 between
the regression coefficient estimate and the sample correlation. How do you interpret
this regression coefficient? What is the relationship between the R 2 of this model and
the sample correlation coefficient between LOGEXPEND and LOGINCOME?
(b) Regress LOGEXPEND on LOGINCOME and MENTHEALTH and perform an F -test to
evaluate the significance of MENTHEALTH. Does the result indicate whether or not to
include MENTHEALTH in the model?
(c) How would you justify the difference in the value of the regression coefficient for
LOGINCOME between models in 2.1(a) and 2.1(b)?
Exercise 2.2. Use the model developed in Exercise 2.1(b) to explore the residuals.
(a) To gain insight into whether AGE might be a significant addition to the model, create
a scatterplot with the standardized residuals plotted on the y-axis and AGE plotted on
the x-axis. Add (i) a horizontal line where the residuals are equal to zero and (ii) a
regression line or a loess curve to help visualize the patterns. What does this residual
analysis suggest?
(b) Create a new model with AGE as the dependent variable and LOGINCOME and
MENTHEALTH as the explanatory variables. Calculate the residuals from this model.
Create a scatterplot of the residuals from the model created in Exercise 2.1(b) on the
y-axis and the residuals from this new model on the x-axis. The resulting graph is
called an added variable plot.
(c) Calculate the sample correlation between the residuals from the models used in
the scatterplot in 2.2(b). This quantity is known as the partial correlation between
LOGEXPEND and AGE, controlling for LOGINCOME and MENTHEALTH.
(d) Regress LOGEXPEND on LOGINCOME, MENTHEALTH, and AGE. Using the t-statistic
for AGE and degrees of freedom for the residual standard error, verify that the partial
correlation computed in 2.2(c) is directly computed by:
t(bj )
r(y, xj |x1 , , xj 1 , xj +1 , , xk ) =
2
t(bj ) + n (k + 1)
Exercise 2.3. To perform further validation of the final case study Model CS3, we
reverse the roles of the training and hold-out data samples to provide insight into the
stability of the estimated model coefficients.
(a) Refit the final case study Model CS3 on the hold-out data and compare the resulting
model coefficients with the coefficients in Model CS3.
62
(b) Use the resulting regression model estimates to compute the predictions on the training
data, and compute the sample correlation coefficient between these predictions and
LOGEXPEND.
Exercise 2.4. As an alternative to the final case study Model CS3, this exercise
examines the use of the derived variable, COMORB.CNT, defined as the sum of the
various comorbidity indicators in the model.
(a) The output of Model CS3 indicates that the coefficients for EMPHYSEMA, CHOLEST,
CANCER, ASTHMA, HIGHBP are of similar magnitude. Modify the model in the
following way: (i) drop the comorbidity indicators EMPHYSEMA, CHOLEST, CANCER, ASTHMA, and HIGHBP and (ii) add COMORB.CNT as an explanatory variable.
Label the resulting Model CS4. Justify the use of these covariates and explain the
results.
(b) Is it meaningful to perform an F -test to compare Models CS3 and CS4? Why or why
not?
(c) Compare the AIC of both models. What conclusion can be drawn as to the choice of
model?
(d) Create a scatterplot matrix to compare the predicted values of the two models on the
hold-out data.
Exercise 2.5. This exercises highlights the dangers of data snooping using simulation.
(a) Generate samples of length 1,000 from the independent standard normal random
variables Y, X1 , . . . , X100 . Next, regress Y on X1 , . . . , X100 .
(b) Given that Y is generated independently from the Xi , how many explanatory variables
would you expect to find significant at the 5% level? Compare your answer to the
output of the model.
Exercise 2.6. This exercise uses simulation to investigate various strategies for handling missing data.
(a) Simulate 1,000 draws for x1 and x2 from a bivariate normal distribution with 1 =
2 = 0, 1 = 2 = 1, and = 0.7. Hint: In R, install the mvtnorm package and use
the following line of code:
x <- rmvnorm(1000,mean=c(0,0),sigma=matrix(c(1,.7,.7,1),
ncol=2))
References
63
(c) Conduct audit checks on the data to verify that the simulated data conform to your
specifications. For example, for the vector of errors, see that the mean equals 0, the
variance equals 1, and the histogram looks normally distributed. For the multivariate
distribution, you can assess the marginals and check on the correlation between x1
and x2 .
(d) Create a scatterplot matrix of the resulting dataset and examine the relationship
between the outcome variable, the explanatory variables, and the error term.
(e) Regress y on x1 and x2 and compare the resulting coefficients with the true parameter
values.
(f) Replace a random 25% of the values of x2 with the missing value NA. Rerun the
regression in 2.6(e) and compare the resulting coefficients with the true parameter
values. Note the number of observations used in estimating the model.
(g) Replace the missing values of x2 with the mean value of the nonmissing values.
Label the resulting variable x2.imp. Regress y on x1 and x2.imp. Note that all 1,000
observations are used to fit this model. Note that the parameter estimates are biased
(i.e., differ from truth) and that the standard errors are larger than in the previous
models.
(h) Now create a binary variable that takes on the value 1 if an observation was meanimputed, and 0 otherwise. Add this binary variable to the model created in 2.6(g) and
create a new model.
(i) Compare the resulting coefficients of all models to the true parameter values. Create
a scatterplot matrix of the modified dataset to help understand the changes in model
coefficients. Note the large number of observations (25% in this case) that are zero.
(j) Add an interaction term between x1 and the mean-imputation indicator variable created above to create a new model. Compare the resulting coefficients with the true
parameter values.
(k) Finally, change = 0.7 to = 0 and repeat steps 2.6(a) to 2.6(j) of this exercise.
Note the effects of the mean-imputation strategy in the scenario when x1 and x2 are
uncorrelated.
References
Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on
Automatic Control 19(6), 716723.
American Diabetes Association (2013a). Diabetes information. http://www.diabetes.
org/diabetes-basics/?loc=GlobalNavDB.
American Diabetes Association (2013b). Diabetes information. http://www.diabetes.
org/living-with-diabetes/complications/.
Brien, R. (2007). A caution regarding rules of thumb for variance inflation factors. Quality &
Quantity 41, 673690.
Centers with Disease Control and Prevention (2013). Body mass index information.
http://www.cdc.gov/healthyweight/assessing/bmi/adult_bmi/
index.html.
64
Dall, T. M., S. E. Mann, Y. Zhang, J. Martin, Y. Chen, and P. Hogan (2008). Economic costs
of diabetes in the U.S. in 2007. Diabetes Care 31(3), 120.
Draper, N. R. and H. S. Smith (1998). Applied Regression Analysis. Wiley, New York.
Freedman, D. A. (2005). Statistical Models: Theory and Practice. Cambridge University
Press, New York.
Frees, E. W. (2010). Regression Modeling with Actuarial and Financial Applications.
Cambridge University Press, New York.
Gelman, A. and J. Hill (2008). Data Analysis Using Regression and Multilevel/Hierarchical
Modeling. Cambridge University Press, New York.
Hastie, T., R. Tibshirani, and J. Friedman (2009). The Elements of Statistical Learning: Data
Mining, Inference and Prediction (2nd edition). Springer-Verlag, New York.
Little, R. and D. Rubin (2002). Statistical Analysis with Missing Data, Second Edition.
Wiley, New Jersey.
Medical Expenditure Panel Survey (2013). Meps website. www.meps.ahrq.gov/.
Rubin, D. (1976). Inference and missing data. Biometrika 63(3), 581592.
Stigler, S. M. (1986). The History of Statistics: The Measurement of Uncertainty before
1900. Belknap Press, Cambridge, MA.
R Core Team (2013). R: A Language and Environment for Statistical Computing.
R Foundation for Statistical Computing, Vienna, Austria. http://www.R-project.
org.
Tukey, J. (1977). Exploratory Data Analysis. Addison-Wesley, Reading, Massachusetts.
3
Regression with Categorical Dependent Variables
Montserrat Guillen
Chapter Preview. This chapter presents regression models where the dependent variable is categorical, whereas covariates can either be categorical or continuous. In the
first part binary dependent variable models are presented, and the second part is aimed
at covering general categorical dependent variable models, where the dependent variable has more than two outcomes. This chapter is illustrated with datasets, inspired
by real-life situations. It also provides the corresponding R programs for estimation,
which are based on R packages glm and mlogit. The same output can be obtained
when using SAS or similar software programs for estimating the models presented in
this chapter.
3.1 Coding Categorical Variables
Categorical variables measure qualitative traits; in other words, they evaluate concepts
that can be expressed in words. Table 3.1 presents examples of variables that are
measured in a categorical scale and are often found in insurance companies databases.
These variables are also called risk factors when they denote characteristics that are
associated with losses.
Categorical variables must have mutually exclusive outcomes. The number of
categories is the number of possible response levels. For example, if we focus on
insurance policies, we can have a variable such as TYPE OF POLICY CHOSEN with
as many categories as the number of possible choices for the contracts offered to the
customer. This categorical variable is valid if the customer can choose one and only
one option. An introduction to categorical variables was given in Chapter 2.
A categorical variable with many categories can be understood as a set of binary
variables that capture the presence or absence of each possible choice. For example,
we can work with the categorical variable TYPE OF VEHICLE. In Table 3.1 ths a
variable has four possible categories (motorbike, car, van, or truck). Equivalently, we
can define four binary variables. The first one indicates whether or not the vehicle is
65
66
Categories
Sex
Bodily injury coverage
Deductible
Policy renewal status
Opinion
Size
Type of vehicle
Number of policies owned by the same customer
Man or woman
Yes or no
Present or absent
Renewed or canceled
Against, neutral, or favorable
Small, medium, or large
Motorbike, car, van, or truck
1, 2, at least 3
a motorbike, the second one indicates whether or not the vehicle is a car, and so on.
For simplification, when using a binary variable we usually denote the presence by
1 and the absence by 0.
When a categorical variable measures choices, then we say that each category
is a possible response. Choices can be unordered, as with sex, or ordered, such as a
small, medium, or large type of answer. Ordered choices can be modeled using special
purpose models, but they can also be treated as unordered responses, if we just ignore
the order.
It is common to call the dependent categorical variable itself the response. If the
response has two possible outcomes, then one outcome is called the event and the
other one is called the baseline, the reference, or the non-event. When using statistical
software, one has to be careful about the way the categories are labeled by internal
procedures, which fix the way models are specified by default. Before starting the
model estimation, one has to be sure about the way categorical responses are handled,
and particularly which is the response that corresponds to the baseline choice.
The baseline category is often assigned by practitioners to the one that is most
frequent in the database or in the population, or to the one that indicates the nonevent. This is done to facilitate the interpretation of coefficients. It seems more natural
to speak about an event occurrence, rather than having to talk about it not ocurring
having too many negative words in a sentence makes it much more complex, and
communicating ideas becomes cumbersome.
3.2 Modeling a Binary Response
Databases contain information on n individuals, who are often also called units, cases,
or observations and who correspond, in the insurance context, to policyholders, firms,
business units, agents, and so on. Let us assume that yi is the categorical response for
individual i. In the simplest case, yi is binary and can only take two possible values,
which we code as 1 for the event, and 0 for the non-event, respectively. We assume yi
follows a Bernoulli distribution with probability i .
67
i = 1, . . . , n.
exp(xi )
1
=
,
1 + exp(xi )
1 + exp(xi )
(3.1)
0.0
Distribution Function
0.2
0.4
0.6
0.8
1.0
68
0
x
and the inverse of this relationship, called the link function in generalized linear
models, expresses xi as a function of i as
i
= logit(i ).
xi = ln
(3.2)
1 i
We call i /(1 i ), the odds and its natural logarithm is the log-odds or the logit.
So, because it clearly follows from (3.2) in the logistic regression model, we say the
log-odds is a linear combination of the predictors.
In the logistic regression model, parameters are interpreted in terms of log-odds
rather than directly on the response. Moreover, the predicted response is the probability
of the event. To put it in simple words, the logistic regression is a model to predict a
probability.
The logistic regression model is popular because the fitted responses lie between 0
and 1, and therefore, the fitted values can always be interpreted easily as the estimated
probability for the modeled event, namely yi = 1. The logistic function is defined
as f (x) = 1/(1 + exp (x)), and it is represented in Figure 3.1. The S-shaped curve
that results from the combination of risk factors, or covariates, is appealing because
in the extremes the effect of the linear combination of risk factors on the estimated
probability smooths down.
The logistic regression model is widely used in insurance, as well as in many
other disciplines. In medicine or pharmaceutics, this model is the basis for analysis
of the cause effects of risk factors or conditions that influence the positive or negative
response of a patient to a treatment.
Parameters are interpreted in the binary logistic regression model setting in a more
complicated way than in the linear regression model context. A bit of practice is
needed to get used to talking about fitted probabilities instead of fitted values.
69
Pr(yi = 1|xi , xik = c)
ln
.
1 Pr(yi = 1|xi , xik = c)
If K is the total number of regressors in the model, including the constant term, then
odds-ratio ORk equals exp(k ), for k = 1, . . . , K.
We can say that the odds for xik = c + 1 are exp(k ) times the odds when xik = c.
If the k-th explanatory variable is continuous, we can also derive that parameter k is
the proportional change in the logarithm of the odds-ratio. This is connected to the
notion of the economic elasticity.
A connected concept to the odds-ratio is the notion of risk-ratio. The risk-ratio is
the ratio of event probabilities when the predictor indicator changes by one unit. It is
especially important in indicating the relative change in the probability of an event
when a risk indicator is present rather than absent. The definition is
RRk =
k = 1, . . . , K.
70
Table 3.2. Number of Policyholders who Choose Other Coverage or Full Coverage
as a Function of Predictors
Other Coverage
y=0
Full Coverage
y=1
Total
woman
man
498(50.30%)
2115(70.27%)
492(49.70%)
895(29.73%)
990
3010
rural
urban
1906(72.83%)
707(51.12%)
711(27.17%)
676(48.88%)
2617
1383
commercial
33(84.62%)
private
2580(65.14%)
MARITAL STATUS
single
467(54.24%)
married
2047(68.85%)
other
99(59.64%)
AGE (YEARS)
48.27
SENIORITY IN COMPANY (YEARS)
9.93
6(15.38%)
1381(34.86%)
39
3961
SEX
DRIVING AREA
VEHICLE USE
394(45.76%)
926(31.15%)
67(40.36%)
43.09
10.88
861
2973
166
46.47
10.88
Notes: Row percents indicate the proportion of policyholders. For quantitative regressors, the
mean is shown.
variable is whether they bought full coverage (event coded as 1) or not (non-event
coded as 0). For this example, full coverage usually refers to a type of contract that
compensates for any physical damage to the insured vehicle, including collision damage as a result of an accident; it also covers stolen on vandalized vehicles as well as
storm or flood damage or an impact with an inanimate object.
Predictors for coverage choice are SEX, DRIVING AREA, VEHICLE USE,
MARITAL STATUS of the policyholder, AGE, and SENIORITY in the company.
Table 3.3 presents the parameter estimates for the logistic regression model to
predict the probability that a policy holder chooses FULL COVERAGE. Parameter
estimates are obtained by the maximum likelihood method. The logistic regression is
a special case of the generalized linear model, and maximum likelihood estimation in
this context is covered in Chapter 5.
We can see that the parameter associated with variable AGE is negative (0.058)
and significantly different from zero (small p-value). This result means that the older
the policyholder is, the lower are the chances he or she buys full coverage. Likewise,
we expect that men would be less prone to buying full coverage than women. The
parameter for MEN is negative (0.961) and significantly different from zero.
Two additional interesting conclusions can be found in our dataset from the model
estimation results. First, customers driving in an urban area have a higher probability
71
Standard
Error
p-Value
Odds-Ratio
(INTERCEPT)
MEN
URBAN
PRIVATE
MARITAL(MARRIED)
MARITAL(OTHER)
AGE
SENIORITY
0.257
0.961
1.173
1.065
0.083
0.161
0.058
0.133
0.486
0.086
0.078
0.469
0.096
0.200
0.004
0.007
0.5959
<0.001
<0.001
0.0232
0.3889
0.4212
<0.001
<0.001
0.773
0.382
3.230
2.901
0.921
1.175
0.944
1.143
2Log-Likelihood
Likelihood ratio test
Pseudo-R 2
2 (2143.905)
875.483 (df = 7, p-value<0.001)
16.96%
Variable
Notes: Dependent variable is the choice of FULL COVERAGE (y = 1) versus OTHER COVERAGE (y = 0).
of buying full coverage compared to those who drive in rural areas. This is because
the parameter for URBAN is positive (1.173) and significant (p-value lower than 1%).
Second, the SENIORITY in the company has a positive influence. Its parameter is
positive (0.133), and the corresponding p-value is below 1%. This means that those
customers who have been in the company for a longer period of time have a higher
probability of buying full coverage than recent or new customers. Marital status of the
policyholder, which separates single, married, and other individuals, does not seem
to have a clear effect on the coverage choice. If we look at the significance in the
p-value column, we see that the p-value for the coefficient of variables corresponding
to marital status equal to MARRIED or OTHER as opposed to SINGLES is large. So,
those coefficients are not significantly different from zero. A zero parameter implies
that the predictor has no influence on the response. Likewise, the p-value for variable
PRIVATE is above 2%. Vehicles for private use are compared to those that have
commercial uses, such as those used for transportation or for business. This factor
has no significant effect on the choice of coverage if we consider significance at 1%,
but it has a significant effect if we fix significance at the 5% level. At this level, the
effect is significant and policyholders who own a private use vehicle have a larger
probability of buying FULL COVERAGE than insureds with a vehicle for commercial
use.
In Table 3.2, we can look at the odds-ratio column and say that when the predictor
increases by one unit, the odds that the insured chooses FULL COVERAGE are multiplied by the odds-ratio value. So, when the odds-ratio is larger than one the odds
72
of FULL COVERAGE versus OTHER COVERAGE increase and they diminish if the
odds-ratio is smaller than one.
n
i=1
73
The method is comparable to the way we would climb a mountain. The likelihood
function would play the role of the mountain to be climbed. A walkers steps lead
us to the summit. The latitude and longitude of terrestrial coordinates would play the
role of the parameter values that are updated at each step until the summit is reached.
Similarly, in the likelihood maximization, we modify the values of the parameters
to locate the summit of the likelihood function. Starting at convenient values for the
parameter vector, this method calculates a new vector of parameter estimates in such
a way that the likelihood function is larger than its initial value obtained from the
initial parameter values. The procedure is repeated until the likelihood evaluated at
new parameter values stabilizes and does not increase when updating the parameter
values. Estimation of generalized linear models is further developed in Chapter 5.
3.3.3 Model Fitting
Once they are obtained, we can predict the
The parameter estimates are denoted by .
probability of the response modeled event. Given the individual characteristics, the
fitted probability is based on the following expression:
i =
exp(xi )
1 + exp(xi )
So, every individual in the dataset has an observed response event, yi , and we
can predict the fitted probability of the modeled event, i . Let us compare those two
values. For instance, an individual in the dataset may have yi equal to 1, and the
estimated probability i may be 0.9 = 90%. That is quite plausible. But the response
can be yi = 1, while the prediction can be much lower, say 20%. In this case the
model prediction would be surprising. In our example, consider a policyholder who
has bought the full coverage, but has an estimated probability of buying it of only
20%. We would not have expected this policyholder to buy this coverage, according
to the fitted probability. We can say that the odds that he buys full coverage are
1 in 4.
In most datasets when predictions and observations are compared, we can find
an unexpected low fitted probability for the event observed responses. The opposite
situation can also occur, when the response is the non-event, whereas the probability
of the event is large.
In linear models, it is sensible to discuss the correlation between an observation and
its fitted value. As pointed out in Chapter 2, squaring this correlation gives the statistic
R 2 , the coefficient of determination. However, in nonlinear models, correlations are
less useful.
In practice, we will be happy if our logistic regression model is a good predictive
tool for the binary response outcome. This means that we will prefer a model in which
74
the predicted probability that is assigned to a case in the original database is coherent
with the observed response. So, when the observed response for individual i is the
event, then we would like our model to predict a large i for individual i. Likewise,
when the observed response for a particular individual is a non-event, then we want
that i for individual i to be small, which means that the probability of a non-event,
namely (1 i ), is high.
When looking at i , we do not really know what we should consider large or small.
In general we fix the discrimination threshold at 50%. That corresponds to odds equal
to one. So, when the estimated probability is larger than 50%, then we would expect
that the response yi is the event, and similarly, when the estimated probability of the
event is smaller than 50%, then we would expect the response to be the non-event.
The discrimination threshold can vary and should not necessarily be equal to 50%.
In fact, it is recommended that the threshold be set equal to the proportion of events in
the dataset to keep the proportion of predicted event cases in line with the observations.
Using the proportion of events in the dataset as the discrimination threshold generally
maximizes the overall prediction performance.
The classification table is the simplest way to present the prediction performance
of a logistic regression model.
Example 3.1 (continued): An example of a classification table is presented in
Table 3.4. We set the threshold at 35%, which is the proportion of full coverage
choice in the sample in the example. Columns represent the predicted responses, and
rows are observed behavior. Each cell in the table counts the number of observations
from the dataset that are classified within each group after the model and threshold
have been used to predict the response behavior. Note that a classification can be done
using an out-of-sample new dataset that is not used in the estimation phase. In our
example, we reclassified the same observations that were used for estimation.
We see that 2,982 cases are correctly classified, as the model predicts the correct
response for 1,858 policyholders who do not have FULL COVERAGE and 1,074 who
have FULL COVERAGE. This means that the overall classification rate is equal to
2,982/4,000=74.55%, which is excellent. However, a number of cases do not have
the expected response. On the one hand, there are 755 cases that do not choose FULL
COVERAGE, but the fitted probability is high and the model predicts that they are
likely to have made this choice. On the other hand, there are 313 cases that have FULL
COVERAGE in the initial database, but the model predicts the opposite response for
them.
There are typical traits from the classification table that we would like to see in
practice when using a logistic regression model for predictive purposes. The number
of cases that are correctly classified has to be high, whereas the number of cases that
are incorrectly classified should be low.
75
Observed y = 0
Observed y = 1
Predicted y = 0
Predicted y = 1
1,858
313
755
1,074
76
so there are less parameters than in the original model. The number of degrees of
freedom for the test statistic is the number of parameters that have been eliminated in
the reduced model compared to the initial one. In the null hypotheses, the parameters
that have been eliminated are zero. In the alternative hypothesis at least one of those
parameters differs from zero. When the sample is large, the likelihood ratio test
statistic follows a 2 distribution.
The expression for the likelihood ratio test statistic is
LRT = 2 (LA LR ),
where LA is the maximized log-likelihood function in the large, initial model and LR
is the maximized log-likelihood function in the reduced model.
In practice an abuse of goodness-of-fit statistics may lead to selecting a logistic
regression model that is difficult to understand. As practitioners, we may be tempted
to start from a model with all possible explanatory variables that can be predictors
of the dependent variable in the dataset and to eliminate variables step by step, using
the likelihood ratio test. This is called a backward elimination procedure. Each time
a variable or a group of variables is eliminated. The likelihood ratio test indicates that
those variables can be deleted from the model if the alternative hypothesis is rejected.
It is highly recommended to be aware of the variables that are being used in the
model specification and their corresponding meaning. Backward elimination may
lead to selecting variables no matter what they mean or whether they are precise. For
instance, an insurer may be confident about the information in the dataset regarding
the driving experience of policyholders. A copy of the driving license can always be
checked. However, the policyholders driving area is a predictor that may be kept in the
model before dropping the driving experience variable or even replace it. Note that
unless a device is installed in the car, it can be difficult to determine what is the
true driving area for a customer. One would prefer driving experience rather than
driving area as a predictor for models in insurance databases, but automatic backward
elimination does not consider the practitioners preferences for predictors or their
information quality.
Pseudo-R squared is a popular goodness-of-fit measure for logistic regression
models, but there are many expressions available in the literature. The simplest version
can take values between zero and one and is easy to interpret. The larger it is, the
better the model fit. The pseudo-R 2 statistic is defined as
LA
2
,
Pseudo R = 1
L0
where LA is the maximized log-likelihood function in the initial model and L0 is
the maximized log-likelihood function in the model with only a constant term. Other
77
Expected
1
2
3
4
5
6
7
8
9
10
22
16
30
31
128
225
219
245
237
234
20.83
44.45
67.28
90.20
114.98
140.23
166.60
198.92
241.17
302.33
X2
289.4653
Df
8
P(>Chi)
<0.001
Group
commonly used statistics are variations of pseudo-R 2 that are defined between 0
and 1.
The Hosmer-Lemeshow test (see Table 3.5) is aimed at evaluating the coherence
between the observed responses and the predicted probabilities. The fitted probabilities obtained from the model estimation are ordered from smallest to highest, and
individuals are grouped in decile groups. Then 10 groups with the same number of
individuals in each group are available.
The Hosmer-Lemeshow test can be calculated as
G
(Og Eg )2
H =
,
n (1 g )
g=1 g g
where Og is the number of cases with the event in group g, Eg is the expected number
of choices in that group, ng is the number of cases in group g, and g is the average
probability predicted by the model in that group.
There is a graphical tool to assess goodness of fit in logistic regression modeling: the
receiver operating characteristic, also known as the ROC curve. The ROC curve is a
graphical plot of the sensitivity, or true positive rate, versus the false-positive rate (one
minus the specificity or true negative rate), for the estimated logistic regression model.
Each point in the curve corresponds to a discrimination threshold level. The best model
fit would be for a model that has a ROC curve as close as possible to the upper left
edge. A nondiscriminant model would have a flat ROC curve, close to the diagonal.
78
ROC analysis provides a way to select possibly optimal models and to discard
suboptimal ones based on classification performance at various threshold levels. For
an objective rule to compare ROC curves, the area under the curve, simply referred
to as AUROC, can be used. A model with the largest AUROC should be preferred to
other possible models.
Example 3.1 (continued): To compute the Hosmer-Lemeshow test in our sample
of 4,000 individuals, we can define groups of 400 individuals. The first group is
formed by the individuals in the dataset who have the lowest estimated probabilities
of the modeled response, as predicted by the model. The second group is formed
by the next bunch of individuals whose estimated probabilities are low, but not as
low as the first group, and so forth. Inside each group the number of event responses
is calculated. So, for example in our case study, we would expect that in the first
group, where the estimated probabilities of buying full coverage are the lowest in
the dataset, the proportion of individuals who bought full coverage in this group
is indeed low. Likewise, in the highest group we would expect the contrary that
most customers would buy full coverage so there the observed proportion would be
high. The Hosmer-Lemeshow test evaluates how close the proportions of observed
event responses within each group are compared to the expected probability for the
modeled event in that group. A model fits perfectly if the observed proportions and
the predicted proportions are equal. The test establishes a distance between observed
and expected proportions. The closer they are, the better is the model fit.
The results for the Hosmer-Lemeshow test are presented in Table 3.5. The test
indicates that the model has a poor fit because the p-value is too low. In large samples,
this test frequently rejects the null hypothesis.
An example is shown in Figure 3.2 of the ROC curve. In this plot, the curve is
distant from the diagonal, which is a good sign. The area under the curve (AUROC)
equals 78.78%, which also indicates a good fit.
79
0.0
0.2
sensitivity
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
1specificity
0.8
1.0
Fig. 3.2. The ROC curve for the logistic regression model in the FullCoverage.
csv dataset. Dependent variable is the choice of FULL COVERAGE (y = 1) versus
OTHER COVERAGE (y = 0).
as following a normal distribution and be set equal to xi plus an error term. In that
case the probit model follows directly as a result of assuming that an individual makes
the decision to chose the event response when the random unobserved utility exceeds
a threshold that depends on xi . So, the probit model arises naturally from a linear
regression model for the underlying unobserved utility.
Other binary regression models are possible, such as the complementary log-log
model. In terms of prediction performance, these other models do not seem to be more
flexible or easily interpretable as the logistic regression model.
80
Standard
Error
p-Value
Odds-Ratio
INTERCEPT(COMPREHENSIVE)
INTERCEPT(FULL COVERAGE)
MEN
URBAN
PRIVATE
MARITAL(MARRIED)
MARITAL(OTHER)
AGE
SENIORITY
1.754
3.094
0.067
0.151
1.478
0.083
0.161
0.001
0.018
0.578
0.580
0.097
0.086
0.562
0.098
0.705
0.004
0.007
0.002
<0.001
0.492
0.079
0.009
0.213
0.207
0.917
0.008
0.173
0.045
1.069
1.163
4.380
0.885
2.433
1.000
1.018
2Log-Likelihood
Likelihood ratio test
4330.819
25.156 (df = 7, p-value = 0.001)
Variable
Pr(yi j |xi )
.
logit(Pr(yi j |xi )) = ln
1 Pr(yi j |xi )
Parameters are interpreted similar to the logistic regression model case. A positive parameter means that if the predictor increases then the cumulated probability
increases. There are as many constant terms as possible choices minus one, because
Pr(yi J |xi ) = 1. Those constant term parameters (j ) are increasing with j because
the cumulative probabilities also increase when more choices are accumulated, but
they can be negative.
Example 3.2 (TypeofCoverage.csv data file). Table 3.6 presents the results of a
sample dataset in which are 2,148 policy holders who are offered three ordered
types of coverage: THIRD PARTY LIABILITY, COMPREHENSIVE, and FULL
COVERAGE. Third-party liability is the minimum compulsory insurance in the market
for those customers. Comprehensive insurance includes additional coverage for stolen
or vandalized vehicles. Full coverage also includes damage to the insured vehicle. We
interpret the parameters and the odds-ratio as before. Here we see that the effect of
predictors is not as neat as in the logistic regression model presented in Example 3.1.
81
Age does not influence the propensity to increase insurance coverage, nor does sex
or marital status; p-values are too large. Private use vehicles do seem to be associated
with higher levels of coverage. Negative signs in the intercept coefficients mean a
resistance to increase coverage, which is certainly explained by the associated price
increase, but price is not included in this model. Only larger seniority in the company
and private as opposed to commercial vehicles indicate a tendency to select extended
(cumulated) coverages.
3.6 Models for Nominal Categorical Dependent Variables
The last section is aimed at specifying the multinomial logit and general categorical
response models, with an emphasis on parameter interpretation.
In nominal categorical dependent variable models, the responses correspond to
choices that are identified with labels and do not respond to a natural ordering. A
classical example is the color chosen when buying a car. Another one is the type of
policy chosen by policyholders in an insurance company. An insurer may offer several
possibilities to the customer, and although we could say that the price of each possible
contract may induce an ordering in the choice, we can ignore price, and study the
choice of products as purely nominal alternatives. A marketing firm could also study
customer behavior by offering the customer a choice of possible insurers to model the
preference for a particular insurance company over some others.
The multinomial logit model (Section 3.6.1) and the conditional logit model (Section 3.6.2) deal with data where the dependent variable contains more than just two
alternatives. The former handles only explanatory variables that are individual specific,
whereas the latter handles explanatory variables that are individual and alternative
specific. Many applications of both types exist in the insurance and risk management
context. For example, if we have three possible insurance policy coverages to choose
from, a model that would just consider the consumers characteristics would only
include individual-specific explanatory variables, whereas a model that would consider, in addition, the differences in quality or price between each coverage would
include individual- and alternative-specific characteristics.
Let us assume that a nominal categorical variable has J possible choices. We recommend using the models in this section with a moderate number of possible choices.
The larger the number of options, the more complex is parameter interpretation.
3.6.1 Specification of the Generalized Logit Model
Let us start with the generalized logit model. This model is often called the multinomial logit model, which we present in the next section and which is a bit more general.
However, the generalized logit model is so widely used that it is often called the
multinomial logit model. The generalized multinomial logit model is used to predict
82
exp(xi j )
, j = 1, . . . , J 1
1
1 + Jk=1
exp(xi k )
Pr(y1 = J |xi ) =
1
.
J 1
1 + k=1 exp(xi k )
83
Parameter Estimate
Standard Error
p-Value
2.774
0.468
0.804
0.006
0.412
0.272
0.267
0.008
<0.001
0.086
0.003
0.470
1.380
0.644
0.071
0.024
0.272
0.178
0.145
0.005
<0.001
<0.001
0.625
<0.001
2 (1104.4)
44.051(df = 6, p-value < 0.001)
Notes: Dependent variable is type of vehicle chosen: MOTORBIKE, FOUR WHEEL DRIVE, and
CAR (baseline).
84
each choice is associated with a time and cost. This model is also known as the conditional logit model because individual characteristics depend on the chosen alternative.
For instance, when underwriting a policy, the price depends on the policyholders risk
factors and on the type of policy or brand that is chosen.
The multinomial logistic regression model specification is
exp(xij )
, j = 1, . . . , J.
Pr(y1 = j |xij ) = J
k=1 exp(xik )
(3.3)
There is only one vector of unknown parameters , but we have J vectors of known
characteristics xi1 , xi2 , . . . , xiJ . For instance, if we offer three insurance policies to a
customer, we would have three different prices, even if the customer finally chooses
one of them. We can then use that price information in the model as a factor that
changes with choice. The regressors can also contain individual characteristics that
do not change with the choice, such as age or driving experience.
The generalized logit model can be understood as a special case of the multinomial
logistic model. If we define xij = (0 , . . . , 0 , xi , 0 , . . . , 0 ), then it is a vector that
contains zeros, Except for the components corresponding to the j -th alternative. Then
we define = (1 , . . . , J ). Because of this special case definition, many authors refer
to the multinomial logit model when they mean to refer to the generalized logit model.
Example 3.4 (VehChoicePrice.csv data file). This example corresponds to a similar
situation to Example 3.3. Customers report the type of vehicle that they choose; however additional information is now available on the price of each possible alternative.
Price is an alternative specific explanatory variable.
Estimation results for a multinomial logistic regression model are shown in
Table 3.8. Not surprisingly, the variable PRICE has a positive (0.056) and significant
effect, which means that customers tend to choose the most expensive alternative.
Indeed, given the nature of the alternatives, higher price vehicles correspond to those
that have more capabilities.
Noticeably, those driving in urban areas tend to find the choice of a four-wheel
drive vehicle less attractive than a regular car, compared to those living in rural areas
(parameter equal to 0.767 and significant). Conclusions similar to those reported for
the generalized logit model, where the alternative specific price was not included in
the model, are also found here. This model found a larger preference for motorbikes
in men compared to women and in younger compared to older respondents.
85
0.411
3.493
0.056
0.331
0.713
0.767
0.082
0.001
0.033
2Log-Likelihood
Likelihood ratio test
2 (1055.8)
141.110 (df = 7, p-value <0.001 )
0.288
0.434
0.006
0.276
0.180
0.270
0.146
0.008
0.006
p-Value
0.153
<0.001
<0.001
0.229
<0.001
0.004
0.574
0.884
<0.001
Notes: Dependent variable is type of vehicle chosen: Motorbike, Four wheel drive and Car
(baseline).
made for the first step, a second choice is possible. This is a flexible setting because
the number and type of possible choices do not have to be the same in each branch.
Moreover, the IIA hypothesis is eliminated in this model.
It is often the case that we have repeated observations for the same individuals.
The mixed logit model can take this panel dimension into account. Panel data models
involving categorical dependent variables are aimed at modeling a sequence of categorical responses. They can be presented as nested logit models. Mixed logit models
assume that parameter vectors are subject specific. This means that the parameters
change from one individual to another. Estimation for this type of models is done
using simulation methods.
3.7 Further Reading
The conditional logit model was introduced by McFadden (1974) in the context of
econometrics. Daniel McFadden and James Heckman were awarded the Nobel Prize
in 2000 for their contributions to transportation economics. Hosmer and Lemeshow
(2000) wrote a book on applied logistic regression in which they present their test.
Long (1997) is a great resource for categorical and limited dependent variables. Hilbe
(2009) is full of guided examples in Stata, and Frees (2010) presents several actuarial
and financial applications of categorical dependent variable models and includes an
excellent summary of the theoretical background. Additional reference material can
be found in Greene (2011), Cameron and Trivedi (2005), and Artis, Ayuso and Guillen
(1999).
86
References
Artis, M., Ayuso, M., and Guillen, M. (1999). Modelling Different Types of Automobile
Insurance Fraud Behaviour in the Spanish Market, 24(1), 6781.
Cameron, A. C. and P. K. Trivedi (2005). Microeconometrics: Methods and Applications.
Cambridge University Press, New York.
Frees, E. W. (2010). Regression Modeling with Actuarial and Financial Applications.
Cambridge University Press, New York.
Greene, W. H. (2011). Econometric Analysis (7th ed.). Prentice Hall, New York.
Hilbe, J. M. (2009). Logistic Regression Models. CRC Press, Chapman & Hall. Boca Raton,
FL.
Hosmer, D. W. and S. Lemeshow (2000). Applied Logistic Regression (2nd ed.). John Wiley
& Sons, New York.
Long, J. S. (1997). Regression Models of Categorical and Limited Dependent Variables.
Sage, Thousand Oaks, CA.
McFadden, D. (1974). The Measurement of Urban Travel Demand. Journal of Public
Economics, 3(4), 303328.
4
Regression with Count-Dependent Variables
Jean-Philippe Boucher
Chapter Preview. This chapter presents regression models where the random variable
is a count and compares different risk classification models for the annual number
of claims reported to the insurer. Count regression analysis allows identification of
risk factors and prediction of the expected frequency given characteristics of the risk.
This chapter details some of the most popular models for the annual number of claims
reported to the insurer, the way the actuary should use these models for inference, and
how the models should be compared.
4.1 Introduction
In the early 20th century, before the theoretical advances in statistical sciences, a
method called the minimum bias technique was used to find the premiums that should
be offered to insureds with different risk characteristics. This techniques aim was
to find the parameters of the premiums that minimize their bias by using iterative
algorithms.
Instead of relying on these techniques that lack theoretical support, the actuarial
community now bases its methods on probability and statistical theories. Using specific probability distributions for the count and the costs of claims, the premium is
typically calculated by obtaining the conditional expectation of the number of claims
given the risk characteristics and combining it with the expected claim amount. In
this chapter, we focus on the number of claims.
4.2 Poisson Distribution
The starting point for modeling the number of claims, a random variable noted Y,
is the Poisson distribution. Introduced by Simeon-Denis Poisson (17811840) in his
1837 work titled Recherches sur la probabilite des jugements en mati`ere criminelle
et mati`ere civile, the Poisson distribution is indeed the basis of almost all analysis of
87
88
count data. The probability function of the Poisson distribution can be expressed as
Pr(Y = y) =
y e
,
y!
where is the mean parameter. The Poisson distribution has an equidispersion property, meaning that the mean E(Y ) of the distribution is equal to its variance Var(Y ).
Because the Poisson distribution is a member of the exponential family, it has some
useful statistical properties (see Chapter 5 for details).
As we see in this chapter, this distribution can be used to model the number of
claims in nonlife insurance. However, it can also be used in other popular areas in
actuarial sciences. For example, count regression can be used with run-off triangles
(see Chapter 18), to model mortality (see Chapter 19) and natality, or to estimate the
severity scores in third-party liability.
89
j
i .
i=1
Consequently, we can link the time arrival (j ) of the j th claim to a claim count
process Y (t):
(j ) t Y (t) j.
Consequently, when the time occurrence of the j th claim is less than t, the number of
claims during the time period t is greater or equal to j . Formally, we have
Pr(Y (t) = j ) = Pr(Y (t) < j + 1) Pr(Y (t) < j )
= Pr((j + 1) > t) Pr((j ) > t)
= Fj (t) Fj +1 (t),
where Fj (t) is the cumulative function of (j ).
If we suppose that the waiting time between two claims is exponentially distributed
with mean 1/, which indicates that (j ) is gamma distributed, we can show that
Pr(Y (t) = y) =
et (t)y
,
y!
90
For classic regression models (see Chapter 2), the exogenous information can be
coded with binary variables. For example, we can model the sex with a variable x that
takes the value 1 if the insured is a man and 0 otherwise.
In statistical modeling, we use the a link function h(xi ), where = (0 ,
1 , . . . , k ) is a vector of regression parameters for the binary explanatory variables
xi = (xi,0 , xi,1 , . . . , xi,k ). Usually in insurance, the link function h() is an exponential function. Consequently, we have i = ti exp(xi ) = exp(xi + ln(ti )) where ti
represents the risk exposure of insured i, as explained in the previous section.
There is a big advantage of using an exponential function for the mean parameter.
It allows the insurer to construct a premium based on multiplicative relativities:
i = ti exp(0 + 1 xi,1 + + 1 xi,k )
= ti exp(0 ) exp(1 xi,1 ) exp(1 xi,k )
= ti p0 ri,1 ri,k ,
where p0 can be viewed as the base premium, and ri,j , j = 1, . . . , k the relativities
applied to insureds having the property j (i.e xi,j = 1).
n
yi ln(exp(xi )) exp(xi ) ln(yi !) ,
i=1
where n is the total number of observations and yi the number of claims of insured i.
We then obtain the first-order condition:
l()
(yi i )xi = 0.
=
i=1
n
(4.1)
The Newton-Raphson algorithm can be used to obtain the using the equation
(4.1). However, classic statistical software such as R or SAS can be used with preprogrammed routines. The properties of the MLE are well known and allow us to
compute the variance of the estimators. The Hessian and the outer-product estimates
91
n
xi xi i
1
.
i=1
=
VarOP []
n
xi xi
(yi i )
1
2
i=1
It is well known that the Poisson distribution has some severe drawbacks, such
as its equidispersion property, that limit its use. Because the estimated parameters
of a Poisson distribution are consistent even if the real distribution is not Poisson,
it is very advantageous to use these estimates. So, we can still use obtained by
the MLE of a Poisson distribution and correct the overdispersion of the Poisson
distribution by adding a multiplicative factor to the variance of the Poisson to
obtain VarOver. [Yi ] = i . We can estimate the parameter by several techniques
including
n
(yi i )2
= i=1
,
n
i=1 i
where the estimator is computed using i , which corresponds to the estimate of the
mean parameter i . In this case, it can be shown that
= VarHess [].
VarOver. []
Based on equation (4.1), other estimation techniques can be used such as the
estimating equation approach, in which the link between the mean and the variance
is broken.
Parameters statistical significance can be tested using classic Wald or likelihood
ratio tests. Deviance, as explained in Chapter 5, can also be used to verify the fit of
the Poisson because this distribution is a member of the linear exponential family.
4.3.2 Empirical Illustration
We use the Singapore Automobile Claims Database [singapore.csv] to apply the
models. The following two covariates have been used in the application:
(1) The No Claim Discount (NCDClass). Based on the accident record, we used three
categories: 0, 1030, and 4050. A high discount means a better accident record than
a lower one.
(2) The age (in years) of the vehicle (Vagecat1). We used five categories: 02, 35,
610, 1115, and >15.
92
Parameter
0
(10, 20, 30)
(02)
(35)
(610)
(1115)
Estimate
3.7748
0.7218
0.3584
1.5896
1.5329
1.1051
0.3489
0
1
2
3
4
5
6
Hess
OP
0.5118
0.1246
0.1264
0.5035
0.5148
0.5179
0.5304
0.5151
0.1227
0.1240
0.5106
0.5222
0.5244
0.5380
1803.841
Log-likelihood
Over.
0.5233
0.1274
0.1292
0.5149
0.5265
0.5296
0.5424
= 1.0475
Results of the application of the Poisson distribution are in Table 4.1. R programs
that computed these results are available on the books website.
Large differences between the Hessian and the outer-product variances indicate a
misspecification of the model, meaning that the Poisson distribution is not indicated
for the data. In our case, the variances of the parameter estimates of are quite close.
However, obtaining a sound theoretical answer to requires construction of a more
formal test. The inclusion of a dispersion parameter increases the Hessian variance
by a factor of 1.0475.
93
( )y e
,
y!
a a a1
exp( a),
(a)
xi
i=1
y
i
i
= 0.
1 + i /a
For the a parameter, it can be shown that the first-order condition can be expressed
as
n y
i 1
i=1
j =0
i
yi i
1
ln 1 +
= 0.
j +a
a
a + i
Additionally, we can show that the and the a estimators are independent, meaning
that
a]
= 0.
CovMLE [,
By changing the parameter of the heterogeneity distribution, it is possible to obtain
other forms of the negative binomial distribution. Indeed, using
f ( ) =
1
exp( ),
( )
with = a(2p) , this general form of the negative distribution has the same mean ,
but because it has a variance of the form + p , it is called a negative distribution
p (NBp).
For p = 2, note that the variance of now depends on the individual characteristics
of the policyholders. When p = 2, the NB2 distribution coincides with the negative
94
binomial 2 distribution seen earlier. When p is set to 1, we get the classic NB1
distribution that has the following probability function:
Pr(Y = y) =
(y + a)
(1 + 1/a)a (1 + a)y ,
(y + 1)(a)
with a = 1/. The NB1 model is interesting because the variance Var[Y ] = + =
is the one used previously to correct the overdispersion of the Poisson (noted
VarOver. [Y ]).
We can even treat the variable p as an unknown parameter to be estimated. This
can be seen as an hyper-model that can allow use of a test to distinguish between the
NB2 and the NB1 distributions. For example, if the confidence interval of the variable
p includes the value 2 but not the value 1, it means that the distribution NB2 should
be preferred to the NB1 distribution.
Finally, note that the NB2 distribution is clearly different from the NB1 or, more
generally, from distributions NBp, with p = 2, in the following three ways:
(1) The NB2 supposes an independence between and a() for the estimation by maximum likelihood (the variance-covariance matrix is 0 outside the diagonal), which is
not the case for other forms of the NB distributions.
(2) For a parameter a() that is known, the NB2 is a member of the exponential linear
family; consequently the MLE of is consistent if the true distribution is not NB2.
(3) An NB2 distribution with a = 1 ( = 1) corresponds to a classic geometric distribution.
2s
Ks (u).
u
95
As for the NB distributions, the PIG distributions can be extended to cover many
forms of variance. If has variance 2p , we get the PIGp distribution with probability mass function:
y 2 0.5 1/2p
Pr(Y = y) =
e
(1 + 23p )s/2 Ks (z),
(4.2)
y! 2p
where s = y 0.5 and z = (1 + 22p )0.5 /k1 . In this case, Var[Y ] = +
2p . Similar to the NB distribution, the parameters of the model can be estimated by maximum likelihood, and the p parameter can also be estimated to use as a
tool to choose between the PIG2 and the PIG1 distributions.
yi !
2
2
96
50
Percent
40
30
20
10
0
0
0.02
0.04
0.06
0.08
0.10
Estimated Parameter Alpha
0.12
0.14
(4.3)
The function g(i ) has to be replaced by the form of variance to be tested, such
as g(i ) = 2i for an NB2 variance form or g(i ) = i for an NB1 variance form. By
construction, the NB2, the PIG2, and the PLN2 are all tested by this score test when
g(i ) = 2i . Similarly, the NB1, the PIG1, and the PLN1 are all tested simultaneously
against the Poisson distribution for g(i ) = i .
Then, the null hypothesis H0 : = 0 is tested against Ha : > 0, yielding the
following score statistic test for the Poisson distribution against heterogeneous models
97
n
1
i=1
n
12
1 2
i g(i ) (yi i )2 yi .
2
i=1
2
2
i g (i )
For various forms g(i ), the statistics TLM are normally distributed with mean 0
and variance 1.
4.4.5 True and Apparent Occurrence Contagion
Strong assumptions are needed when we suppose a Poisson distribution for the number
of claims. One of these assumptions is the absence of occurrence dependence. A
positive (negative) occurrence means that the event of a claim increases (decreases) the
probability of another claim. It has been shown that positive occurrence dependence
can lead to a binomial negative distribution. The negative binomial distribution has
also been shown to be the limit distribution of apparent occurrence dependence, where
the modification of the probability of another claim arises from the recognition of the
accident proneness of an individual. Consequently, we cannot deduce the kind of
dependence implied by the data when we suppose a negative binomial distribution.
This impossibility of distinguishing between true and apparent contagion has been
called the impossibility theorem.
In the actuarial literature, the apparent occurrence dependence explanation seems
to be the most accepted thesis. This indicates that although past events do not truly
influence the probability of reporting a claim, they provide some information about
the true nature of the driver. The heterogeneity term of the models can be updated
according to the insureds accident history (see Chapter 7 for heterogeneity analysis
for panel data or Chapter 14 for the a posteriori analysis of heterogeneity).
4.4.6 Numerical Application
Negative binomial distributions that have been applied to the Singapore dataset are
are approximately the same for the
shown in Table 4.2. The mean parameters
NB2, the NB1, and the Poisson distributions. Fit of the distributions can be verified
by the log-likelihood, which shows that the introduction of another parameter by
the NB2 and the NB1 improves the adjustment slightly. However, because another
parameter () has been added to the distributions NB2 and NB1, compared with
the Poisson distribution, we cannot compare the log-likelihood directly. To compare
the fit, information criteria that penalize models with a large number of parameters
should be used. The classical criteria include Akaike information criteria (AIC =
2 log(L) + 2k, where L is the likelihood and k is the number of parameters of the
model) and the Bayesian information criteria (BIC = 2 log(L) + log(n)k, where k
98
Table 4.2. MLE for the NB2 and the NB1 Distributions
NB2
NB1
Parameter
Estimate
(Std.Err.)
0
1
2
3
4
5
6
Log-likelihood
3.7735
(0.5146)
0.7191
(0.1272)
0.3572
(0.1286)
1.5906
(0.5061)
1.5304
(0.5180)
1.1048
(0.5209)
0.3481
(0.5332)
0.4137
(0.2328)
1801.7336
Estimate
(Std.Err.)
3.7509
(0.5122)
0.7112
(0.1259)
0.3470
(0.1277)
1.5729
(0.5037)
1.5247
(0.5151)
1.0895
(0.5184)
0.3374
(0.5310)
0.0273
(0.0227)
1802.9360
and n represent, respectively, the number of parameters of the model and the number
of observations).
Another way of comparing the NB distributions with the Poisson distribution is to
use the score test defined earlier. Applications of these tests lead to values of 2.25 and
1.23, where these statistics are normally distributed with mean 0 and variance 1 under
H0 . Consequently, the Poisson distribution is rejected against the NB2 distribution,
but not against the NB1 distribution.
The difference between the Poisson, the NB1, and the NB2 distribution can be
visualized by looking at the variance function. Figure 4.2 illustrates the variance of
0.60
0.55
0.50
0.45
Variance
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
0.00
0.05
0.10
0.15
0.20
Poisson
0.25
Mean
0.30
NB1
0.35
0.40
0.45
NB2
0.50
99
Table 4.3. MLE for the PIG2 and the PIG1 Distributions
PIG2
PIG1
Parameter
Estimate
(Std.Err.)
0
1
2
3
4
5
6
Log-likelihood
3.7731
(0.5146)
0.7187
(0.1272)
0.3569
(0.1286)
1.5903
(0.5062)
1.5308
(0.5180)
1.1048
(0.5209)
0.3481
(0.5332)
0.4170
(0.2382)
1801.7263
Estimate
(Std.Err.)
3.7506
(0.5123)
0.7109
(0.1260)
0.3467
(0.1278)
1.5727
(0.5037)
1.5248
(0.5151)
1.0897
(0.5184)
0.3376
(0.5310)
0.0277
(0.0232)
1802.9296
each distribution compared with its mean. We can see that the differences between the
models concern the riskiest drivers, where the NB2 distribution, having an exponential
variance function, supposes a bigger variance for these clients.
The PIG and the PLN distributions have also been applied to the Singapore dataset,
with results shown in Tables 4.3 and 4.4, respectively. Covariates are again approximately equal to the ones obtained with the Poisson and the NB distributions. As
explained earlier, the same score test used to test the NB2 and the NB1 distributions
must be applied to the models PIG2-PLN2/PIG1-PLN1. Consequently, we can again
say that the Poisson distribution is rejected against the PIG2 and against the PLN2,
whereas it is not rejected against the PIG1 and the PLN1.
Differences between models can be observed through the mean and the variance of
the annual number of claims for some insured profiles. Several insured profiles have
Table 4.4. MLE for the PLN2 and the PLN1 Distributions
PLN2
PLN1
Parameter
Estimate
(Std.Err.)
0
1
2
3
4
5
6
2
Log-likelihood
3.7732
(0.5146)
0.7186
(0.1272)
0.3567
(0.1286)
1.5904
(0.5062)
1.5312
(0.5180)
1.1050
(0.5209)
0.3483
(0.5332)
0.3462
(0.1693)
1801.7411
Estimate
(Std.Err.)
3.7506
(0.5124)
0.7108
(0.1260)
0.3466
(0.1278)
1.5728
(0.5039)
1.5251
(0.5153)
1.0901
(0.5186)
0.3379
(0.5311)
0.0277
(0.0232)
1802.9236
100
x1
x2
x3
x4
x5
x6
x7
1
2
3
4
0
0
0
1
0
0
0
0
0
0
0
1
0
0
1
0
0
1
0
0
0
0
0
0
0
0
0
0
been selected and are described in Table 4.5. Profile 1 is classified as a good driver,
whereas profile 4 usually exhibits bad loss experience. Other profiles are medium risk.
The premium and the variance for each profile are given in Table 4.6, which shows
that the differences between premiums are smaller than the differences between the
variances. So, although we cannot see differences in premium corresponding to the
expected value (E[Y ]), in the case where the premiums are calculated using a premium
principle based on the standard error (E[Y ] + Var[Y ]) or on the variance (E[Y ] +
Var[Y ]) or, when using other transformations of the distribution (Esscher premium,
for example = E[SehY ]/E[ehY ], for a giver h), the choice of the distribution has an
impact.
Finally, using the estimated parameters for the NB2, the PIG2, and the PLN2 (that
do not depend on the covariates), we can illustrate the heterogeneity distribution of
the underlying gamma, inverse-Gaussian, and lognormal. This is shown in Figure 4.3,
where we see a close similarity between the distributions. Figure 4.4 illustrates the
heterogeneity distribution of profiles 1, 2, 3, and 4, which depends on i . For all distributions, the inverse-Gaussian shows a fatter tail than the gamma and the lognormal
distributions. This property of the inverse-Gaussian does not have a big impact on the
count analysis for cross-section data, but may lead to the highest penalties for bad
drivers (see Chapter 9 on credibility for details).
Table 4.6. Comparison of a Priori Claim Frequencies
1st Profile
2nd Profile
3rd Profile
4th Profile
Dist.
Mean
Var.
Mean
Var.
Mean
Var.
Mean
Var.
Poisson
NB2
NB1
PIG2
PIG1
PLN2
PLN1
0.0229
0.0230
0.0235
0.0230
0.0235
0.0230
0.0235
0.0229
0.0232
0.0241
0.0232
0.0242
0.0232
0.0242
0.0693
0.0693
0.0699
0.0694
0.0699
0.0693
0.0699
0.0693
0.0713
0.0718
0.0714
0.0718
0.0710
0.0718
0.1063
0.1061
0.1079
0.1062
0.1080
0.1061
0.1080
0.1063
0.1108
0.1109
0.1109
0.1110
0.1100
0.1110
0.2314
0.2313
0.2306
0.2313
0.2306
0.2314
0.2306
0.2314
0.2535
0.2369
0.2536
0.2370
0.2499
0.2370
101
1.5
1.4
1.3
1.2
1.1
1.0
g(x)
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0
x
Gamma
Inverse-Gaussian
Lognormal
1.5
1.0
1.0
g(x)
g(x)
1.5
0.5
0.5
0.0
0.0
0
1
Gamma
3
x
Inverse-Gaussian
1
Gamma
Lognormal
1.5
1.0
1.0
3
x
Inverse-Gaussian
2
3
x
Inverse-Gaussian
Lognormal
g(x)
g(x)
1.5
0.5
0.5
0.0
0.0
0
1
Gamma
3
x
Inverse-Gaussian
4
Lognormal
1
Gamma
Lognormal
102
+ (1 ) Pr[V = 0] for n = 0
(1 ) Pr[V = y]
for n = 1, 2, . . .
(4.4)
+ (1 )e
(1
y
) e y!
for y = 0
for y = 1, 2, . . .
From this, we can determine the first two moments of the ZIP distribution: E[Y ] =
(1 ) and Var[Y ] = E[Y ] + E[Y ]( E[Y ]). Covariates can be included in the
parameter with a log link function, as done with the Poisson and the other models with
heterogeneity. Moreover, covariates can also be included in the parameter, with a
logit, a probit, or other appropriate link functions.
The ZIP model could be useful for modeling purposes because it accounts for
overdispersion. Since the only difference between the ZIP model and the Poisson
distribution is found when Y = 0, it is easy to adjust the MLE equations of the Poisson
distribution to find the parameters of the ZIP model. Other count distributions can
also be used with the zero-inflated distributions, such as the NB, the PIG, or the PLN,
leading to distributions called ZI-NB, ZI-PIG, or ZI-PLN, respectively.
103
There are mainly two ways to explain how the zero-inflated distribution can differentiate between the claim and the accident:
(1) A first interpretation to justify the use of a zero-inflated distribution is the assumption
that each year, a number of insureds will not claim at all, whatever the case. In this
situation, one might question why these insureds procure insurance. Some explanations refer to their fear of insurance, their having minimal protection (mandatory
insurance), or their being insured only for major risk (with probability close to 0).
(2) Another way of interpreting the zero-inflated model assumes that the number of
accidents is Poisson distributed. In addition, it considers the probability of each
accident being reported. The model assumes that the first accident of each insured
year indicates the way the insured will act for the rest of the year. Accordingly, if the
first accident is reported, the succeeding ones will also be reported. If the first accident
is not reported, the following accidents will not be reported. Those who will not claim
their first accident, because they made an effort to financially support their decision,
tend to defend the way they act and consequently will not claim other accidents.
When all the insureds analyzed in the portfolio have a risk exposure of 1, meaning
that we observe all the clients for complete one-year contracts, the two interpretations
of the ZIP models generate the same model. Both models are equivalent.
However, when some insureds in the portfolio do not have complete coverage, the
model differs depending on which interpretation we choose. In the first interpretation,
because the insurance portfolio supposes that a portion of the insureds will not report
at all, the parameter that corresponds to the extra probability of having no claim
does not need to be modeled using risk exposure. However, in the second case, the
parameter should include the risk exposure because it is linked to the probability of
reporting the first accident. As seen in Section 4.2.2, the risk exposure of a Poisson
distribution is proportional to the observed time length. It is also important to understand that the Poisson in the ZIP model is also modeled with risk exposure, meaning
that mean parameter of the Poisson is equal to t.
104
Table 4.7. MLE for the Poisson and the Zero-Inflated Poisson Distribution
Poisson
Par.
Est.
0
1
Log-likelihood
ZIP1
(Std.Err.)
Est.
2.2147
(0.0381)
.
.
.
.
.
.
2533.8862
ZIP2
(Std.Err.)
1.7444
(0.1553)
0.3731
(0.0937)
.
.
.
.
2529.7992
Est.
(Std.Err.)
1.6262
(0.1199)
.
.
0.0505
(0.2400)
0.9578
(0.2633)
2516.2043
Phi
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Exposure
ZIP1
ZIP2
Fig. 4.5. Values of for both ZIP models, depending on the risk exposure.
References
105
4.6 Conclusion
A wide selection of models can be used to model count data and, more precisely,
to model the number of claims by an insured. We chose to focus on the Poisson
distribution, on its generalization by the introduction of a random heterogeneity term,
and on zero-inflated models. It is also possible to use other advanced models to model
claim count. Different interpretations of the parameters can be done depending on the
chosen model. The choice of the best distribution should be supported by specification
tests and a goodness-of-fit test. Such count distributions can serve as a basis for the
modeling of time series of count, for panel data, or for hierarchical models.
References
Boucher, J.-P., M. Denuit, and M. Guillen (2009). Number of accidents or number of claims?
An approach with zero-inflated Poisson models for panel data. Journal of Risk and
Insurance 76(4), 821846.
106
5
Generalized Linear Models
Curtis Gary Dean
(5.1)
The k predictor variables are given, nonrandom variables whose values can change
with i. The error terms i are the differences between the response variables and their
107
108
predicted values:
i = yi (0 + 1 xi1 + 2 xi2 + + k xik ).
Two key assumptions are that error terms i have expected value 0, E[i ] = 0, and
that the variance of i is constant and does not change across observations i, a property
referred to as homoskedasticity: Var[i ] = 2 . The errors i are usually assumed to
be independent and normally distributed.
Taking expected values in (5.1) yields
E[yi ] = 0 + 1 xi1 + 2 xi2 + + k xik .
(5.2)
Taking the variance of both sides of (5.1), Var[yi ] = Var[i ] = 2 . The normality
of error terms implies that response variables yi are normally distributed about their
respective means E[yi ].
The coefficients in the linear model are estimated by the method of least squares.1
Normality is not a requirement to construct a linear model using least squares, but is
important for hypothesis testing and constructing confidence intervals. Linear models
have shown their value in modeling, but in many situations linear models need to be
generalized as demonstrated in the next section.
Problems with Predicting Number of Claims
Suppose that yi represents the number of claims for risk i in a portfolio of n risks. The
actuary may want to predict the expected number of claims for each risk i, E[yi ], based
on k risk characteristics xi1 , xi2 , . . . , xik . Multiple linear regression may not be the
best tool for this job. (For an in-depth discussion of modeling counts, see Chapter 4.)
Here are three problems with applying the standard linear model:
(1) The Poisson is commonly used to model the number of claims. If yi is Poisson
distributed, then Var[yi ] = E[yi ]: the variance is not constant across observations, but
depends on the expected value of the response variable. The assumption of constancy
of the variance across risks, Var[yi ] = Var[i ] = 2 , does not hold.
(2) When modeling the expected number of claims, the left-hand side of equation (5.2)
needs to be non-negative, but this cannot be guaranteed in the linear model. It is quite
possible that some combination of the predictors xij could result in a negative value
for b0 + b1 xi1 + b2 xi2 + + bk xik .2
(3) Rather than building an additive model where the contributions of risk characteristics
xi1 , xi2 , . . . , xik are added, perhaps a multiplicative model is more appropriate. For
example, a bad driver may drive predominantly in a territory with a high rate of
1
2
The true coefficients 0 , 1 , . . . , k are unknowable, but are estimated by quantities b0 , b1 , . . . , bk determined from
the data.
The calculated coefficient values b0 , b1 , . . . , bk are used to compute quantities in a model.
109
accidents. Should the contributions from poor driving ability be added to territory
effects, or should these contributions be multiplied?
The linear model needs to be adjusted. Letting ln(E[yi ]) equal the linear combination of predictor variables addresses items 2 and 3. Equation (5.2) becomes
ln(E[yi ]) = 0 + 1 xi1 + + k xik which gives
E[yi ] = e0 e1 xi1 ek xik .
With this model the expected number of claims for risk i, E[yi ], will not be negative,
and the predictive model is multiplicative. This adjustment to the linear model raises
another issue: how to determine coefficients j in this nonlinear equation.
Linear models can be extended to account for heteroskedasticity (i.e., Var[yi ] is
not constant across risks as in item 1). One technique is to use weighted least squares
with weights wi = 1/Var[yi ] that are inversely proportional to variances to estimate
coefficients 0 , 1 , 2 , . . . , k . Less weight would be given when the variance is
greater. This approach needs a model for how Var[yi ] changes across risks i. If we
assume that yi is Poisson distributed for each risk i, then it follows that Var[yi ] = E[yi ]
but we have the problem that Var[yi ] is unknown until E[yi ] is calculated.
Some of the complications arising from predicting the number of claims with linear
models can be addressed by moving on to generalized linear models. Predicting the
number of claims is only one of many applications where linear models need to be
extended. Generalized linear models were developed to unify a variety of statistical
models.
5.1.2 Generalized Linear Model Assumptions
Generalized linear models (GLMs) generalize linear regression in two important ways:
(1) The independent response variables yi can be linked to a linear function of predictor
variables xij with a nonlinear link function.
(2) The variance in the response variables yi is not required to be constant across risks,
but can be a function of yi s expected value.
A function is strictly monotonic if it is strictly increasing or strictly decreasing. If f (x) is strictly increasing then
x1 < x2 implies f (x1 ) < f (x2 ). Strictly decreasing is defined similarly.
110
(5.3)
The predictor variables xij are still combined into a linear function, but the response
variable E[yi ] can be a nonlinear function of this linear combination. The linear
function of predictor variables is often assigned the symbol :
i = 0 + 1 xi1 + + k xik .
(5.4)
The function can be a probability density function (e.g., normal), probability mass function (e.g., binomial), or a
mixture (e.g., Tweedie).
Formula (5.5) is a special case of the more general formula f (y; ) = exp[(r( )h(y) + s(y) + q( )] that defines
the exponential family. Note that h(y) has been replaced by a linear term y in (5.5) Distributions whose density
functions can be written as (5.5) are members of the linear exponential family.
111
is called the dispersion parameter or, sometimes, nuisance parameter because the
mean of the distribution does not depend directly on . The functions b( ), a(), and
c(y, ) determine the type of distribution; for example, normal, Poisson, and so on.
The Poisson distribution, with one parameter, also fits into this family as shown in
Example 5.1.
The mean and variance of the distribution are simply
E[y] = b (),
Var[y] = a()b (),
(5.6)
(5.7)
where b ( ) is the first derivative with respect to and b () is the second derivative.
Derivations of the formulas for the mean and variance are shown in the appendices.
Distributions in this exponential family include the normal, binomial, Poisson,
exponential, gamma, inverse-Gaussian, and the compound Poisson-gamma. With a
little algebra the common forms of these distributions can be rewritten in exponential
family form (5.5).
Example 5.1. The Poisson distribution is a member of the exponential family with
probability mass function
y e
y!
y
e
= exp ln
y!
y ln
= exp
ln y! .
1
f (y; ) =
112
y 2 /2
y
2
= exp
+ ln( 2 ) .
2
2 2
The parameter corresponds to in (5.5) and = 2 . Rewriting the pdf with
parameters and ,
2
y
y 2 /2
+ ln( 2 ) .
f (y; , ) = exp
2
2
y
+ ln( 2 ) . The mean and variSo, b( ) = 2 /2, a() = , and c(y, ) = 2
ance are
d( 2 /2)
= =
d
d 2 ( 2 /2)
= = 2.
Var[y] = a()b ( ) =
d 2
E[y] = b ( ) =
The most complicated term, c(y, ), is not relevant in the computation of the mean
and the variance.
Table 5.1 displays five common distributions in the exponential family. The pdfs
for distributions are shown on the left side of the table in commonly displayed forms.
The exponential family parameters for each distribution can be derived from the pdfs
on the left.
Suppose that response variable yi represents the loss severity for the i th risk where
loss severity = losses / number of losses. A loss severity computed from several
observed losses will be a better predictor of the expected loss severity than would
a single loss. If the variance of a single loss is Var[L], then the variance in loss
severity based on m losses would be Var[L]/m. Greater weights should be attached to
observed severities based on more losses. This same argument would apply to other
types of response variables whose value may depend on an average.
Example 5.3. Let vj be independent random variables with vj N(, 2 ) for j =
1, . . . , w. If y = (v1 + + vw )/w is the average, then E[y] = and Var[y] = 2 /w.
113
b( )
a()
1
(y )2
Normal:
exp
2 2
2 2
y
Poisson: e /y!
n y
a
p (1 p)ny
Binomial :
y
Gammab : y 1 ey / ()
!
(y )2
Inverse-Gaussian:
exp
2y 3
2y2
c(y, )
y2
+ ln(2 )
ln(y!)
n
ln
y
1 y
1
ln ln y
1
12 ln(2 x 3 ) +
y
/2
ln
ln[p/(1 p)]
n ln(1 + e )
1
22
ln()
12
Notes: Other sources may show different parameterizations. Common forms of pdfs are displayed so that readers can make their own
calculations and reconcile formulas.
a
n is assumed to be a known, fixed quantity.
b
n With this parameterization the mean and variance are E[y] = / and Var[y] = / 2 .
114
Using the results of Example 5.2, the probability density function for y is
2
y 2 /2
y
f (y; , , w) = exp
+ ln( 2 /w) .
/w
2/w
where = 2 again. Note that the pdf for random variable y depends on the weight
w. As the weight w increases, the variance of y decreases.
For response variables yi it is standard practice, as shown in Examples 5.2 and
5.3, to define parameter and function a() so that a() can be replaced by /wi in
equation (5.5). The pdf for yi becomes
yi i b(i )
f (yi ; i , , wi ) = exp
+ c(yi , /wi ) .
(5.8)
/wi
Parameter is called the scale parameter or dispersion and is usually assumed to be
constant for all response variables in the sample. The weights can be identically one
or can be inputs into the GLM. In insurance applications the number of claims or
exposures are common measures for weights wi .
The Appendix shows how to put the gamma and binomial distributions into the
exponential family form shown in (5.5).
5.2.1 The Variance Function and the Relationship between
Variances and Means
The variance formula from (5.7) with /wi replacing a() is
Var[yi ] =
b (i ).
wi
(5.9)
V (i ),
wi
(5.10)
where V (i ) = b (b1 (i )). Function V (i ) is called the variance function. The
variance function defines the relationship between the variance and the mean for a
distribution in the exponential family (see Table 5.2).
115
V ()
Distribution
V ()
0 = 1
(1 )
Tweedie
gamma
inverse-Gaussian
p , 1 < p < 2
2
3
The Tweedie6 distribution is the name commonly used for the compound Poissongamma distribution. Note that its exponent p lies between that of the Poisson and
gamma. Intuitively that makes sense because the number of losses is Poisson distributed and the size of each loss is gamma distributed.
The variance of response variable yi controls how much weight to put on observation yi when fitting the GLM. As is demonstrated in Section 5.4, when Var[yi ] is
larger, less weight will be given to observation yi in calculating GLM coefficients
b0 , b1 , . . . , bk than when Var[yi ] is smaller:
Weight for observation i
1
wi
=
.
Var[yi ]
V (i )
Maurice Charles Kenneth Tweedie (19191996) was a medical statistician. Tweedie distributions that bear his name
are a particular subclass of an exponential dispersion model. The compound Poisson-gamma is included in the
subclass.
116
identity
log
logit
probit
complementary log-log
inverse
inverse squared
g()
g 1 ()
Range of g 1 ()
ln()
ln[/(1 )]
1 ()
ln((ln(1 ))
1/
1/2
e
e /(1 + e )
()
1 ee
1/
1/
(, )
(0, )
(0, 1)
(0, 1)
(0, 1)
(, 0) (0, )
(0, )
117
(5.11)
Canonical parameters i are equal to a linear function of the predictors. The chosen
link function g() is called a canonical link function.
Canonical links generate linear equations (5.11) for unknown parameters i .7 A
canonical link may be a good choice for modeling a particular problem, but the
modeler should not feel compelled to select a canonical link. The overall fit of the
model and other considerations such as intuitive appeal may be more important.
Example 5.4. Parameters and functions for the gamma distribution are shown in
Table 5.1. For a gamma distribution b( ) = ln( ) so b () = 1/ , which implies
= 1/ . Solving for gives = 1/.
If link function g() = 1/ is used with a gamma distribution, then the predictive
equations can be written as
i = 0 + 1 xi1 + + k xik .
Link function g() = 1/ is a canonical link for the gamma. If one uses the link
g() = 1/ instead, then
i = 0 + 1 xi1 + + k xik .
But multiplying through by 1 will reverse the sign of the coefficients on the right,
showing that the i can still be calculated as linear functions of the predictors. The
same goes for any constant times on the left. Multiply through by the reciprocal.
Most sources will get rid of the negative sign of 1/ and display g() = 1/ as the
canonical link for the gamma distribution.
On page 32 of McCullagh and Nelder (1997) they explain that canonical links result in the existence of a sufficient
statistic for the model.
118
a
()
j
i
i=1
=
n
i=1
1
i
b(i )
yi
.
ai ()
j
j
Note that only the i s are functions of the j s. The yi s and do not depend on the
j s.
The connection between the i s and j s is a little complicated. The following
three equations are the thread:
i = b (i ),
g(i ) = i ,
i = 0 + 1 xi1 + + k xik .
The chain rule for differentiation says
i i i
=
.
j
i i i j
The subscript i is applied to ai () and ci (yi , ) to recognize that a weight wi can be included in the model as shown
in Example 5.3 and equation (5.8).
119
Equation (5.7) says that the first two terms in the denominator are the variance
of yi : Var[yi ] = ai ()b (i ). This can be rewritten using the variance function and
weight, as shown in equation (5.10), as Var[yi ] = (/wi )V (i ), giving
l(y; ) wi (yi i )xij
=
= 0.
j
V (i )g (i )
i=1
n
(5.13)
for j = 0, 1, . . . , k. Note that xi0 = 1 because 0 is the intercept. The j s are included
in (5.13) through
i = g 1 (0 + 1 xi1 + + k xik ).
There are k + 1 equations with k + 1 unknowns.
What is going on inside (5.13)? The weighted sum of the differences (yi i )
should equal 0. This is the MLE solution for the best coefficients b0 , b1 , . . . , bk to
predict E[yi ].
The weights applied to (yi i ) are wi xij /(V (i )g (i )). Values for wi are
specified by the modeler; for example, wi = number of claims if yi represents claim
severity. A larger wi gives more weight to the difference (yi i ). A larger variance
at data point i as captured by variance function V (i ) reduces the weight placed on
the difference (yi i ). Note that has no effect on the solution if it is constant
across risks. The g (i ) in the denominator makes an adjustment for the effect of the
link function.
Note that (5.13) works for any distribution in the exponential family defined by
(5.5). The variance function V (i ) is the only characteristic of the distribution that is
required to make the fit.
Statistical packages have numerical methods to maximize the log-likelihood function (5.12). A technique commonly used is referred to as iteratively reweighted least
squares. One also sees the name Fisher Scoring algorithm. For those interested in the
details of these numerical techniques, see Dobson and Barnett (2008); Gill (2000);
McCullagh and Nelder (1997); and Nelder and Wedderburn (1972).
Example 5.5. Suppose that a GLM model were built assuming a Poisson distribution,
log link g(i ) = ln(i ), and constant weights wi = 1. The Poisson assumption implies
V (i ) = i and = 1. The log link gives g (i ) = 1/i . Equations (5.13) become
n
i=1
(yi i )xij = 0
120
i=1
n
(yi i )2 .
i=1
where xi0 = 1. Setting the partial derivatives equal to zero yields a system of k + 1
equations to solve for the bj s:
n
(yi i )xij = 0.
i=1
These are the same as equation (5.13) if one assumes that (1) weights wi = 1, (2)
the link function is the identity implying g (x) = 1, and (3) the distribution is normal
and homoskedasticity holds. Requirement (3) means that V (i ) = 1 (see Table 5.2)
and = 2 is a constant and can be canceled. So, classic linear regression is a special
case of a GLM. Weighted least squares is also a special case of (5.13).
5.4.1 Quasi-Likelihood
Here we take a quick look at a generalization of the results of the prior section. Inside
(y)
the sum in equation (5.13) is the fraction V
where we have dropped the subscripts.
()
Integrating with respect to and assigning limits gives
yt
dt.
(5.14)
Q(, y) =
y V (t)
121
(y)
Note that Q/ = V
. Function Q(, y) is called the log quasi-likelihood,
()
though log is usually dropped from the name. The quasi-likelihood for dataset
y1 , . . . , yn is
Q(, y) =
n
Qi (i , yi ).
(5.15)
i=1
r Link function g(x) is strictly monotonic and differentiable with inverse function g 1 (x).
r The expected values of yi , i = E[yi ], are predicted by the equations
g(i ) = i or i = g 1 (i ) for i = 1, . . . , n.
122
Insured
Age
Gender
Marital
Status
Territory
Collision
Coverage
Purchased
Homeowners
30
51
23
42
60
:
F
M
F
F
M
:
M
M
S
M
S
:
T01
T03
T06
T11
T03
:
Y
Y
N
Y
N
:
1
0
0
1
0
:
5.6 Applications
This section presents three stylized examples that demonstrate basic GLM modeling
in insurance.
5.6 Applications
123
Policy
Years
Gender
Territory
Claims
5
5
4
3
4
3
5
5
3
2
4
5
M
F
M
F
F
F
M
M
M
F
M
F
East
East
East
West
East
West
West
West
East
East
West
West
0
0
1
1
0
1
0
2
1
1
1
0
0 +1 xGender(i) +2 xTerritory(i)
(5.16)
124
GENDER and TERRITORY are categorical variables, and in this simple example,
each has two levels or categories. The predictors xGender(i) and xTerritory(i) are binary
indicator variables that take on values of either 0 or 1. Choosing the base frequency
to be FEMALE and EAST, the predictors take on values
#
0, Female,
xGender(i) =
1, Male,
#
0, East,
xTerritory(i) =
1, West.
5.6 Applications
125
increases, then both E[yi ] and Var[yi ] will increase. Policy year offsets ln(mi ) capture
these effects in the log-link Poisson model.
The measure of variability, Var[yi ], in claim counts yi increases as policy years
increase. Putting policy years mi into the model as weights has the opposite and wrong
effect. A bigger weight mi means that more weight will be given to observation yi
because you have told the model that yi is less variable. You also neglected to tell the
model that E[yi ] will be bigger for risks with larger mi . The model will identify risks
with more experience as poorer risks. So, a risk with many policy years of experience
will tend to be classified as a poor risk, and the model will assign a lot of weight to
this observation.
Now the glm function can be run with Table 5.5 as input. Although statistical
software can handle categorical variables, GENDER (FEMALE=0 and MALE=1) and
TERRITORY (EAST=0, WEST=1) are switched to numeric for this example. Here is
the one line of code to run the GLM in R.
freqmodel <- glm(Claims ~ Gender + Territory,
family = poisson(link=log), data = ClaimCountData,
offset=log(Years))
The R code is similar to that in the prior section except that we are modeling with a
Poisson distribution and log link. We also include an offset log(Years) to account
for the varying number of experience years across policies. The program will take the
natural log of YEARS from Table 5.5 and use that as an offset.
The coefficients are b0 = 2.2214, b1 = 0.3282, and b2 = 0.4152.9 The expected
frequencies are
i = e2.2214 (e0.3282 )xGender(i) (e0.4152 )xTerritory(i) .
The four possibilities for annual claim frequencies are (FEMALE,EAST) = 0.1085,
(FEMALE,WEST) = 0.1643, (MALE,EAST) = 0.1506, and (MALE,WEST) = 0.2281.
A quick check for reasonability can be performed. For each policy in Table 5.5
multiply the number of years by the appropriate i from the GLM model and add up
the 12 results. The total is 8.00, which exactly matches the eight claims shown in the
Claims column of the table.
For more complete coverage of modeling claim counts see Chapter 4, which discusses regression models that model count data. In addition to the Poisson distribution,
it discusses more general distributions including the negative binomial and the zeroinflated Poisson.
9
The unknown coefficients 0 , 1 , and 2 have been relabeled as b0 , b1 , and b2 now that specific values have been
estimated.
126
Industry
1
2
3
4
5
6
:
Retail
Retail
Service
Office
Restaurant
Contract
:
Territory
Number
of Claims
Total $
Claims
Average
Claim
B
B
C
A
D
C
:
1
1
3
2
1
2
:
850
2,070
5,430
400
1,100
18,560
:
850
2,070
1,810
200
1,100
9,280
:
5.6 Applications
127
Note the similarities and differences of this R code with that in the prior section
where we modeled frequency. In this example we use a gamma distribution rather
than Poisson for family and use weights rather than offset. Another difference is that
this example has categorical variables INDUSTRY and TERRITORY, whereas in the
frequency example GENDER and TERRITORY are binary variables.
.
= freq severity =
premium
exposure
number of losses
The frequency and severity components can be modeled separately, and then the
predicted frequencies and severities can be multiplied together to produce pure premiums. For loss ratios the denominator for frequency is premium, so frequency would
be number of losses per unit of premium. Note that if the loss ratio is modeled, the
offset in the frequency model as in Section 5.6.2 would be premium, the measure of
exposure to loss.
Frequency and severity may be affected differently by risk characteristics so coefficients and predictors in frequency and severity models may be different. Modeling
frequency and severity separately may provide more insight into the loss process than
modeling the losses directly.
5.6.4.2 Model Losses Directly
A compound Poisson-gamma distribution is often used to model aggregate losses. If
the number of losses n is Poisson distributed and loss amounts uj are i.i.d. gamma
distributed, then aggregate loss y can be written as
y = u1 + u2 + + un .
Frequency and severity are combined into a single distribution and model.
128
129
5.7.1 Deviance
A statistical measure called deviance is commonly used to evaluate and compare
GLMs and it is based on log-likelihoods. The log-likelihood function for linear exponential family distributions is
n
yi i b(i )
l(y; ) =
+ ci (yi , ) ,
ai ()
i=1
where vector represents the n canonical parameters 1 , 2 , . . . , n . When a particular
GLM is constructed lets call it model M coefficients bj are calculated to maximize
the log-likelihood. As explained earlier in the chapter, canonical parameters can be
computed using these coefficients and predictive variables xi1 , xi2 , . . . , xik for the
chosen distribution and link function. If a canonical link is chosen for Model M then
the canonical parameters have the simple linear form iM = b0 + b1 xi1 + + bk xik
for i = 1, . . . , n; otherwise the function is nonlinear. Let l(y; M ) denote the value of
the log-likelihood for these parameters iM .
If the n parameters i were free to take on any values, what values would maximize
the log-likelihood? The log-likelihood will be maximized if each term of the sum is
maximized. Take the partial derivative of term i with respect to i and set it equal to
zero:
yi i b(i )
+ ci (yi , ) = yi b (i ) = 0.
i
ai ()
The canonical parameters will satisfy yi = b (iS ) where the superscript S indicates
that we have a saturated model with a separate parameter iS for each observation
yi . The mean of a distribution is = b ( ) (see Appendix B) so in the saturated
model the estimated mean Si of distribution i equals the value of observation i:
Si = b (iS ) = yi .10
10
The conclusion that the log-likelihood will be maximized by Si = yi is also apparent from equation (5.13). Note
that all of the (yi i ) terms in the sum are zero when i = yi . The log-likelihood can also be written as a
function of the distribution means in place of the canonical parameters: l(y; ). In this form the saturated model
log-likelihood l(y; S ) can be written simply as l(y; y).
130
The difference between the log-likelihoods of the saturated model S and model
M is
n
iS iM ) (b(i S ) b(iM ))
(
y
i
.
l(y; S ) l(y; M ) =
ai ()
i=1
This quantity is non-negative because l(y; S ) l(y; M ). If ai () = /wi then scaled
deviance D (y; M ) for model M is defined as
D (y; M ) = 2[l(y; S ) l(y; M )]
n
1
S
=
2wi [yi (iS iM ) (b(i ) b(iM ))].
i=1
+ ln(2 )
f (yi ; i , ) = exp
2
with i = i and = 2 .
With the identity link the predicted means for the n response variables in Model
M are M
i = b0 + b1 xi1 + + bk xik . With a normal distribution the identity link is
the canonical link so iM = M
i . In saturated model S the fitted means are the actual
observations: iS = Si = yi . The deviance for model M is
D(y; M ) =
n
2
2
2[yi (yi M
M
i ) (yi /2 (
i ) /2)]
i=1
n
2
(yi M
i ) .
i=1
The deviance is the residual sum of squares, a goodness of fit measure used in linear
regression.
Example 5.8. Generalized linear model Q assumed Poisson distributions f (yi ) =
i yi ei /yi ! for response variables yi , log links, and no weights. The distribution for
131
yi i ei
ln yi !
f (yi ; i ) = exp
1
n
2[yi (ln yi ln Q
Q
i ) (yi
i )].
i=1
ln Q
i
with
= b0 + b1 xi1 + + bk xik for i = 1, . . . , n. Note that the deviance goes
to zero as fitted means Q
i approach observations yi .
The output from the glm function in R displays null deviance and residual deviance.
Null deviance is the residual deviance for a GLM with only a constant term. The null
deviance provides an upper bound on residual deviance; 0, the deviance of the saturated
model, is the lower bound: 0 D (y; M ) D (y; Null ).
The residual deviance can be used to compare two nested models. Suppose model P
has p explanatory variables, xi1 , . . . , xip .11 Suppose that model Q has q explanatory
variables that include all of the explanatory variables of model P plus q p > 0
additional explanatory variables. The difference in the residual deviances is equivalent
to a likelihood-ratio statistic:
L(y; Q )
.
D (y; P ) D (y; Q ) = 2[l(y; Q ) l(y; P )] = 2 ln
L(y; P )
This statistic has an approximate 2 distribution with q p degrees of freedom.
Example 5.9. In Section 5.6.1 a model was created to predict the probability that an
auto insurance policyholder would also buy a homeowners policy. Suppose that the
actuary has built two models and wants to compare:
Model P : g(i ) = 0 + 1 AGEi + 2 GENDERi
Model Q: g(i ) = 0 + 1 AGEi + 2 GENDERi + 3 MARITALSTATUSi
+ 4 COLLCOVERAGEi .
Model P is more restrictive than model Q in that model P says 3 and 4 must
be 0, whereas coefficients 3 and 4 can have nonzero values in model Q. Setting
this up as a hypothesis test, the null hypothesis is H0 : 3 = 4 = 0. The alternative
hypothesis is H1 : At least one of 3 or 4 is nonzero.
11
132
The test statistic is T = D (y; P ) D (y; Q ). The critical value for a 2 test
with two degrees of freedom with an = 5% significance level is c = 5.99. The null
hypothesis will be rejected if T > 5.99. If one chooses model Q instead of model P
based on this 2 test, then the change in the deviance has to be large enough to justify
the choice, in this case, T > 5.99.
2r(r + 1)
.
nr 1
As the sample size becomes large AICC approaches AIC. In many insurance applications that include individual policies or claims, the AIC and AICC may be essentially
the same because n is large.
BIC is short for Bayesian information criterion. It is also referred to as SBC, the
Bayesian criterion. Like AIC and AICC it makes an adjustment to the log-likelihood:
BIC = 2[l(y; M ) + r ln(n)].
In addition to these three measures, there are other measures that make adjustments
to the log-likehood for number of parameters and sample size.
5.9 Appendix A. Binomial and Gamma Distributions in Exponential Family Form 133
5.8 Conclusion
The goal of this chapter is to provide an introduction to generalized linear models.
Many useful topics were not covered, and for those topics that were addressed much
was left out. This chapter is only a starting point for those who will use GLMs.
GLMs greatly extend the power of regression models. Data that actuaries deal
with constantly such as claim frequencies, loss amounts, pure premiums, and loss
ratios can be response variables in GLMs. The confining assumptions of classic linear
regression such as normality, homoskedasticity, and linearity are eliminated.
GLMs allow modelers to combine many predictors into one model and consider all
of their effects simultaneously, thereby making it more efficient to construct actuarial
models while improving the accuracy of models. Though this chapter has focused
on ratemaking, GLMs can be applied in many other areas of interest to actuaries,
including reserving, marketing, underwriting, and claims analysis.
p
[n ln(1 p)]
y ln 1p
n
.
+ ln
= exp
1
y
p
This is the exponential
family form with = ln 1p
, b() = n ln(1 p),
n
.
It
assumed
that
n
is
a
known
and
fixed quantity. Inverting
a() =1, and
c(y)
=
ln
y
p
134
linear predictor. The logit link is referred to as the canonical link for the binomial
distribution.
The mean for any distribution in exponential family form is b (). Tying this back
to familiar parameters n and p,
d
[n ln(1 + e )]
d
e
=n
1 + e
= np.
b ( ) =
y 1 ey
.
()
y 1 ey/
y 1
ln
= exp
+ ( 1) ln y ln () + ln .
1/
This is exponential family form with = 1/, b( ) = ln , a() = 1/, and
c(y, ) = ( 1) ln y ln () + ln .
Because = 1/ and = / is positive, it follows that < 0. Substituting
1/ in for , then b() = ln(1/ ) = ln( ). Calculating the mean for the
distribution,
b ( ) =
d
1
( ln( )) =
= .
d
The variance is a()b ( ), which is (1/)(1/ 2 ). Converting back to original parameters this is /( 2 ), which is the variance Var[y].
5.10 Appendix B. Calculating Mean and Variance from Exponential Family Form 135
a()
The score function is used to compute maximum likelihood estimates and also in
approximating the error in maximum likelihood estimates. Two properties of the
score function that we use without proof are
l(, ; y)
Ey
= 0, Identity 1.
2
l(, ; y)
l(, ; y) 2
+ Ey
Ey
= 0. Identity 2.
2
136
Using identity 1 with the score function for the exponential family pdf given earlier,
l(, ; y)
0 = Ey
y b ()
= Ey
a()
=
E[y] b ()
.
a()
2
y b ()
y b ( ) 2
= Ey
Ey
a()
a()
2
y 2yb ( ) + (b ( ))2
b ()
=
E
Ey
y
(a())2
a()
E[y 2 ] 2E[y]b ( ) + (b ( ))2 = a()b ()
E[y 2 ] (E[y])2 = a()b ()
Var[y] = a()b ().
This is the variance for distributions that can be put into the linear exponential family
form.
References
Anderson, D., S. Feldblum, C. Modlin, D. Schirmacher, E. Schirmacher, and N. Thandi
(2005). A Practitioners Guide to Generalzied Linear Models: A Foundation for Theory,
Interpretation, and Application.
Dobson, A. J. and A. G. Barnett (2008). An Introduction to Generalized Linear Models
(Third ed.). Chapman & Hall/CRC.
Gill, J. (2000). Generalized Linear Models: A Unified Approach. Sage University Papers
Series on Quantitative Applications in the Social Sciences, series no. 07-134. Sage
Publications.
Jrgensen, B. (1987). Exponential dispersion models. Journal of the Royal Statistical
Society. Series B (Methodological) 49(2), 127162.
Klugman, S. A., H. H. Panger, and G. E. Willmot (2008). Loss Models: From Data to
Decisions (Third ed.). Wiley.
McCullagh, P. and J. Nelder (1997). Generalized Linear Models (Second ed.). Chapman &
Hall/CRC.
References
137
6
Frequency and Severity Models
Edward W. Frees
Chapter Preview. Many insurance datasets feature information about frequency, how
often claims arise, in addition to severity, the claim size. This chapter introduces
tools for handling the joint distribution of frequency and severity. Frequency-severity
modeling is important in insurance applications because of features of contracts,
policyholder behavior, databases that insurers maintain, and regulatory requirements.
Model selection depends on the data form. For some data, we observe the claim
amount and think about a zero claim as meaning no claim during that period. For
other data, we observe individual claim amounts. Model selection also depends on
the purpose of the inference; this chapter highlights the Tweedie generalized linear
model as a desirable option. To emphasize practical applications, this chapter features
a case study of Massachusetts automobile claims, using out-of-sample validation for
model comparisons.
139
140
about the insurance claims event. A separate file, often known as the claims file,
records details of the claim against the insurer, including the amount. (There may
also be a payments file that records the timing of the payments, although we do not
deal with that here.) This recording process makes it natural for insurers to model the
frequency and severity as separate processes.
Regulatory and Administrative. Insurance is a closely monitored industry sector.
Regulators routinely require the reporting of both claims numbers and amounts. This
may be due to the fact that there can be alternative definitions of an amount (e.g., paid
versus incurred), and there is less potential error when reporting claim numbers.
At a broad level, it is clear that insurers need very different administrative systems for handling small, frequently occurring, reimbursable losses (e.g., prescription
drugs) versus rare occurrence, high-impact events, such as inland marine. Every insurance claim means that the insurer incurs additional expenses, suggesting that claims
frequency is an important determinant of expenses.
There are considerable differences of opinion concerning the importance of frequency models for allocated loss adjustment expenses (ALAE), costs that can be
associated with a specific claim (e.g., legal defense fees and claims adjuster costs).
According to Werner and Modlin (2010), it is common to assume that ALAE vary by
the amount of the claim rather than by frequency.
6.2 Sampling and the Generalized Linear Model
6.2.1 Sampling
For a sampling basis, begin by thinking about the policyholder and claims databases
that an insurer maintains. An insurer enters new contracts with insureds and administers claims continuously over time. For some purposes, it is helpful to consider a
continuous-time stochastic process as a sampling model; this is the perspective of the
loss reserving described in Chapter 18 and the survival modeling presented in Chapter 19 (where we examine processes that influence policy retention). In contrast, this
chapter focuses on collections of policies without an emphasis on the exact calendar
start date of the policy. In particular, it only considers closed claims, leaving issues
of claim development to those chapters. Ratemaking and reinsurance are the primary
purposes of this chapter rather than reserving and policy retention.
To establish notation, for each policy {i }, the potentially observable responses are
r Ni the number of claims (events),
r yij , j = 1, . . . , Ni the amount of each claim (loss), and
r Si = yi1 + + yiNi , the aggregate claim amount.
By convention, the set {yij } is empty when Ni = 0.
141
For a specific accounting period (such as a year), the sample of observable responses
may consist of the following:
(1) Si , so that only aggregate losses are available. For example, when examining losses
for commercial insurance, it is common that only aggregate losses are available.
(2)
(Ni , Si ), so that the number and amount of aggregate losses are available.
(3) Ni , yi1 , . . . , yi,Ni , so that detailed information about each claim is available. For
example, when examining
personal
automobile claims, losses for each claim are
available. Let yi = yi1 , . . . , yi,Ni be the vector of individual losses.
We can use ideas from conditional probability to decompose the distribution into
frequency and severity components. To be specific, consider the third data form. Suppressing the {i} subscript, we decompose the distribution of the dependent variables
as
f (N, y) = f (N) f (y|N)
joint = frequency conditional severity,
where f (N, y) denotes the joint distribution of (N, y). This joint distribution equals
the product of the two components:
1. claims frequency: f (N ) denotes the probability of having N claims, and
2. conditional severity: f (y|N ) denotes the conditional density of the claim vector y
given N .
The second data form follows similarly, replacing the vector of individual losses y
with the aggregate loss S. We can even decompose the first data form by breaking
off the zero event through the indicator notation ri = I(Si > 0) for the frequency
component and conditioning on ri = 1 for the severity component. We examine this
data form using two-part models in Section 6.3.1.
Through this decomposition, we do not require independence of the frequency
and severity components as is traditional in the actuarial science literature. There are
many ways to model dependence when considering the joint distribution f (N, y). For
example, one may use a latent variable that affects both frequency N and loss amounts
y, thus inducing a positive association. Copulas are another tool used regularly by
actuaries to model nonlinear associations. The conditional probability framework is
a natural method of allowing for potential dependencies and provides a good starting
platform for empirical work.
6.2.2 Generalized Linear Model
A natural starting platform for empirical modeling of both frequency and severity
components is the generalized linear model (GLM) introduced in Chapter 5. Indeed,
142
one reason for the popularity of this modeling framework is that it has the flexibility
to address both frequency and severity models.
Thinking of a generic dependent variable yi (without regard to whether it represents
frequency or severity), we focus on logarithmic links so that the mean function is
E yi = exp(xi ).
In some instances, the mean is known to vary proportionally with a variable that we
label as Ei , for exposure; see the next subsection for more discussion of exposures.
To incorporate exposures, one can always specify one of the explanatory variables
to be ln Ei and restrict the corresponding regression coefficient to be 1; this term is
known as an offset. With this convention, the link function is
ln i = ln Ei + xi .
Example 6.1 (Relativities). In this example, we consider a small fictitious dataset
that appears in Werner and Modlin (2010). The data consist of loss and loss adjustment
expenses (LossLAE), decomposed by three levels of an amount of insurance (AOI)
and three territories (Terr). For each combination of AOI and Terr, we have available
the number of policies issued, given as the exposure.
AOI
Low
Medium
High
Low
Medium
High
Low
Medium
High
Terr
Exposure
LossLAE
1
1
1
2
2
2
3
3
3
7
108
179
130
126
129
143
126
40
210.93
4458.05
10565.98
6206.12
8239.95
12063.68
8441.25
10188.70
4625.34
Our objective is to fit a generalized linear model (GLM) to the data using LossLAE
as the dependent variable. We would like to understand the influence of the amount
of insurance and territory on LossLAE.
We now specify two factors and estimate a generalized linear model using a
gamma distribution with a logarithmic link function. In the R output that follows,
the relevel command allows us to specify the reference level. For this example,
a medium amount of insurance (AOI = medium) and the second territory (Terr =
2) are chosen as the reference levels. Logarithmic exposure is used as an offset variable
143
so that cells (combinations of the two categorical variables) with larger numbers of
exposures/policies will have larger expected losses.
Selected R Output
> Sampdata$AOI = relevel(Sampdata$AOI, ref = "Medium")
> Sampdata$Terr = factor(Sampdata$Terr)
> Sampdata$Terr = relevel(Sampdata$Terr, ref = "2")
> summary(glm(LossLAE AOI + Terr, offset = log(Exposure),
data = Sampdata, + family = Gamma(link = "log")))
Call:
glm(formula = LossLAE AOI + Terr, family = Gamma(link = "log"),
data = Sampdata, offset = log(Exposure))
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 4.180e+00 1.975e-06 2116446
<2e-16
AOIHigh
3.577e-01 2.164e-06 165302
<2e-16
AOILow
-3.147e-01 2.164e-06 -145448
<2e-16
Terr1
-4.601e-01 2.164e-06 -212656
<2e-16
Terr3
2.123e-01 2.164e-06
98109
<2e-16
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1
***
***
***
***
***
1
on 8
on 4
degrees of freedom
degrees of freedom
Variable
Parameter
Estimate
Relativity (exponential
parameter estimate)
Intercept
AOILow
AOIMedium
AoIHigh
Terr1
Terr2
Terr3
4.18
0.3147
0
0.3577
0.4601
0
0.2123
65.366
0.730
1
1.430
0.631
1
1.237
144
Exposure Basis
Personal Automobile
Homeowners
Workers Compensation
Commercial General Liability
Commercial Business Property
Physicians Professional Liability
Professional Liability
Personal Articles Floater
145
r consider any preexisting exposure base established within the industry, and
r for some lines of business, be proportional to inflation. In this way, rates are not sensitive
to the changing value of money over time because these changes are captured in exposure
base.
To illustrate, consider personal automobile coverage. Instead of the exposure basis
earned car year, a more accurate measure of the quantitative exposure to loss might
be the number of miles driven. However, this measure is difficult to determine at the
time the policy is issued and is subject to potential manipulation by the insured.
For frequency and severity modeling, it is customary to think about the frequency
aspect as proportional to exposure and the severity aspect in terms of loss per claim
(not dependent on exposure). However, this does not cover the entire story. For many
lines of business, it is convenient for exposures to be proportional to inflation. Inflation
is typically viewed as unrelated to frequency, but as proportional to severity.
6.2.3.1 Small Exposures
We begin by considering instances where the units of exposure may be fractions.
To illustrate, for our automobile data, Ei will represent the fraction of the year that
a policyholder had insurance coverage. The logic behind this is that the expected
number of accidents is directly proportional to the length of coverage. This can also
be motivated by a probabilistic framework based on collections of Poisson-distributed
random variables known as Poisson processes (see, for example, Klugman, Panjer,
and Willmot, 2008).
For binary outcomes, this situation is less clear. One way to handle exposures is to
let the logit link function depend on exposures. To this end, the basic logit link function
ez
ez
is (z) = 1+e
z . Define an exposure-weighted logit link function to be i (z) = Ei 1+ez .
With this definition, the probability of a claim is
Pr(ri = 1) = i = i (xi ) = Ei
exp(xi )
.
1 + exp(xi )
(6.1)
For more discussion, see de Jong and Heller (2008, p. 102); for illustrative SAS code,
see de Jong and Heller (2008, p. 162).
Variations. There are alternative ways of incorporating partial year exposures,
none clearly superior to the others. Equation (6.1) is based on a uniform distribution
of failures within a year. Two other measures are as follows:
1. A constant hazard rate within the year assumption, resulting in i,H (z) = 1 (1
e z Ei
) .
1+ez
2. A hyperbolic assumption (known
as the Balducci assumption for an Italian actuary),
z
resulting in i,B (z) =
e
Ei 1+e
z
e
1(1Ei ) 1+e
z
146
For some applications, the event of a claim is a relatively infrequent event, and
the analyst would like to use all the information available in a claims database. One
may wish to over-sample policyholders with claims; the idea is to draw a larger
proportion of a subset of the population that is of interest in the study. Appendix
Section 6.7 provides details of this type of sampling scheme.
6.2.4 Grouped versus Individual Data
A discussion of large exposures leads naturally into a discussion of grouped versus
individual data.
6.2.4.1 Using Offsets to Handle Large Exposures
To begin, recall that sums of independent Poisson random variables also have a
Poisson distribution. So, when summing random variables from independent policies,
it is sensible to think of exposures as large positive numbers. Thus, it is common to
model the number of accidents per thousand vehicles or the number of homicides per
million population.
For a Poisson distribution, we can use the (logarithmic) number of policies in a
group as an offset variable. Mathematically, if we are thinking about Ei independent
Poisson variables in group i, each with mean i , then the sum will also be Poisson
distributed with mean Ei i . For the Poisson distribution, the variance equals the
mean, so both the mean and the variance grow proportionally to the exposure Ei .
When using a Poisson distribution with a logarithmic link function, one only needs to
specify an offset variable ln Ei to automatically account for the growing variability.
However, for other distributions, this need not be the case. In the GLM linear
exponential family, we saw that the variance can be expressed as a function of the
mean, v(). To be specific, consider the gamma distribution where v() = 2 /
and is a dispersion parameter. If we are thinking about Ei independent gamma
random variables in group i, each with mean and variance , then the sum will
also be gamma distributed with mean Ei and variance Ei . When using a gamma
distribution with a logarithmic link function and offset variable ln Ei , the mean will
grow proportionally to the exposure Ei , but the variability will grow proportionally
to Ei2 , not Ei . So, an offset by itself cannot handle large exposures.
6.2.4.2 Using Variable Scale Parameters to Handle Exposures
For a general distribution in the linear exponential family, suppose that we have
m independent variables from the same distribution with location parameter and
scale parameter . Then, basic arguments given in Section 6.6 show that the sample
average comes from the same distributional family with location parameter and
scale parameter /m. To apply this result, let us consider a problem that analysts
regularly face: the use of grouped versus individual data.
147
To be specific, think about a sample i = 1, . . . , n categories, or groups. For example, each group could be formed by the intersection of the amount of insurance,
territory, and so forth. For the ith group, we have Ei independent observations with
the same distribution from the linear exponential family. It has, for example, location
parameter i , mean i , and scale parameter (that may or may not depend on i). One
could run an analysis with a dataset based on individual observations j = 1, . . . , Ei ,
i = 1, . . . , n.
However, with these assumptions, then the average outcome from the ith group
comes from the same exponential family with the same mean i (or location parameter
i ) but with a scale parameter /Ei . An alternative method of analysis would be to
use the smaller grouped data sample consisting of only n observations, using the
reciprocal of exposure as the weight. The Section 6.6 result guarantees the following:
r Estimates of location/mean parameters (e.g., the regression coefficients) will be the
same. For ratemaking purposes, analysts typically focus on location parameter estimates.
r Only in the case when the scale parameter is known (e.g., binomial, Poisson) would
other inferential aspects (standard errors, t-statistics, p-values, and so forth) be the
same. Individual data analysis provides more accurate estimates of scale parameters
than the corresponding grouped data analysis.
The book is website provides a demonstration of this comparison using the statistical
package R.
6.2.4.3 Large Exposures for Frequency and Severity Models
As noted earlier, for frequency and severity modeling, it is customary to think about
the frequency aspect as proportional to exposure and the severity aspect in terms
of loss per claim. Let us make this advice a bit more specific in the context of an
individual versus grouped analysis.
Suppose that individual data consist of a sample i = 1, . . . , n groups, with
j = 1, . . . , Ei independent observations within each group i. For observation {ij },
the dependent variables consist of (Nij , Sij ): the frequency and total amount of
claims.
If explanatory/rating variables are available at the individual observation level, then
aggregating information up to the group level is problematic because one loses the
information in individual level variables.
Instead, assume that explanatory/rating variables are available only at the group
level and we wish to model aggregate frequency and severity variables {Ni =
n
n
j =1 Nij , Si =
j =1 Sij }.
r For claims frequency, one alternative is to use a Poisson model with the response Ni
and offset ln Ei .
148
r For claims frequency, another alternative is to use a count member from the exponential
distribution family (e.g., binomial) with the response Ni /Ei and scale parameter /Ei .
r For claims severity, use a severity member from the exponential distribution family
(e.g., gamma) with the response Si /Ni and scale parameter /Ni .
As noted earlier, these modeling strategies provide reliable estimates of location
(mean) parameters but not scale parameters. This is a comparative advantage of
analysis with individual level analysis; other advantages include the following:
r Group-level analysis was important before modern day computing and databases
became available. Currently, however, the computing requirements for an individuallevel analysis rarely present a substantial barrier.
r Group-level analysis precludes the examination of individual observations. Often,
a highly unusual observation (outlier) can provide important information to the
analyst.
r The equivalence between the two procedures relies on a number of unverifiable assumptions, including the independence of observations within a group. In some instances,
we can think of reasons for positive associations among observations from the same
category. In this case, the variance of the sum grows faster than linearly, and so specifying the scale parameter as inversely proportional to the exposure may give too large
a weight to categories with large exposure.
149
There is usually overlap in the sets of explanatory variables, where variables are
members of both x1 and x2 . Typically, one assumes that 1 and 2 are not related so
that the joint likelihood of the data can be separated into two components and run
separately.
6.3.1.1 Tobit Models
Another way of modeling a large proportion of zeros is to assume that the dependent
variable is (left) censored at zero. Chapter 19 on survival models provides a more
complete introduction to censored regression. This section emphasizes the application
to two-part data.
With censored regression models, we use an unobserved, or latent, variable y that
is assumed to follow a linear regression model of the form
yi = xi + i .
(6.2)
The responses are censored or limited in the sense that we observe yi = max(yi , 0).
Estimation of this model is typically done by assuming normally distributed disturbances i and using maximum likelihood estimation.
It is straightforward to extend this model to allow for limiting values that vary
by policy. In actuarial applications, we think about di as representing a (known)
deductible that varies by policyholder.
One drawback of the tobit model is its reliance on the normality assumption of
the latent response. A second, and more important, drawback is that a single latent
variable dictates both the magnitude of the response and the censoring. There are
many instances where the limiting amount represents a choice or activity that is
separate from the magnitude. For example, in a population of smokers, zero cigarettes
consumed during a week may simply represent a lower bound (or limit) and may
be influenced by available time and money. However, in a general population, zero
cigarettes consumed during a week can indicate that a person is a nonsmoker, a choice
that could be influenced by other lifestyle decisions (for which time and money may
or may not be relevant).
150
For the second form, we have aggregate counts and severities (Ni , Si ) (or use the
notation yi instead of Si ). Then, the two-step frequency-severity model procedure is
as follows:
r 1. Use a count regression model with Ni as the dependent variable and x1i as the set
of explanatory variables. Denote the corresponding set of regression coefficients as 1 .
Typical models include the Poisson and negative binomial models.
r 2. Conditional on Ni > 0, use a GLM with Si /Ni as the dependent variable and x2i as
the set of explanatory variables. Denote the corresponding set of regression coefficients
as 2 . Typical models include the gamma regression with a logarithmic link and a
dispersion parameter proportional to 1/Ni .
form of dependent variables, we have individual claims yi =
For the third
yi1 , . . . , yi,Ni available. In this case, the first step for the count model is the same.
The second step for severity modeling becomes
2*. Conditional on Ni > 0, use a regression model with yij as the dependent variable and
x2i as the set of explanatory variables. Denote the corresponding set of regression
coefficients as 2 . Typical models include the linear regression (with logarithmic
claims as the dependent variable), gamma regression, and mixed linear models. For
the mixed linear models, one uses a subject-specific intercept to account for the
heterogeneity among policyholders.
151
y 0.
n=1
Because the sum of i.i.d. gammas is a gamma, Sn (not SN ) has a gamma distribution
with parameters n and . Thus, for y > 0, the density of the Tweedie distribution is
fS (y) =
n=1
n n n1 y
y
e .
n! (n)
(6.3)
At first glance, this density does not appear to be a member of the linear exponential family. To see the relationship, we first calculate the moments using iterated
expectations as
E SN =
and
Var SN =
(1 + ).
2
2p
,
(2 p)
2p
,
p1
and
1
= (p 1)p1 .
(6.4)
1
exp S(y, p, ) =
y n=1
y
1/(p1) (2p)(p1)
n!(n)
n
.
Thus, the Tweedie distribution is a member of the linear exponential family. Easy
calculations show that
E SN =
and
Var SN = p ,
(6.5)
where 1 < p < 2. The Tweedie distribution can also be viewed as a choice that is
intermediate between the Poisson and the gamma distributions.
For the Tweedie GLM, we might use xi,T as a set of covariates and T as the corresponding set of regression coefficients. With a logarithmic link, i = exp(xi,T T ). For
the distribution function, there is no closed-form expression, but we could compute
this directly, for example, using the R function ptweedie.
152
i = exp(xi,F F ),
using a logarithmic link function. Here, xi,F is a set of covariates to be used in the
frequency modeling, and F is the corresponding set of regression coefficients.
For the severity, we might use a gamma regression also with a logarithmic link
function. Thus, we would model loss amounts as
= E yij = exp(xi,S S ),
yij gamma(, i ), where
i
for j = 1, . . . , Ni . Similar to frequency, xi,S is a set of covariates to be used in the
severity modeling, and S is the corresponding set of regression coefficients. Thus,
the frequency and severity models need not employ the same set of covariates.
Putting the frequency and severity components together yields the aggregate loss
SN,i = yij + + yi,Ni .
This has mean
E SN,i = E Ni E yij = exp xi,F F + xi,S S
and variance
Var SN,i = i
(6.6)
(6.7)
Note that for frequency-severity modeling, two parameters, i and i , vary with i.
To compute the distribution function, one could use the Tweedie for SN with the R
function ptweedie. This would be done by reversing the relations in (6.4) to get
p=
+2
,
+1
i = i
,
i
and i i = i
(1 + ).
i2
(6.8)
Note that if one begins with the frequency-severity model formulation, the scale
parameter depends on i.
6.4 Application: Massachusetts Automobile Claims
We investigate frequency-severity modeling using an insurance automobile claims
dataset studied in Ferreira and Minikel (2010; 2012). These data, made public by
the Massachusetts Executive Office of Energy and Environmental Affairs (EOEEA),
summarize automobile insurance experience from the state of Massachusetts in year
153
Total
13,905
293
706
14,603
268
685
8,600
153
415
15,609
276
627
14,722
183
549
9,177
96
471
76,616
1,269
3,453
700
700
433
830
814
713
4,190
2,806
18,410
3,104
19,360
1,644
11,245
2,958
20,300
2,653
18,921
1,307
11,764
14,472
100,000
2006. The dataset consists of approximately 3.25 million policies representing more
than a half-billion dollars of claims.
Because the dataset represents experience from several insurance carriers, it is not
surprising that the amount of policyholder information is less than typically used by
large carriers that employ advanced analytic techniques. Nonetheless, we do have
basic ratemaking information that is common to all carriers, including primary driver
characteristics and territory groupings. At the vehicle level, we also have mileage
driven in a year, the focus of the Ferreira and Minikel study.
154
Total
Loss
Claim
Number
Earned
Exposure
Annual
Mileage
Total
With Valid
Annual Miles
115.95
159.67
354.68
187.27
114.14
127.48
0.040
0.055
0.099
0.065
0.038
0.043
0.871
0.894
0.764
0.800
0.914
0.870
12,527
14,406
12,770
13,478
7,611
11,858
76,616
1,269
3,453
4,190
14,472
100,000
69,201
1,149
2,786
3,474
13,521
90,131
youthful drivers are more than three times greater than for adult drivers. We can think
of this total loss as a pure premium.
Table 6.2 shows that the average claim frequency is 4.3%. Specifically, for the
100,000 policies, 95,875 had zero claims, 3,942 had one claim, 176 had two claims,
and 7 had three claims. The table also reports important differences by rating group,
where the average number of losses for inexperienced youthful drivers are about 2.5
times greater than for adult drivers.
Table 6.2 also summarizes information on the earned exposure, defined here as the
amount of time that the policy was in force in the study, and annual mileage. Annual
mileage was estimated by Ferreira and Minikel (2010; 2012) based on Massachusetts
Commonwealths Registry of Motor Vehicles mandatory annual safety checks, combined with automobile information from the vehicle identification number (commonly
known using the acronym VIN). Interestingly, Table 6.2 shows that only about 90%
of our data possess valid information about the number of miles driven, so that about
10% are missing this information.
Table 6.3 provides similar information but by territory. Here, we see that the average
total loss and number of claims for territory 6 are about twice that for territory 1.
Table 6.3. Averages by Territory
Number of Policies
Territory
1
2
3
4
5
6
Total
Total
Loss
Claim
Number
Earned
Exposure
Annual
Mileage
Total
With Valid
Annual Miles
98.24
94.02
112.21
126.70
155.62
198.95
127.48
0.032
0.036
0.037
0.044
0.051
0.066
0.043
0.882
0.876
0.870
0.875
0.866
0.842
0.870
12,489
12,324
12,400
11,962
10,956
10,783
11,858
18,410
19,360
11,245
20,300
18,921
11,764
100,000
16,903
17,635
10,092
18,331
16,944
10,226
90,131
0.00
0.15
0.30
Loss
155
20000
40000
10
Fig. 6.1. Loss distribution. The left-hand panel shows the distribution of loss; the
right-hand panel shows the same distribution but on a (natural) logarithmic scale.
There are 4,125 (= 100,000 95,875) policies with losses. To get a better handle
on claim sizes, Figure 6.1 provides smooth histograms of the loss distribution. The
left-hand panel is in the original (dollar) units, indicating a distribution that is rightskewed. The right-hand panel shows the same distribution on a logarithmic scale
where we see a more symmetric behavior.
Do our rating factors affect claim size? To get some insights into this question,
Figure 6.2 shows the logarithmic loss distribution by each factor. The left-hand panel
shows the distribution by rating group; the right-hand panel shows the distribution by
territory. Neither figure suggests that the rating factors have a strong influence on the
size distribution.
8
6
4
10
By Territory
10
By Rating Group
Fig. 6.2. Logarithmic loss distribution by factor. The left-hand panel shows the
distribution by rating group; the right-hand panel shows the distribution by territory.
156
Negative Binomial
Relativity
(Poisson)
Effect
Estimate
t-Statistic
Estimate
t-Statistic
(Intercept)
2.636
70.92
2.639
69.67
Rating Group
B
I
M
S
0.344
1.043
0.541
0.069
2.85
18.27
8.58
1.49
0.343
1.038
0.539
0.069
2.79
17.64
8.38
1.48
1.41
2.84
1.72
0.93
Territory
1
2
3
4
5
0.768
0.641
0.600
0.433
0.265
14.02
12.24
9.87
8.81
5.49
0.766
0.640
0.598
0.432
0.264
13.79
12.04
9.70
8.64
5.37
0.46
0.53
0.55
0.65
0.77
Notes: Both models use logarithmic exposure as an offset. Estimated negative binomial dispersion parameter is 2.128. Reference levels are A for Rating Group and 6 for Territory.
Table 6.4 summarizes the results from two frequency models: Poisson and negative binomial regression models. For both models, we used a logarithmic link with
logarithmic exposure as an offset variable. Focussing on the Poisson fit, we see that
the t-statistics indicate strong statistical significance for several levels of each factor:
rating group and territory. Additional tests confirm that they are statistically significant factors. Although not reported in Table 6.4, we also ran a model that included
interactions among terms. The interaction terms were statistically insignificant with
a p-value = 0.303 level. Hence, we report on the model without interactions, in part
because of our desire for simplicity.
We also ran an analysis including annual mileage. This variable turned out to
be strongly statistically significant, with a t-statistic equal to 12.08. However, by
including this variable, we also lost 9,869 observations due to missing values in
annual mileage. From the perspective taken in the Ferreira and Minikel (2010; 2012)
study, mileage is the key variable of interest, and so the analyst would wish to retain
this variable. From another perspective, including the mileage variable might result
in analyzing a biased sample; that is, the roughly 10% population without mileage
might differ dramatically from the 90% with mileage. Because of the biased sample
concern, we treat the potential inclusion of the mileage variable as an interesting
follow-up study.
For some audiences, analysts may wish to present the more flexible negative binomial regression model. Table 6.4 shows that there is little differences in the estimated
157
Estimate
t-Statistic
With Number
Estimate
t-Statistic
7.986
137.33
7.909
76.87
Rating Group
B
I
M
S
0.014
0.222
0.013
0.036
0.08
2.49
0.13
0.50
0.020
0.218
0.015
0.038
0.11
2.43
0.15
0.52
Territory
1
2
3
4
5
Claim Number
0.026
0.137
0.004
0.029
0.019
0.31
1.67
0.04
0.38
0.26
0.027
0.138
0.005
0.029
0.018
0.071
0.31
1.68
0.05
0.38
0.23
0.90
2.432
2.445
Notes: Reference levels are A for Rating Group and 6 for Territory.
coefficients for this dataset, indicating that the simpler Poisson model is acceptable
for some purposes. We use the Poisson distribution in our out-of-sample analysis in
Section 6.4.3, primarily because it provides an analytic expression for the predictive
distribution (using the fact that a Poisson sum of independent gammas has a Tweedie
distribution).
Table 6.5 summarizes the fit of a gamma regression severity model. As described
in Section 6.3.2, we use total losses divided by the number of claims as the dependent
variable and the number of claims as the weight. We fit a gamma distribution with a
logarithmic link and the two factors, rating group and territory. Table 6.5 shows small
t-statistics associated with the levels of rating group and territory only inexperienced drivers are statistically significant. Additional tests indicate that the territory
factor is not statistically significant and the rating group factor is marginally statistically significant with a p-value = 0.042. This is an interesting finding.
Table 6.5 also shows the result of using claim number as an explanatory variable in
the severity model. For our data, the variable was not statistically significant and so was
not included in subsequent modeling. Had the variable been statistically significant,
a proxy would need to be developed for out-of-sample prediction. That is, although
we can condition on claim number and it may be a sensible explanatory variable
of (average) severity, it is not available a priori and so cannot be used directly for
out-of-sample prediction.
158
Estimate
(Intercept)
t-Statistic
Relativity
5.356
63.47
Rating Group
B
I
M
S
0.340
1.283
0.474
0.033
1.28
9.39
3.22
0.36
1.41
3.61
1.61
0.97
Territory
1
2
3
4
5
0.743
0.782
0.552
0.480
0.269
6.53
6.92
4.37
4.44
2.50
0.48
0.46
0.58
0.62
0.76
Notes: This model uses logarithmic exposure as an offset. Estimated dispersion parameter is 2.371. Reference levels are A for
Rating Group and 6 for Territory.
159
500 1000
10000
0
Frequency
25000
500
1000
1500
500
1000
1500
Fig. 6.3. Out-of-sample mean performance. The left-hand panel shows the distribution of the out-of-sample predictions calculated using the pure premium, or Tweedie,
model. The right-hand panel shows the strong relationship between the scores from
the frequency-severity and the pure premium models.
and with values of the independent variables from the held-out validation sample. The
predictions for the Tweedie model followed similarly using the coefficients reported
in Table 6.6.
Figure 6.3 compares the predictions for frequency-severity and the pure premium
models. The left-hand panel shows the distribution of our pure premium predictions.
The right-hand panel shows the strong relationship between the two predictions; it
turns out that the correlation is approximately 0.999. For the purposes of predicting
the mean, which is typically the primary focus of ratemaking, these two models yield
virtually indistinguishable predictions. Both models provided some ability to predict
total losses; the (Spearman) correlation between held-out losses and (either) predictor
turned out to be 8.2%.
Because the mean predictor did not provide a way of discriminating between the
pure premium and frequency-severity models, we also looked to tail percentiles.
Specifically, in the Tweedie regression model, in Section 6.4.2 we cited p = 1.5
and = 2.371 and described how to estimate i for each observation i. Then, in
Section 6.3.3 we noted that one could use the pTweedie function in R to get the
distribution function. We did this for each held-out observation and evaluated it
using the actual realized value. Recall the probability integral transform, a result
in probability theory that says that when a continuous random variable is evaluated
using its distribution function, the resulting transformed random variable should have
a uniform (on [0,1]) distribution. Thus, if our distribution function calculation is
approximately correct, then we can expect the held-out transformed random variables
to have an approximate uniform distribution.
The procedure for Poisson frequency is similar to that for gamma severity models,
but is more complex. In Section 6.3.3, we noted that a Poisson sum of gamma random
160
Pure Premium
Frequency-Severity
0.50912
0.85888
0.93774
0.97092
0.99294
0.99528
0.99784
0.42648
0.79766
0.86602
0.90700
0.93948
0.97722
0.99870
variables has a Tweedie distribution. So, even though we estimate the frequency and
severity parameters separately, they can still be combined when we look at the loss
distribution. In display (6.8), we show explicitly how to get Tweedie parameters from
the Poisson frequency and gamma severity models. Then, as with the Tweedie GLM,
we can calculate the transformed (using the distribution function) actual realized
value.
Table 6.7 provides the comparisons for selected percentiles. Both models provide
disappointing results below the 98th percentile; perhaps this is to be expected for a
distribution with approximately 96% zeros. For the 99th percentile and above, the
Tweedie does a good job tracking the actual held-out losses. In comparison, the
frequency-severity approach is only competitive at the 99.9th percentile. On the one
hand, this table suggests that fitting the tails of the distribution is a more complex
problem that requires more refined data and sophisticated models. On the other hand,
the similarity of results in Figure 6.3 when predicting the mean suggests a robustness
of the GLM procedures that gives the analyst confidence when providing recommendations.
161
With this notation, it can be readily shown (e.g., Frees, 2010, chapter 13) that the
moment-generation function can be expressed as
b( + s) b( )
sy
.
M(s; , ) = E e = exp
i=1
%
$
b( + s m ) b()
= M(s; , /m).
= exp
/m
m
"
162
Thus, the sample average is from the same linear exponential family with parameters
and /m.
6.7 Appendix B. Over-Sampling Claims
If you work with government surveys such as the Medical Expenditure Survey (MEPS)
in Chapter 2 or the Survey of Consumer Finances (SCF) in Frees (2010), you will
see that it is common for such surveys to use unequal probabilities when drawing
samples from larger populations. For example, the MEPS data over-sample poor and
minority individuals; the SCF over-samples the wealthy. The idea is to draw a larger
proportion of a subset of the population that is of interest in the study. In insurance, it
is common to over-sample policyholders with claims.
Specifically, consider the two-part model introduced in Section 6.2.1 and presented
in more detail in Section 6.3.1. Suppose that we have a very large database consisting
of {ri , yi , xi }, i = 1, . . . , N observations. We want to make sure to get plenty of
ri = 1 (corresponding to claims or cases) in our sample, plus a sample of ri = 0
(corresponding to nonclaims or controls). Thus, we split the dataset into two pieces.
For the first piece consisting of observations with ri = 1, take a random sample with
probability 1 . Similarly, for the second piece consisting of observations with ri = 0,
take a random sample with probability 0 . For example, we might use 1 = 1 and
0 = 0.2, corresponding to taking all of the claims and a 20% sample of nonclaims.
Thus, the sampling weights 0 and 1 are considered known to the analyst. This
over-sampling procedure is sometimes known as the case-control method.
How does this sampling procedure affect the inference in a two-part model? Think
about this question from a likelihood perspective. To develop the likelihood, let
{si = 1} denote the event that the observation is selected to be included in the sample
and {si = 0} means that it is not included. Suppressing the {i} subscript, we decompose
the likelihood of the dependent variables that can be observed as
f (r, y|s = 1) = f (r|s = 1) f (y|s = 1, r)
observable likelihood = conditional frequency conditional severity.
For the conditional severity, it is common to assume that f (y|s = 1, r) = f (y|r)
given the absence or presence of a claim, the selection mechanism has no effect on
the amount. This is an assumption that may need to be verified, but does seem to
commonly hold.
For the conditional frequency, here are some basic probability calculations to show
how the conditional (on selection) claim frequency relates to the (population) claim
frequency. Conditional on {ri = 1}, we have that Pr(si = 1|ri = 1) = 1 , a Bernoulli
163
Pr(ri = 1|si = 1) =
Now, using the logit form in equation (6.1), we can express the odds-ratio as
Ei
i
Ei
1+exp(xi )
=
=
.
E
i
1 i
1
+
exp(x
)
E
1 1+exp(x
i
i
)
(6.9)
Thus,
Pr(ri = 1|si = 1) =
=
Ei
1 1+exp(x
)Ei
i
Ei
1 1+exp(x
i )Ei
+ 0
1 Ei
1 Ei + 0 (1 + exp(xi ) Ei )
1 Ei
,
ci + 0 exp(xi )
Ei
1 + i exp(xi )
(6.10)
1 Ei
and i = 0 /ci = 1 Ei +00(1Ei ) .
where Ei = 1 Ei /ci = 1 Ei +
0 (1Ei )
In summary, equation (6.10) has the same form as equation (6.1) with a new
definition of exposure and the introduction of an offset term, ln i , assuming that
you use logistic regression (not probit) for your claim frequency modeling. If all of
your exposures are identically equal to 1 (Ei 1), then i is a constant and you simply
reinterpret the constant in the systematic component xi (which we typically ignore).
If exposures are not constant, then equation (6.10) gives a straightforward method of
adjusting the exposure and introducing an offset term, allowing you to run the usual
logistic regression software without the need for specialized software routines.
164
References
Bowers, N. L., H. U. Gerber, J. C. Hickman, D. A. Jones, and C. J. Nesbitt (1997). Actuarial
Mathematics. Society of Actuaries.
Brockman, M. J. and T. S. Wright (1992). Statistical motor rating: making effective use of
your data. Journal of the Institute of Actuaries 119, 457543.
Coutts, S. M. (1984). Motor insurance rating, an actuarial approach. Journal of the Institute
of Actuaries 111, 87148.
Cragg, J. G. (1971). Some statistical models for limited dependent variables with application
to the demand for durable goods. Econometrica 39(5), 829844.
de Jong, P. and G. Z. Heller (2008). Generalized Linear Models for Insurance Data.
Cambridge University Press, Cambridge, UK.
Ferreira, J. and E. Minikel (2010). Pay-as-you-drive auto insurance in Massachusetts: A risk
assessment and report on consumer, industry and environmental benefits. In
Conservation Law Foundation, Boston. http://www.clf.org/wp-content/uploads/2010/
12/CLF-PAYD-Study November-2010.pdf.
Ferreira, J. and E. Minikel (2012). Measuring per mile risk for pay-as-you-drive automobile
insurance. Transportation Research Record: Journal of the Transportation Research
Board 2297, 97103.
Frangos, N. E. and S. D. Vrontos (2001). Design of optimal bonus-malus systems with a
frequency and a severity component on an individual basis in automobile insurance.
ASTIN Bulletin 31(1), 122.
Frees, E., X. Jin, and X. Lin (2013). Actuarial applications of multivariate two-part
regression models. Annals of Actuarial Science 7(2), 258287.
Frees, E., G. Meyers, and A. D. Cummings (2012). Predictive modeling of multi-peril
homeowners insurance. Variance 6(1), 1131.
Kahane, Y. and H. Levy (1975). Regulation in the insurance industry: Determination of
premiums in automobile insurance. Journal of Risk and Insurance 42, 117132.
Klugman, S. A., H. H. Panjer, and G. E. Willmot (2008). Loss Models: From Data to
Decisions. John Wiley & Sons, Hoboken, NJ.
Mullahy, J. (1998). Much ado about two: Reconsidering retransformation and the two-part
model in health econometrics. Journal of Health Economics 17, 247281.
Pinquet, J. (1997). Allowance for cost of claims in bonus-malus systems. ASTIN
Bulletin 27(1), 3357.
Pinquet, J. (1998). Designing optimal bonus-malus systems from different types of claims.
ASTIN Bulletin 28(2), 205229.
Renshaw, A. E. (1994). Modeling the claims process in the presence of covariates. ASTIN
Bulletin 24(2), 265285.
Tweedie, M. C. K. (1984). An index which distinguishes between some important
exponential families. In Statistics: Applications and New Directions. Proceedings of the
Indian Statistical Golden Jubilee International Conference (Editors J. K. Ghosh and
J. Roy), pp. 579604. Indian Statistical Institute, Calcutta.
Weisberg, H. I. and T. J. Tomberlin (1982). A statistical perspective on actuarial methods for
estimating pure premiums from cross-classified data. Journal of Risk and Insurance 49,
539563.
Werner, G. and C. Modlin (2010). Basic Ratemaking (4th ed.). Casulty Actuarial Society.
Part II
Predictive Modeling Methods
7
Longitudinal and Panel Data Models
Edward W. Frees
Chapter Preview. This chapter considers regression methods where the analyst has
the ability to follow a unit of analysis, such as a policyholder, over time. In the
biomedical literature, this type of information is known as longitudinal data and, in
the econometric literature, as panel data.
7.1 Introduction
7.1.1 What Are Longitudinal and Panel Data?
Regression is a statistical technique that serves to explain the distribution of an
outcome of interest (y) in terms of other variables, often called explanatory or
predictor (xs) variables. For example, in a typical ratemaking exercise, the analyst
gathers a cross-section of policyholders and uses various rating variables (xs) to
explain losses (ys).
In this chapter, we assume that the analyst has the ability to follow each policyholder
over time. In many situations, a policyholders past loss experience can provide an
important supplement to the information gleaned from available rating variables. We
use the notation i to represent the unit of analysis (e.g., policyholder) that we will
follow over time t. Using double subscripts, the notation yit refers to a dependent
variable y for policyholder i at time t, and similarly for explanatory variables xit .
Consider four applications that an actuary might face that fall into this framework:
(1) Personal lines insurance such as automobile and homeowners: Here, i represents the
policyholder that we follow over time t, y represents the policy loss or claim, and
the vector x represents a set of rating variables. One could use yi1 > 0 to mean that
a claim has occurred in period 1 this often signals that we are likely to observe a
claim in period 2.
(2) Commercial lines insurance such as commercial auto: Here, i represents the commercial customer (policyholder) and y represents claims per premium (i.e., the loss
167
168
ratio). Without intervening loss reduction measures, a high loss ratio in period 1 might
signal a high loss ratio in period 2.
(3) Insurance sales: Here, i represents a sales agent and y represents annual sales.
Although agent information (e.g., years of experience) and sales territory information
can be useful, often prior sales history is the most important variable for predicting
sales.
(4) Customer retention: In any line of business, the analyst may follow a customer i over
time. We might use yi1 = 1 to indicate that customer i bought a policy in period 1, and
we wish to predict yi2 = 1 or 0, whether or not a customer buys a policy in period 2.
Of course, sometimes actuaries and other analysts attack these and related problems
without a formal background in statistics, using well-established benchmarks and
procedures that have stood the test of time. For example, in personal lines auto
insurance, it is common to use the presence of a claim to indicate a rate increase by
omitting the safe driver discount. In commercial lines, actuaries have a long history
of incorporating past claims experience using credibility techniques. In contrast,
statistical methods provide a disciplined way to reexamine and reevaluate old methods
and can also be brought to bear on emerging problems.
7.1 Introduction
Year 1
Years 1 and 2
Loss
60
60
70
70
80
80
90
90
100
100
110
110
Loss
169
60
65
70
Rating Variable
75
60
65
70
Rating Variable
75
Fig. 7.1. Loss and rating variable. The left-hand panel shows the positive period 1
relationship between the loss and a rating variable. The right-hand panel shows the
loss and rating variables for both periods 1 and 2, with lines connecting the periods for
each policyholder. Most of the lines have a negative slope, indicating that increases
in the rating variable result in a decrease in the loss variable.
In contrast, the right-hand panel of Figure 7.1 shows the plot of the rating variable
and the loss for the combined years 1 and 2, but with a line connecting year 1 and year
2 results for each policyholder. The line emphasizes the dynamic effect, moving from
the year 1 rating variable to the year 2. To interpret this plot, the data were constructed
so that the year 1 rating variable is generally smaller than the year 2 result, so you
can think of the change over time as moving from left to right. Most of the lines have
a negative slope, indicating that increases in the rating variable result in a decrease
in the loss variable. That is, the overall dynamic effect is negative, in contrast to the
positive static effect.
How can these data happen? We can think of one scenario in which the rating
variable naturally increases from year 1 to year 2 due to inflation. Further, suppose
that an unobserved loss reduction measure has been introduced that serves to reduce
expected losses for all policyholders. This would mean that each policyholder could
expect to have an increase in the horizontal x axis and a decrease in the vertical y
axis, resulting in a negative slope.
For another scenario, suppose that x represents a loss prevention measure such
as a burglar alarm, and y represents theft losses in a home or commercial building.
170
We might interpret the left-hand panel as both x and y being positively related to
an unobserved variable such as the safety of the home/buildings neighborhood.
That is, in very safe neighborhoods, theft losses y tend to be low, and expenditures
on burglar alarms x tend to be low (why pay a lot for a burglar alarm in such
a safe neighborhood?) and, conversely for unsafe neighborhoods, resulting in the
overall positive slope. However, for each home or building, the introduction of a more
extensive burglar alarms means that expenditures x increase while expected losses y
tend to decrease.
Regardless of the scenario or model generating the data, the important point is
that static analysis without paying attention to temporal effects gives a grossly biased
inference about the effects of the rating variable on the losses.
7.1.2.2 Efficiency and Sharing of Information
Example 7.1 illustrates a potential type of bias that one can encounter if the analyst
does not think carefully about the longitudinal or panel nature of the data. Bias is a
serious concern as we saw in Example 7.1, the static relationship summarizing the
rating variables effect on losses (positive) displayed the opposite tendency to that of
the dynamic relationship (negative).
Analysts also use longitudinal and panel data techniques because they wish to
optimize the use of information by promoting efficient inference and sharing of
information among different quantities of interest.
Suppose that observations from different years are independent of one another.
Even in this case, it is clear that an analyst would prefer two years of information
about, for example, the relationship between a rating and a loss variable. Having
two years of information, as compared to one, enables the analyst to provide more
precise (efficient) estimates of parameters that summarize this relationship, with
correspondingly smaller standard errors. Having multiple years of information allows
the analyst to make efficient estimates about loss prediction for each policyholder,
which is the goal of credibility theory.
However, it is more common for longitudinal and panel data to exhibit features
of clustering, where observations from the same unit of analysis tend to be similar
or close to one another in some sense. Referring to Example 7.1, it is commonly
observed in insurance claims that yi2 tends to be related to yi1 ; that is, year 2 claims are
related to year 1 claims for the same policyholder. By recognizing and incorporating
this relationship into our models, we can (i) develop more efficient estimators and (ii)
use the information in prior years (e.g., year 1) to predict current year (e.g., year 2)
losses.
Example 7.2 (Bivariate Normal Example). It is helpful to assume normality to reaffirm basic concepts. Suppose that we have a sample of size n of logarithmic losses
7.1 Introduction
171
from two years yi = (yi1 , yi2 ) that we assume are bivariate normal. That is, yit is normally distributed with mean it = 0 + 1 xit and variance t2 , for years t = 1 and
t = 2, and is the correlation between yi1 and yi2 . Here, xi1 and xi2 are known rating
variables. Assume that we have a sufficiently large sample size n so that estimation
errors of the parameters 0 , 1 , 12 , 22 , and are not a significant issue.
If we want to predict year 2 losses given the information in year 1, standard
probability theory tells us that the conditional distribution is normal. Specifically, we
have
2
2
2
yi2 |yi1 N i2 + (yi1 i1 ) , 2 (1 ) .
1
Without information about yi1 , the optimal predictor of yi2 is its mean, i2 =
0 + 1 xi2 . However, with information about prior year losses, yi1 , we can do better.
The optimal predictor of yi2 given yi1 is its conditional mean, i2 + 21 (yi1 i1 ).
For the conditional predictor, the stronger the relationship between the two years,
the larger is the value of , and the smaller is the variance of the conditional
distribution.
The conditional predictor outperforms the original (marginal) predictor because
we are sharing information over the two years through the correlation parameter
. Of course, this improved performance relies on sufficient knowledge of the year
2 parameters, which is an empirical question. Nonetheless, this example opens the
gateway so that you can see how this type of information might improve standard
cross-sectionalbased prediction.
172
Looking at the data this way, we see that longitudinal data analysis can be thought of
a special case of
regression, where each subject features a multivariate
multivariate
response yi = yi1 , . . . , yiTi .
The term panel study was coined in a marketing context when Lazarsfeld and
Fiske (1938) considered the effect of radio advertising on product sales. Traditionally,
hearing radio advertisements had been thought to increase the likelihood of purchasing
a product. Lazarsfeld and Fiske considered whether those who bought the product
would have been more likely to hear the advertisement, thus positing a reverse in
the direction of causality. They proposed repeatedly interviewing a set of people (the
panel) to clarify the issue.
Baltes and Nesselroade (1979) trace the history of longitudinal data and methods
with an emphasis on childhood development and psychology. They describe longitudinal research as consisting of a variety of methods connected by the idea that the
entity under investigation is observed repeatedly as it exists and evolves over time.
Moreover, they trace the need for longitudinal research to at least as early as the 19th
century.
7.2 Linear Models
Longitudinal and panel data consist of repeated observations of a subject. To see the
variety of ways to represent this type of data, we begin by working in a traditional
linear model framework, where mean effects are linear in the parameters. Naturally,
many actuarial applications fall into the nonlinear model context, but the linear model
framework provides a convenient pedagogic framework on which we can base our
discussions.
For period 1, think about a sample of the form {yi1 , xi1 } where we have a dependent
variable of interest (y) and a collection of variables (x) that could be used to explain y.
This is the usual regression set-up. For our extension to longitudinal and panel data,
we have a similar sample for the other periods of the form {yit , xit }. Unlike some
surveys (known as repeated cross-sections), we can link the i across periods.
Linear Model 1. Cross-Sectional Model. Here, we assume that all observations are
independent and have a common variance Var yit = 2 and regression function:
E yit = + 1 xit,1 + 2 xit,2 + + k xit,k
= + xit .
This model is helpful because it is regularly used as the strawman in any panel
data analysis, one that can be easily defeated by using more sophisticated techniques.
However, it is useful as a benchmark because many consumers are familiar and
comfortable with cross-sectional regression analysis.
173
The model is useful in some isolated cases. A good illustration of this point is
the capital asset pricing model, known by the acronym CAPM, of stock returns.
In this application, i represents a firm whose stock price return, yit , is followed over
time t. (Sometimes it is the return in excess of the risk-free rate.) The only explanatory
variable is the return based on a market index. Essentially, the argument from financial
economics is that any patterns in the errors would be discovered, taken advantage of,
and disappear in a liquid market. This model, without clustering or other relationships
over time, is a good representation in this application.
Linear Model 2. Fixed Effects Model. As with Linear Model 1, assume independence
among all observations and a common variance Var yit = 2 . The difference is in the
regression function
E yit = i + xit ,
where the intercept parameter has a subscript i attached to it this means that each
subject has its own parameter. Intuitively, the parameter i is meant to capture all of
the effects
that are not observable in the other explanatory variables
for the ith subject
xit = xit,1 , . . . , xit,k . With at least one parameter for each subject, you can see
where repeated observations are necessary to estimate parameters of this model; in
technical terms, this model is not identifiable with cross-sectional data where Ti 1.
The model is popular in part because it is easy to estimate with commonly available
software. Think of the subject as a categorical variable, a factor, and so the varying
intercepts i correspond to different levels of the factor. One can replace categorical
variables with an appropriate set of binary variables; for this reason, panel data
estimators are sometimes known as least squares dummy variable model estimators.
This is known as a fixed effects model because the quantities i are fixed,
although unknown to the analyst, parameters to be estimated. This is contrast to the
next model where we interpret the quantities to be random variables.
Linear Model 3. Random Effects Model. To allow intercepts to be random, instead
of writing down only the regression function, consider the model equation
yit = i + xit + it ,
where it is an identically and independently distributed error, or disturbance, term
with mean zero and variance 2 . In this random effects formulation, the intercept
i is a random variable, independent of {it }.
The random effects model is commonly used in actuarial practice (although not by
this name). As seen in Chapter 9, it is the basis of widely used credibility formulas
for prediction.
Both the fixed and random effects models account for the clustering through common intercept parameters. For example, both yi1 and yi2 are based in part on the
174
175
Coverage (000s)
Claims (000s)
Logarithmic Coverage
Logarithmic Claims
Mean
Median
Standard
Deviation
Minimum
Maximum
30,277
14.724
16.272
8.029
11,545
5.744
16.262
8.656
54,901
32.517
1.426
2.710
25
0
10.145
0
427,727
290.206
19.874
12.578
176
1 43 41 13 29 25 101 3
6 26
12
10
8
6
4
2
Logarithmic Claims
44 85 89 79 20 82 96 34 45 36 54 38 95 60 66 17 31 11 98 93 75 55 4 12 50 74 59 81 65
12
10
8
6
4
2
91 104 22 87 94 77 103 53 106 84 83 67 51 64 56 47 10 49 102 86 28 92 78 16 88 73 33 40 68
12
10
8
6
4
2
Year
Fig. 7.2. Trellis plot of logarithmic group term life claims from 19931996.
0.884
the intercept is 8.8657
= 12.420, and the slope for logarithmic coverage is 10.303
=
10.303
1.268, consistent with the other models. The t-statistic for the lagged dependent
variable model is 5.35, indicating strong statistical evidence of serial autocorrelation.
For the correlated errors model, the estimate of auto-covariance turns out to be =
0.351 which is qualitatively similar to findings from the lagged dependent variable
model. For the model fits, the cross-sectional model turns out to have an R 2 = 0.434
goodness-of-fit statistics. The corresponding measure for the fixed effects model
is R 2 = 0.711, a statistically significant improvement. Additional examination of
model diagnostics (not displayed here) shows that all four longitudinal data models
are superior to our strawman, the ordinary cross-sectional regression model.
177
Table 7.2. Summaries of Five Model Fits to the Group Term Life Data
Intercept
t-statistic
Logarithmic Coverage
t-statistic
Lagged Claims
t-statistic
CrossSectional
Fixed
Effects
Models
Random
Effects
12.337
9.49
1.252
15.73
12.286
1.65
1.264
3.04
12.882
7.48
1.282
12.14
Lagged
Dependent
Correlated
Errors
8.657
5.80
0.884
8.36
0.303
5.35
12.567
7.73
1.265
12.69
model context. This section provides a roadmap of how to think about modeling
choices when your data cannot be reasonably be represented using a linear model.
178
Intuitively, our ability to estimate global parameters that are common to all subjects
is corrupted by the inability to estimate the subject-specific effects i . In contrast, in
the linear model context, the estimation procedure sweeps out the intercept terms
when producing estimates of , producing reliable estimates even when the estimates
of i are not.
7.3.1.2 Markov Transition Models
In the biomedical and econometrics literatures, the random effects model is the most
common method of analyzing longitudinal binary outcomes. However, in actuarial
applications, it is helpful to think about Markov transition modeling. With these models, actuaries can account for persistency by tracing the development of a dependent
variable over time and representing the distribution of its current value as a function of
its history. In this context, think about the events {y = 1} and {y = 0} as representing
two states. Persistency connotes the tendency to remain in a state over time. It is
the same idea as clustering, yet applied to a state space context.
To be more precise about the development over time, define Hit to be the history
of the ith subject up to time t. For example, if the explanatory variables are assumed
to be nonstochastic, then we might use Hit = {yi1 , . . . , yi,t1 }. With this information
set, we may partition the likelihood for the ith subject as
f (yit )
Ti
"
f (yit |Hit ),
t=2
where f (yit |Hit ) is the conditional distribution of yit given its history and f (yi1 ) is
the marginal (unconditional) distribution of yi1 .
For a Markov model (of order 1), one assumes that the entire history is captured
by the most recent outcome, so that
f (yit |Hit ) = f (yit |yi,t1 ).
For binary outcomes, we can write the conditional probability of a 1, given yi,t1 = 0,
as
it,0 = Pr yit = 1|yi,t1 = 0 = (0 + xit 0 )
and, given yi,t1 = 1, as
it,1 = Pr yit = 1|yi,t1 = 1 = (1 + xit 1 ).
In this context, it,0 and it,1 are examples of transition probabilities, quantifying
the probability of moving or transiting from one state to another. Here, we have
used one set of parameters {0 , 0 } when the origin (determined by yi,t1 ) state
is 0 and another set, {1 , 1 }, when the origin state is 1. Whether this is appropriate
179
is an empirical question, although many datasets are rich enough to support this
parameterization.
With this notation, the conditional distribution is given as
Ti
i
ln f (yit |yi,t1 ).
t=2
As with the lagged dependent variable linear model, this modeling choice does lose
the period 1 observations; in this sense, it is a partial likelihood. For many problems
of interest, the interesting part is estimating the transition probabilities (e.g., what is
the probability of not retaining a customer?), and so one loses little by focusing on
the partial likelihood.
180
The general random effects GLM model has the same limitations discussed in the
special case of binary outcomes:
r This is a computationally intensive formulation that is tractable only with modern-day
software and hardware.
r A fixed effects version is often not a reliable alternative.
See Chapter 16 for further discussion.
7.3.2.2 Categorical Outcomes and Markov Transition Models
Instead of only two (binary) outcomes, actuaries may want to model transitions to
more than two states. For example, one might want to follow an automobile driver
over time and assess whether the driver belongs to one of the three states: preferred,
standard, or substandard. As another application, one might want to follow a bond
over time and model how it moves among ratings AAA, AA, A, and so forth. On
the one hand, it is possible to extend the generalized linear model framework to handle
categorical outcomes using multinomial logits. On the other hand, Markov transition
modeling provides greater flexibility and is easier to explain and to implement. See
Chapter 20 for further discussion.
References
181
these techniques can be applied to other sampling schemes that do not have a time
element. The key is the replication of observations, not the observation over time. For
example, you may organize your unit of analysis i to represent different geographic
regions of a country (e.g., counties). Then, there may be several policies within each
county, a type of replication without a time element.
Extending this line of thought, it is possible to think about several layers of a hierarchy. For example, in commercial automobile lines, one might consider a customer as
the unit of analysis that we follow over time. For each customer, there may be several
vehicles types insured (e.g., trucks, automobiles, motorcycles), and within each type,
several vehicles. This type of analysis is known as multilevel modeling.
7.5 Further Reading
The ideas in this chapter are expanded on in the book-length treatment of Frees (2004).
Readers may also wish to refer to Diggle et al. (2002) and Hsiao (2002), both of which
provide excellent introductions to longitudinal and panel data analysis.
References
Baltes, P. B. and J. R. Nesselroade (1979). History and rational of longitudinal research. In
P. B. Baltes and J. R. Nesselroade (Eds.), Longitudinal Research in the Study of
Behavior and Development. Academic Press, New York.
Diggle, P. J., P. Heagarty, K. Liang, and S. L. Zeger (2002). Analysis of Longitudinal Data
(2nd ed.). Oxford University Press, Oxford.
Frees, E. W. (2004). Longitudinal and Panel Data: Analysis and Applications in the Social
Sciences. Cambridge University Press, Cambridge.
Hsiao, C. (2002). Analysis of Panel Data (2nd ed.). Cambridge University Press, Cambridge.
Lazarsfeld, P. F. and M. Fiske (1938). The panel as a new tool for measuring opinion. Public
Opinion Quarterly 2, 596612.
8
Linear Mixed Models
Katrien Antonio and Yanwei Zhang
Chapter Preview. We give a general discussion of linear mixed models and continue
by illustrating specific actuarial applications of this type of model. Technical details on
linear mixed models follow: model assumptions, specifications, estimation techniques,
and methods of inference. We include three worked-out examples with the R lme4
package and use ggplot2 for the graphs. Full code is available on the books website.
(8.1)
183
This model completely pools the data, ignores the panel structure, and produces
identical estimates for all subjects i (for a given xit ). The linear fixed effects model
(Linear Model 2 in Section 7.2) specifies
E yit = i + xit ,
(8.2)
where each subject i has its own unknown but fixed intercept i . Hence, the
name fixed effects model. Independence among all observations is assumed, and
Var(yit ) = 2 . This regression model does not pool information and estimates each i
separately using least squares or maximum likelihood. This approach often results in
overfitting and unreasonable i s (see Gelman 2006). The linear random effects model
(see Linear Model 3 in Section 7.2) is an alternative approach, balancing between no
pooling and complete pooling of data. It allows for random intercepts, with model
equation
yit = i + xit + it ,
(8.3)
where it (0, 2 ).1 The subject specific intercept i is now a random variable with
zero mean and variance 2 . Hence, the name random effects model. Moreover, the
model in (8.3) is a first example of a linear mixed model (LMM), with a combination
(mix) of fixed and random effects in the linear predictor. The errors it with variance
2 structure represent variability within subject i, whereas the random intercepts with
variance 2 represent variation between subjects. Compared with the no pooling and
complete pooling examples, the linear mixed model has many interesting features, as
explained later.
Mixed or Multilevel Models for Clustered Data. Panel data are a first example of
so-called clustered data. As mentioned in Section 7.4, predictive modeling in actuarial
science (and in many other statistical disciplines) is based on data structures going
beyond the cross-sectional as well as panel data design. Section 8.3 in this chapter
includes multiple motivating examples. Mixed (or multilevel) models are statistical
models suitable for the analysis of data structured in nested (i.e., hierarchical) or nonnested (i.e., cross-classified, next to each other instead of hierachically nested) clusters
or levels. In this chapter we explain the use of linear mixed models for multilevel
data. A discussion of nonlinear mixed models is found in Chapter 16. Chapters 7 (on
longitudinal and panel data), 9 (on credibility), and 11 (on spatial statistics) include
additional examples of clustered data and their analysis with mixed models.
Textbook Examples. A standard textbook example of multilevel data is the students
in schools data structure. Extended versions are the students in classes in schools
or students followed repeatedly over time, in classes in schools examples, where
1
184
each example adds an extra level of observations to the data hierarchy. Reflecting the
actuarial audience of this book, we consider the example of a collection of vehicles
j (with j = 1, . . . , ni ) insured under fleets i (with i = 1, . . . , m). Let yij be the loss
observed for vehicle j in fleet i (in a well defined period of exposure). Denote with
x1,ij covariate information at vehicle-level (our level 1): x1,ij is, for example, the cubic
capacity or vehicle age of car j in fleet i. x2,i is a predictor at fleet-level (our level 2)
and could, for example, refer to the size of the fleet or the business in which the fleet is
operating. The socalled varying intercepts model is a basic example of a multilevel
model. It combines a linear model at vehicle-level (i.e., level 1),
yij = i + 1,0 + x1,ij 1,1 + 1,ij ,
j = 1, . . . , ni ,
(8.4)
i = 1, . . . , m,
(8.5)
i = 1, . . . , m.
(8.6)
Here 2,i (0, 22 ) and 1,ij (0, 12 ) are mean zero, independent error terms, representing variability (or heterogeneity) at both levels in the data. Written as a single
model equation, the combination of (8.4) and, for example, (8.5), is
yij = 1,0 + 2,i + x1,ij 1,1 + 1,ij .
(8.7)
This regression model uses an overall intercept, 1,0 ; a fleet-specific intercept, 2,i ;
a vehicle-level predictor x1,ij with corresponding regression parameter, 1,1 ; and an
error term 1,ij . We model the fleet-specific intercepts, 2,i , as random variables,
which reflects heterogeneity among fleets in an efficient way, even for a large number
of fleets. Indeed, by assigning a distribution to these error terms, we basically only
need an estimate for the unknown parameters (i.e., the variance component 22 ) in
their distribution. The other regression parameters, 1,0 and 1,1 , are considered fixed
(in frequentist terminology); we do not specify a distribution for them. The model in
(8.7) is again an example of a linear mixed model (LMM). Mixed refers to the
combination of fixed and random effects, combined in a model specification that is
linear in the random (2,i ) and in the fixed effects (1,0 and 1,1 ).
Allowing for varying slopes and intercepts results in the following model equations:
yij = i,0 + x1,ij i,1 + 1,0 + x1,ij 1,1 + 1,ij ,
(8.8)
i = 1, . . . , m, j = 1, . . . , ni , with
i,0 = 2,i,0 ,
i,1 = 2,i,1 .
(8.9)
185
(8.10)
In addition to having random intercepts (2,i,0 ), the model also allows the effect of
predictor x1,ij on the response to vary by fleet. This is modeled here by the random
slopes 2,i,1 .
Main Characteristics and Motivations. The examples of varying intercepts and
varying slopes reveal the essential characteristics of a multilevel model: (1) varying
coefficients and (2) a regression model for these varying coefficients (possibly using
group-level predictors). Motivations for using multilevel modeling are numerous (see
Gelman and Hill 2007); we illustrate many throughout this chapter. Because data
often are clustered (e.g., students in schools, students in classes in schools, cars in
fleets, policyholder data over time, policies within counties), statistical methodology
should reflect the structure in the data and use it as relevant information when building statistical models. Using traditional (say linear or generalized linear models, as
in Chapters 2 and 5) regression techniques, the clustering in groups is either ignored
(complete pooling) or groups are analyzed separately (no pooling), resulting in overfitting because even small clusters will get their own regression model. The multilevel
model enhances both extremes: for example, in the varying intercepts model from
(8.7), complete pooling corresponds with 22 0, and no pooling corresponds with
22 . Multilevel modeling is a compromise between these two extremes, known
as partial pooling. In this case, we impose a distributional assumption on 2,i (with
variance 22 ) and estimate 22 from the data. This takes heterogeneity among clusters into account, making appropriate cluster-specific predictions and structuring the
dependence between observations belonging to the same cluster. Moreover, predictions related to new clusters become readily available. Whereas in classical regression
cluster-specific indicators cannot be included along with cluster-specific predictors,
multilevel models allow doing this in a convenient way (see (8.6)). When specifying
regression models at different levels in the data, interactions between explanatory
variables at different levels (so-called crosslevel effects) may appear. This feature
is often mentioned as another advantage of multilevel models.
Whats in a Name?: Labels and Notation. Multilevel models carry many labels
in statistical literature. They are sometimes called hierarchical, because data are
often hierarchically structured (see the students in schools example) and because of
the hierarchy in the model specifications. However, non-nested models, with levels
structured next to each other instead of being hierarchically nested, can also be
analyzed with the multilevel methodology. Multilevel models are also known as
random effects or mixed models, because they combine (a mix of) fixed and random
effects. This distinction is only applicable when using frequentist methodology and
186
i = 1, . . . , N
j = 1, . . . , m,
(8.11)
and
yi = j [i] + j [i] xi + i ,
j = a0 + b0 uj + j 1 ,
i = 1, . . . , N
j = 1, . . . , m
j = j 2 .
(8.12)
Observations in the dataset are indexed with i, where N is the total number of
observations j denotes the fleets in the dataset, and j [i] is the fleet to which observation
i belongs; xi refers to covariate information available at vehicle-level (i.e., level 1 in
(8.4)); and uj refers to covariate information available at fleet-level (i.e., level 2
in (8.6)).
The notation used from Section 8.2 on is motivated by generality and is inspired
by Frees (2004). This notation allows model equations to be written in a structured
way, with clear reference to the particular level in the data to which the parameter/predictor is attached. Moreover, this notation can be used for any number of levels
in a concise way. Section 8.2 explains the connection between this particular notation
and the matrix notation (and corresponding manipulations) that is often developed
in statistical literature on mixed models. When discussing examples, we replace this
general notation with a more intuitive one, explicitly referring to the structure of the
data under consideration.
187
tool in actuarial science, with applications ranging from ratemaking to reserving and
smoothing. We deploy some of these examples within the framework of linear mixed
models, whereas others are more appropriate for analysis with generalized linear
mixed models.
Example 8.1 (Credibility Models). Credibility theory is an a posteriori ratemaking
technique. Credibility models are designed to predict an insureds risk premium by
weighting the insureds own loss experience and the experience in the overall portfolio. An extensive discussion of credibility models is available in Chapter 9. Credibility
models have a natural and explicit interpretation as special examples of mixed models.
Frees, Young, and Luo (1999) demonstrate this connection by reinterpreting credibility
models using mixed model parlance. This mapping greatly increases the accessibility
and usefulness of such models. Indeed, the complete machinery (including computational methods and software) of mixed models becomes available for the analysis of
these actuarial models. The famous Hachemeister dataset (see Hachemeister 1975)
has often been used in credibility literature. This dataset considers 12 periods, from
the third quarter of 1970 to the second quarter of 1973, of bodily injury losses covered
by a private passenger auto insurance; it registers the total loss and corresponding
number of claims for five states. Figure 8.1 shows a trellis plot of the average loss
per claim, followed over time, per state. The plot also shows a linear regression line
and corresponding confidence intervals (in gray). In Section 8.3 we use linear mixed
models to analyze this dataset and predict the next years average claim per state.
Further analysis with focus on credibility theory follows in Chapter 9.
Example 8.2 (Workers Compensation Insurance: Losses). The dataset is from
the National Council on Compensation Insurance (United States) and contains losses
due to permanent partial disability (see Klugman 1992); 121 occupation or risk classes
are observed over a period of seven years. The variable of interest is the Loss paid
out (on a yearly basis) per risk class. Possible explanatory variables are Year and
Payroll. Frees, Young, and Luo (2001) and Antonio and Beirlant (2007) present
mixed models for the pure premium, PP=Loss/Payroll. For a random subsample of
10 risk classes, Figure 8.2 shows the time series plot of Loss (left) and corresponding
Payroll (right).
Example 8.3 (Workers Compensation Insurance: Frequencies). The data are
from Klugman (1992); see Scollnik (1996), Makov et al. (1996), and Antonio and
Beirlant (2007) for further discussion). Frequency counts in workers compensation
insurance are observed on a yearly basis for 133 occupation classes followed over 7
years. Count is the response variable of interest. Possible explanatory variables are
Year and Payroll, a measure of exposure denoting scaled payroll totals adjusted for
188
1
Fig. 8.1. Trellis plot of average losses per period and a linear regression line with
corresponding confidence intervals (in gray); each panel represents one state. Source:
Hachemeister (1975).
inflation. Figure 8.3 shows exploratory plots for a random subsample of 10 occupation
classes. Statistical modeling should take into account the dependence between observations on the same occupation class and reflect the heterogeneity among different
classes. In ratemaking (or tariffication) an obvious question for this example would
be: What is the expected number of claims for a risk class in the next observation
period, given the observed claims history of this particular risk class and the whole
portfolio?
Example 8.4 (Hierarchical Data Structures). With panel data a group of subjects
is followed over time, as in Examples 8.2 and 8.3. This is a basic and widely studied
example of hierarchical data. Obviously, more complex structures may occur. Insurance data often come with some kind of inherent hierarchy. Motor insurance policies
grouped in zip codes within counties within states are one example. Workers compensation or fire insurance policies operating in similar industries or branches is another
one. Consider the manufacturing versus education branch, with employees in manufacturing firms indicating larger claims frequencies, and restaurants versus stores,
189
1.25e+08
1.00e+08
Payroll
Loss
2e+06
1e+06
7.50e+07
5.00e+07
2.50e+07
0e+00
0.00e+00
2
4
Year
4
Year
Fig. 8.2. Time series plot of losses (left) and payroll (right) for a random sample of
10 risk classes: workers compensation data (losses).
with restaurants having a higher frequency of fire incidents than stores, and so on. A
policyholder holding multiple policies (e.g., for theft, motor, flooding), followed over
time, within the same company, is an example of a hierarchical data structure studied
in the context of multidimensional credibility (see Buhlmann and Gisler 2005).
Another detailed multilevel analysis (going beyond the panel data structure) is
Antonio, Frees, and Valdez (2010). These authors model claim count statistics for
vehicles insured under a fleet policy. Fleet policies are umbrella-type policies issued
to customers whose insurance covers more than a single vehicle. The hierarchical or
multilevel structure of the data is as follows: vehicles (v) observed over time (t), nested
within fleets (f ), with policies issued by insurance companies (c). Multilevel models
allow for incorporation of the hierarchical structure of the data by specifying random
effects at vehicle, fleet, and company levels. These random effects represent unobservable characteristics at each level. At the vehicle level, the missions assigned to a
5000
100
4000
Payroll
Count
75
50
25
3000
2000
1000
0
2
4
Year
4
Year
Fig. 8.3. Time series plot of counts (left) and payroll (right) for a random sample of
10 risk classes: workers compensation data (counts).
190
Table 8.1. Number of Payments (left) and Average Loss Per Combination of Status
and Experience Risk Class: Credit Insurance Data
Experience
Status
1
2
3
Experience
status
40
54
39
43
53
39
41
48
44
1
2
3
180.39
172.05
212.30
246.71
232.67
269.56
261.58
253.22
366.61
vehicle or unobserved driver behavior may influence the riskiness of a vehicle. At the
fleet level, guidelines on driving hours, mechanical check-ups, loading instructions,
and so on, may influence the number of accidents reported. At the insurance company
level, underwriting and claim settlement practices may affect claims. Moreover, random effects allow a posteriori updating of an a priori tariff, by taking into account the
past performance of vehicle, fleet, and company. As such, these models are relevant
for a posteriori or experience rating with clustered data. See Antonio et al. (2010) and
Antonio and Valdez (2012) for further discussion.
Example 8.5 (Non-Nested or Cross-Classified Data Structures). Data may also
be structured in levels that are not nested or hierarchically structured, but instead act
next to each other. An example is the dataset from Dannenburg, Kaas, and Goovaerts
(1996) on private loans from a credit insurer. The data are payments of the credit
insurer to several banks to cover losses caused by clients who were no longer able to
pay off their loans. These payments are categorized by civil status of the debtors
and their work experience. The civil status is single (1), divorced (2), or other (3),
and the work experience is less than 2 years (<2, category 1), from 2 up to 10 years
(2 and <10, category 2), and more than 10 years (10, category 3). Table 8.1 shows
the number of clients and the average loss paid per risk class. Boxplots of the observed
payments per risk class are in Figure 8.4. Using linear mixed models we estimate the
expected loss per risk category and compare our results with the credibility premiums
derived by Dannenburg et al. (1996).
Example 8.6 (Loss Reserving). Zhang, Dukic, and Guszcza (2012) analyze data
from the workers compensation line of business of 10 large insurers, as reported
to the National Association of Insurance Commissioners.2 Common accident years
available are from 1988 to 1997. Losses are evaluated at 12-month intervals, with
the highest available development age being 120 months. The data have a multilevel
2
191
800
Payment
600
400
200
0
1:1 1:2 1:3 2:1 2:2 2:3 3:1 3:2 3:3
Status:Experience
Fig. 8.4. Boxplots of payments versus combination of status and experience: credit
insurance data.
10
Comp #7
Comp #3
Comp #8
12
36
60
84
10
8
Comp #2
Comp #6
12
36
60
84
Cumulative loss
Comp #1
Comp #4
Comp #9
Comp #5
Comp #10
1
2
3
4
5
6
7
8
12
36
60
84
10
8
12
36
60
84
10
12
36
60
84
10
structure with losses measured repeatedly over time, among companies and accident
years. A plot of the cumulative loss over time for each company clearly shows a
nonlinear growth pattern (see Figure 8.5). Predicting the development of these losses
beyond the range of the available data is the major challenge in loss reserving. Figure 8.5 reveals that the use of a nonlinear growth curve model is an interesting path
to explore. Random effects are included to structure heterogeneity among companies
and accident years.
Evaluation time
Fig. 8.5. Observed growth of cumulative losses for the 10 companies in the study.
The colored lines represent accident years.
192
(8.13)
with an N 1 vector of error terms with covariance matrix (see later discussion
for examples), which is independent of u. This is the hierarchical specification of a
linear mixed model. For a given u, the conditional mean and variance are
E[ y|u] = X + Zu,
Var[ y|u] = .
(8.14)
(8.15)
showing that fixed effects enter the (implied) mean of Y and random effects structure
the (implied) covariance matrix of y.
Usually, normality is assumed for u and , thus,
u
0
D 0
N
,
.
(8.16)
0
0
193
(8.17)
This implies the marginal model y N(X, V ), but not vice versa. When interest is
only in the fixed effects parameters , the marginal model can be used. When there is
explicit interest in and u, the specification in (8.13) and (8.17) should be used.
Examples 8.7 and 8.8 focus on particular examples of two and three level data and
explain in detail the structure of vectors and matrices in (8.13) and (8.15).
Example 8.7 (A Two-Level Model for Longitudinal Data.). Yij represents the j th
measurement on a subject i (with i = 1, . . . , m and j = 1, . . . , ni ); m is the number
of subjects under consideration; and ni the number of observations registered on
subject i. xij (p 1) is a column vector with fixed effects covariate information
from observation j on subject i. Correspondingly, zij (q 1) is a column vector
with covariate information corresponding with random effects. (p 1) is a column
vector with fixed effects parameters, and ui (q 1) is a column vector with random
effects regression parameters. These are subject-specific and allow one to model
heterogeneity between subjects. The combined model is
yij = xij + zij ui + ij .
fixed
random
(8.18)
random
G Rqq
i (0, i )
i Rni ni .
(8.19)
i = 1, . . . , m,
ui (0, G)
i (0, i ),
(8.20)
194
where
xi1
zi1
..
..
ni p
, Zi = . Rni q ,
X i := . R
xin
zin
i
i
yi1
, X = . R
,
= . RN1 ,
y = ... R
ym
Xn
m
Z1 0n1 q . . . 0n1 q
u1
0n q Z2
..
2
N (mq)
, u = . R(mq)1 . (8.22)
Z= .
R
..
..
.
um
Zm
0nm q
The covariance matrix of the combined random effects vector u, on the one hand, and
the combined residual vector , on the other hand, are specified as
G
1
mqmq
NN
..
..
D=
, =
.
(8.23)
R
R
.
.
m
Z1 G Z1 + 1 . . .
0
..
=
.
+ m
0
Zm G Zm
V1
..
=
,
.
(8.24)
Vm
with Vi = Zi G Zi + i .
Example 8.8 (A Three-Level Example.). yij k is the response variable of interest, as
observed for, say, vehicle k, insured in fleet j by insurance company i. At vehicle-level
195
(8.25)
Hereby, predictors z1,ij k and x1,ij k may depend on the insurance company, fleet, or
vehicle. 1 is a vector of regression parameters that do not vary by company or fleet;
they are fixed effects regression parameters. Parameters ij vary by company and
fleet. We model them in a level-two equation:
ij = Z2,ij i + X 2,ij 2 + 2,ij .
(8.26)
X 2,ij and Z2,ij may depend on the company or fleet, but not on the insured vehicle.
The regression parameters in i are company-specific and modeled in (8.27):
i = X 3i 3 + 3i ,
(8.27)
where the predictors in X 3i may depend on the company, but not on the fleet or vehicle.
The combined level 1, 2, and 3 models lead to the following model specification:
Yij k = z1,ij
k (Z2,ij (X 3,i 3 + 3i ) + X 2,ij 2 + 2,ij ) + x1,ij k 1 + 1,ij k
(8.28)
z1,ij
z1,ij
where xij k = (x1,ij
k
k X 2,ij
k Z2,ij X 3i ), = (1 2 3 ) , z i,j,k =
z1,i,j,k
Z2,ij ) and uij = (2,ij
3,i
) . Formulating this three-level model in
(z1,i,j,k
matrix notation follows from stacking all observations Yij k .
More examples of LMM specifications are in McCulloch and Searle (2001). A standard notation for a klevel model is in Frees (2004; see Chapter 5, Appendix A).
8.2.2 The Structure of Random Effects
Since the random effects u often correspond to factor predictors, the design matrix Z is
often highly sparse, with a high proportion of elements being exactly zero. Moreover,
the covariance matrix D is highly structured and depends on some parameter vector
that is to be estimated. We consider the following structures for the random effects:
r Single random effect per level: This is the simplest yet most common case where the
random effect corresponds to a certain level of a single grouping factor. For example,
we may have the state indicator in the model, and each state has its own intercept (i.e.,
y ~ (1|state) (in R parlance)). We illustrate this structure in Section 8.3 with the
workers compensation losses data.
r Multiple random effects per level: Another common case is that the model has both
random intercepts and random slopes that vary by some grouping factor. For example,
each state in the model has its own intercept and also its own slope with respect to some
predictor (i.e., y ~ (1 + time|state)). In general, the multiple random effects are
196
correlated, and so the matrix D is not diagonal. We illustrate this structure in Section 8.3
with the workers compensation losses data.
r Nested random effects: In the nested classification, some levels of one factor occur
only within certain levels of a first factor. For example, we may have observations within
each county and then of the counties within each state. The county from state A never
occurs for state B, so counties are nested within states, forming a hierarchical structure
(i.e., y ~ (1|county/state)). Antonio et al. (2010) provide an example of this
type of structuring.
r Crossed random effects: This happens when each level of each factor may occur with
each level of each other factor. For example, we may have both state and car make in the
model, and cars of different makes can occur within each state (i.e., y ~ (1|state)
+ (1|make)). The credit insurance example in Section 8.3 is an example of crossed
random effects.
(8.29)
See Frees (2004) or Czado (2004) for a formal derivation of this result.
Predicting u. In the sense of the minimal mean squared error of prediction (MSEP),
the best predictor (BP) of u is the conditional mean E[u|Y ]. This predictor obviously
requires knowledge of the conditional distribution u|Y . The BP is often simplified by
restricting the predictor to be a linear function of Y : the best linear predictor (BLP).
197
(8.30)
where V = Var( y) and C = Cov(u, y ). BP(u) and BLP(u) are unbiased, in the sense
that their expected value equals E[u]. Normality is not required in BP or BLP, but
with ( y u) being multivariate normally distributed, the BP and BLP coincide. See
McCulloch and Searle (2001) and Chapter 9 for more details.
In the context of the LMM sketched in (8.17) the predictor of u is usually called
the best linear unbiased predictor (BLUP). Robinson (1991) describes several ways
to derive this BLUP. For instance, under normality assumptions,
Cov( y, u ) = Cov(X + Zu + , u )
= Cov(X, u ) + ZVar(u, u ) + Cov(, u )
= Z D,
which leads to the multivariate normal distribution
y
X
V
N
,
u
0
D Z
ZD
D
.
(8.31)
Using either properties of this distribution4 or the result in (8.30), the BLUP of u
follows:
BLUP(u) := u = D Z V 1 ( y X).
(8.32)
Y
Y
Y Y Z
N
,
we know Z|Y N( Z|Y , Z|Y ) where Z|Y = Z +
Z
Z
ZY Z
1
1
ZY Y (Y Y ) and Z|Y = Z ZY Y Y Z .
Namely, with X =
198
(8.34)
Q(, u) = 0 and
Q(, u) = 0
u
1
1
X 1 Z
X y
X X
.
=
1
1
1
Z X Z Z+ D
Z 1 y
u
(8.35)
(8.36)
using Vi = Zi D Zi + i and ni the cluster size. y i is a weighted mean of the global
profile X i and the data observed on cluster i, yi . y i is a socalled shrinkage estimator.
Actuaries will recognize a credibility type formula in (8.36).
The prediction of a future observation is discussed in detail in Frees (2004 and in
Section 4.4). The case of nondiagonal residual covariance matrices i requires special
u +
attention. For instance, with panel data the BLUP for yi,Ti +1 is xi,T
+ zi,T
i +1
i +1 i
BLUP(i,Ti +1 ). From (8.30) we understand that the last term in this expression is zero
when Cov(i,Ti +1 , i ) = 0. This is not the case when serial correlation is taken into
account. Chapter 9 carefully explains this kind of prediction problems.
Estimating Variance Parameters. The parameters or variance components used in
V are in general unknown and should be estimated from the data. With the vector
of unknown parameters used in V = Z D()Z + D(), the log-likelihood for (, )
is (with c a constant)
(, ) = log {L(, )}
1
= ln |V ( )| + ( y X) V ()1 ( y X) + c.
2
(8.37)
()
= (X V ( )1 X)1 X V ()1 y.
(8.38)
199
2
1 22
ln X V ( )1 X 2 + constants.
2
(8.42)
(8.43)
A well-known example of REML versus ML considers the case of a random sample X1 , . . . , XN N(, 2 ). The
resulting estimators for the unknown variance 2 are
2
ML
=
N
N
1
1
2
2 , REML
2,
(Xi X)
=
(Xi X)
N
N 1
i=1
(8.40)
i=1
2
accounts for the
with X the sample mean. The REML estimator is unbiased for 2 . The (N 1) in REML
estimation of by X.
200
However, this estimate ignores the extra variability originating from the estimation of
. Kacker and Harville (1984) among others discuss attempts to quantify this extra
variability through approximation, but only a fully Bayesian analysis accounts for all
sources of variability (see Chapter 16 for a demonstration of a Bayesian analysis of a
generalized linear mixed model).
The covariance of the empirical BLUP in (8.32) is equal to
= Cov( D Z V 1 ( y X ))
Cov(u)
0
1 1 1
XV
= D Z V 1 V 1 X X V 1 X
Z D.
(8.44)
However, the estimator in (8.44) ignores the variability in the random vector u.
Therefore, as suggested by Laird and Ware (1982), inference for u is usually based
on Cov(u u). Estimates of the precision of other predictors involving and u are
based on
Cov
,
(8.45)
u u
and are available in McCulloch and Searle (2001, section 9.4 (c)). Accounting for the
variability induced by estimating the variance components would require once
again a fully Bayesian analysis. Using Bayesian statistics, posterior credible intervals
of cluster-specific effects follow immediately. They are useful in understanding the
between-cluster heterogeneity present in the data.
Inference. We consider testing a set of s (s p) hypotheses concerning the fixed
effects parameters in :
H0 : C =
versus H1 : C = .
(8.46)
][C ]
[C ] [CVar()C
(8.47)
)
the log-likelihood obtained with ML
is approximately s2 distributed. With (,
)
the log-likelihood with ML in the
in the restricted model (i.e., under H0 ) and (,
unrestricted model, the likelihood ratio test (LRT) statistic for nested models
)
)],
(,
2[(,
(8.48)
8.3 Examples
201
Testing the necessity of random effects requires a hypothesis test involving the variance components. For example, in the varying intercepts model from (8.7), we want
to investigate whether the intercepts of different subjects are significantly different.
This corresponds with
H0 : 22 = 0
versus H1 : 22 > 0.
(8.49)
However, because 0 is on the boundary of the allowed parameter space for 22 , the
likelihood ratio test statistic should not be compared with a 12 distribution, but with
a mixture 12 02 + 12 12 . When testing a hypothesis involving s fixed effects parameters
2
and one variance component, the reference distribution is 12 s2 + 12 s+1
. When more
variance components are involved, the complexity of this problem increases; see
Ruppert et al. (2003) and related work from these authors.
8.3 Examples
8.3.1 Workers Compensation Insurance Losses
We analyze the data from Example 8.2 on losses observed for workers compensation
insurance risk classes. The variable of interest is Lossij observed per risk class i and
year j . The distribution of the losses is right skewed, which motivates the use of
log (Lossij ) as the response variable. To enable out-of-sample predictions, we split
the dataset into a training (without Lossi7 ) versus test set (the Lossi7 observations).
We remove observations corresponding with zero payroll from the dataset. Models are
estimated on the training set, and centering of covariate Year is applied. Throughout
our analysis we include log (Payroll)ij as an offset in the regression models, since
losses should be interpreted relative to the size of the risk class.
Complete Pooling. We start with the complete pooling model introduced in (8.1).
The model ignores the clustering of data in risk classes and fits an overall intercept
(0 ) and an overall slope (1 ) for the effect of Year:
log (Lossij ) = log (Payrollij ) + 0 + 1 Yearij + ij
ij N(0, 2 ) i.i.d.
We fit the model with lm in R.
>fitglm.CP <- lm(log(loss)~ yearcentr, offset=log(payroll),
data=wclossFit)
>summary(fitglm.CP)
Coefficients:
Estimate Std. Error
t value Pr(>|t|)
(8.50)
(8.51)
202
(Intercept) -4.34023
0.04105 -105.733
<2e-16 ***
yearcentr
0.03559
0.02410
1.477
0.14
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Residual standard error: 1.062 on 667 degrees of freedom
Multiple R-squared: 0.7282,
Adjusted R-squared: 0.7278
F-statistic: 1787 on 1 and 667 DF, p-value: < 2.2e-16
According to this R output, 0 = 4.34 (with s.e. 0.041), 1 = 0.036 (with s.e. 0.024),
and = 1.062.
No Pooling. The fixed effects linear regression model in (8.2) estimates an intercept
for each of the 118 risk classes in the dataset. According to model equation (8.52),
the intercepts 0,i are unknown, but fixed, whereas the error terms ij are stochastic:
log(Lossij ) = log (Payrollij ) + 0,i + 1 Yearij + ij
ij N(0, 2 ) i.i.d.
(8.52)
We fit this model in R by identifying the risk class variable as a factor variable.
>fitglm.NP <- lm(log(loss)~0+yearcentr+factor(riskclass),
offset=log(payroll),
data=wclossFit)
>summary(fitglm.NP)
Coefficients:
Estimate Std. Error
0.03843
0.01253
-3.49671
0.22393
-3.92231
0.22393
-4.48135
0.22393
-4.70981
0.22393
t value Pr(>|t|)
3.067 0.00227 **
-15.615 < 2e-16 ***
-17.516 < 2e-16 ***
-20.012 < 2e-16 ***
-21.032 < 2e-16 ***
yearcentr
factor(riskclass)1
factor(riskclass)2
factor(riskclass)3
factor(riskclass)4
...
Residual standard error: 0.5485 on 550 degrees of freedom
Multiple R-squared: 0.9986,
Adjusted R-squared: 0.9983
F-statistic: 3297 on 119 and 550 DF, p-value: < 2.2e-16
8.3 Examples
4
6
8
12.5
203
2
4
6
8
15.0
17.5 20.0
log(Size)
22.5
12.5
15.0
17.5 20.0
log(Size)
22.5
Fig. 8.6. Point estimates for risk class specific intercepts, plus/minus one standard
error. Results from no pooling approach (left) and linear mixed model (right). The
dashed line is y = 4.34 (i.e., the overall intercept from the complete pooling model).
> anova(fitglm.CP,fitglm.NP)
Analysis of Variance Table
Model 1: log(loss) ~ yearcentr
Model 2: log(loss) ~ 0 + yearcentr + factor(riskclass)
Res.Df
RSS Df Sum of Sq
F
Pr(>F)
1
667 751.90
2
550 165.48 117
586.42 16.658 < 2.2e-16 ***
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Figure 8.6 (left) shows the estimates 0,i , plus/minus one standard error, against
the size (on log-scale) of the risk class. The size of a risk class is here defined as
6
j =1 Payrollij . The no pooling model estimates risk class specific intercepts with
reasonable precision.
Linear Mixed Models: Random Intercepts. A linear mixed model with random risk
class specific intercepts is a meaningful alternative to the no pooling model in (8.52).
The regression equation is
log (Lossij ) = log (Payrollij ) + 0 + u0,i + 1 Yearij + ij
u0,i N(0, u2 ) i.i.d.
ij N(0, 2 ) i.i.d.
(8.53)
Random intercepts u0,i are independent across risk classes and independent of the
error terms ij . In R we use the lme4 package to fit this linear mixed model. The
204
The R output shows the following parameter estimates: 0 = 4.32 (s.e. 0.089),
1 = 0.037 (s.e. 0.013), u = 0.94, and = 0.55. In Figure 8.6 (right) we plot the
point predictions for the ui,0 s, and their corresponding standard errors, against size
of the risk class. To create this plot we refit the linear mixed model and do not include
an intercept.
The point estimates of the random intercepts obtained with the no pooling model
in (8.52) and the linear mixed model in (8.53) are similar in this example. For the
standard errors of the random intercepts in the LMM, we use the following instructions
str(rr1 <- ranef(lmm0, condVar = TRUE))
my.se.risk = sqrt(as.numeric(attributes(rr1$riskclass)$postVar)),
8.3 Examples
205
(8.44). We are aware of the fact that they do not account for all sources of variability
involved.
Linear Mixed Models: Random Intercepts and Slopes. We now extend the LMM in
(8.53) and allow for random slopes as well as random intercepts. This is an example of
the multiple random effects per level setting from Section 8.2.2. The model equation
is
log (Lossij ) = log (Payrollij ) + 0 + u0,i + 1 Yearij + u1,i Yearij + ij ,
ui N(0, D( )) i.i.d.
ij N(0, 2 ) i.i.d.
(8.54)
The random effects vector ui is now bivariate, say with Var(ui,0 ) = 0 , Var(ui,1 ) = 1 ,
and Cov(ui,0 , ui,1 ) = 01 . Random effects are independent across risk classes and
independent of the error terms ij . We fit this model with lmer as follows.
> lmm2 <- lmer(log(loss) ~ (1+yearcentr|riskclass)+yearcentr
+offset(log(payroll)),
data=wclossFit)
> print(lmm2)
Linear mixed model fit by REML
Formula: log(loss) ~ (1 + yearcentr | riskclass) + yearcentr
+ offset(log(payroll))
Data: wclossFit
AIC BIC logLik deviance REMLdev
1451 1478 -719.4
1429
1439
Random effects:
Groups
Name
Variance Std.Dev. Corr
riskclass (Intercept) 0.885937 0.941242
yearcentr
0.003171 0.056312 -0.195
Residual
0.290719 0.539184
Number of obs: 669, groups: riskclass, 118
Fixed effects:
Estimate Std. Error t value
(Intercept) -4.32030
0.08929 -48.38
yearcentr
0.03715
0.01340
2.77
Correlation of Fixed Effects:
(Intr)
yearcentr -0.072
206
> anova(lmm1,lmm2)
Data: wclossFit
Models:
lmm1: log(loss) ~ (1 | riskclass) + yearcentr + offset
(log(payroll))
lmm2: log(loss) ~ (1 + yearcentr | riskclass) + yearcentr +
offset(log(payroll))
Df
AIC
BIC logLik Chisq Chi Df Pr(>Chisq)
lmm1 4 1438.5 1456.6 -715.27
lmm2 6 1440.9 1468.0 -714.46 1.6313
2
0.4423
When performing the corresponding LRT the software automatically refits lmm1 and
lmm2 with ML (instead of REML), as required (see our discussion in Section 8.2.3.1).
This explains why the AIC, BIC, and logLik values differ from those printed in
the box. The observed Chisq test statistic and reported pvalue indicate that H0 :
12 = 0 cannot be rejected. The model with only random intercepts is our preferred
specification.
Out-of-Sample Predictions. We compare outofsample predictions of Lossi7 ,
for given Payrolli7 , as obtained with models (8.50), (8.52), and (8.53). Figure 8.7 plots observed versus fitted losses (on log scale) for (from left to right)
the complete pooling, the random intercepts, and the no pooling linear regression
model.
8.3 Examples
lmm1
NP
log(Pred)
CP
207
10.0
12.5
15.0
10.0
12.5 15.0
log(Obs)
10.0
12.5
15.0
Fig. 8.7. Out-of-sample predictions for Lossi7 versus observed losses, as obtained
with model (8.50) (CP, complete pooling), (8.53) (lmm1, random intercepts), and
(8.52) (NP, no pooling): losses on workers insurance compensation data.
(8.55)
Apart from the normality assumption, actuaries recognize this as the so-called
Buhlmann credibility model, as Chapter 9 explains.
Random Intercepts, Including Weights. Our response variable is average loss per
claim, constructed as total loss (per state and quarter) divided by the corresponding
number of claims. This average loss is more precise when more claims have been
observed. We therefore include the number of observed claims as weights (wij ) in our
LMM:
Ratioij = 0 + ui,0 + ij
(8.56)
208
The model equation and variance assumptions (apart from normality) correspond
with the BuhlmannStraub credibility model. Including weights goes as follows in
R lme4.
> lmmBS <- lmer(ratio ~ (1|state),weights=weight,data=hach)
> print(lmmBS)
Linear mixed model fit by REML
Formula: ratio ~ (1 | state)
Data: hach
AIC BIC logLik deviance REMLdev
1301 1307 -647.5
1306
1295
Random effects:
Groups
Name
Variance Std.Dev.
state
(Intercept)
22.326
4.725
Residual
47928.954 218.927
Number of obs: 60, groups: state, 5
Fixed effects:
Estimate Std. Error t value
(Intercept) 1688.934
2.265
745.6
The risk (or: credibility) premium for state i is 0 + u i,0 , and is available in R as
follows.
## get fixed effects
fe <- fixef(lmmBS)
## get random intercepts
re <- ranef(lmmBS)
## calculate credibility premiums in this lmm
pred.lmmBS <- fe[1]+re$state
> t(pred.lmmBS)
1
2
3
4
5
(Intercept) 2053.18 1528.509 1790.053 1468.113 1604.815
8.3 Examples
209
Random Intercepts and Slopes, including Weights. We extend the random intercepts model to a random intercepts and slopes model, using the period of observation
as regressor:
Ratioij = 0 + ui,0 + 1 periodij + ui,1 periodij + ij
ui N(0, D( )) i.i.d.
ij N(0, 2 /wij ) i.i.d.
(8.57)
Our analysis uses periodij as the quarter (j = 1, . . . , 12) of observation. The use of a
centered version of period is discussed in Chapter 2. In R the (1+period|state)
instruction specifies random intercepts and slopes per state.
> lmmHach <- lmer(ratio ~ period+(1+period|state),weights=weight,
data=hach)
> lmmHach
Linear mixed model fit by REML
Formula: ratio ~ period + (1 + period | state)
Data: hach
AIC BIC logLik deviance REMLdev
1242 1255 -615.1
1247
1230
Random effects:
Groups
Name
Variance
Std.Dev. Corr
state
(Intercept) 4.1153e+00
2.02863
period
1.9092e-01
0.43695 1.000
Residual
1.6401e+04 128.06735
Number of obs: 60, groups: state, 5
Fixed effects:
Estimate Std. Error t value
(Intercept) 1501.5452
1.1265 1333.0
210
period
0.2172
127.7
Using LMM (8.57) the state specific risk premium for the next time period is
(8.58)
> t(pred.lmmHach)
[,1]
[,2]
[,3]
[,4]
[,5]
[1,] 2464.032 1605.676 2067.279 1453.923 1719.48.
These premiums correspond with the results (obtained with SAS) reported in Frees
et al. (1999, table 3, columns Prediction and standard errors). These authors also
investigate linear mixed models as a user-friendly and computationally attractive
alternative for actuarial credibility models. The traditional Hachemeister credibility
premiums are available in R as follows (see also the Base results in table 3 from
Frees et al. (1999).
fitHach <- cm(~state, hachemeister,regformula = ~time, regdata =
data.frame(time = 1:12),ratios = ratio.1:ratio.12,
weights = weight.1:weight.12)
pred.Hach <- predict(fitHach, newdata = data.frame(time = 13))
# cred.premium
# > pred.Hach
# [1] 2436.752 1650.533 2073.296 1507.070 1759.403
Once again, with linear mixed models we obtain premiums that are close to, but
do not replicate, the traditional actuarial credibility results. Differences in parameter
estimation techniques explain why these results are not identical.
Using an LRT we verify whether model (8.57) should be reduced to the model with
random intercepts only. The pvalue indicates that this is not case.
lmmHach2 <- lmer(ratio ~ period+(1|state),weights=weight,data=hach)
anova(lmmHach,lmmHach2)
#Data: hach
#Models:
#lmmHach2: ratio ~ period + (1 | state)
#lmmHach: ratio ~ period + (1 + period | state)
8.3 Examples
211
#
Df
AIC
BIC logLik Chisq Chi Df Pr(>Chisq)
#lmmHach2 4 1272.2 1280.5 -632.08
#lmmHach
6 1258.7 1271.2 -623.32 17.521
2 0.0001568 ***
#--#Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Figure 8.8 illustrates the fit of a complete pooling (dark gray, dashed line), a no
pooling (black, dashed line), and an LMM with random intercepts and slopes (black,
solid line). The regression equations for the complete and no pooling model are
Ratioij = 0 + 1 periodij + ij
ij N(0, 2 /wij ),
(8.59)
and
Ratioij = 0,i + 1,i periodij + ij
ij N(0, 2 /wij ),
(8.60)
respectively.
1
2500
2000
1500
Ratio
1000
4
2500
2000
1500
1000
2.5 5.0 7.5 10.0 12.5 2.5 5.0 7.5 10.0 12.5
Period
Fig. 8.8. Fit of a complete pooling (dark gray, dashed line), a no pooling (black,
dashed line), and an LMM with random intercepts and slopes (black, solid line):
Hachemeister data (no centering of period).
212
181.05
172.11
225.29
238.18
229.16
282.24
277.77
268.8
323.68
(2)
(12)
(123)
(8.61)
(2)
(12)
Paymentij t = m + ui + uj + uij + ij t
2
u(1)
i N(0, 1 )
2
u(2)
j N(0, 2 )
2
u(12)
ij N(0, 12 )
ij t N(0, 2 ),
(8.62)
where i and j run over all levels in factors 1 (status) and 2 (experience) and we
assume all random variables to be independent.
213
The resulting risk premiums as obtained with lme4 are very close to the credibility
premiums in Table 8.2.
experience
#status
1
2
3
1 181.0253 238.1813 277.7692
2 172.1086 229.1551 268.7954
3 225.2921 282.2424 323.6784
Our analysis directly uses Payment as response variable to facilitate the comparison
between the credibility and linear mixed model calculations. However, the positivity
and right skewness of Payment suggest the use of a lognormal or gamma distribution
for this response.
8.4 Further Reading and Illustrations
We recommend Czado (2004), Gelman and Hill (2007), Frees (2004), McCulloch
and Searle (2001), Ruppert et al. (2003), and Verbeke and Molenberghs (2000) as
214
further readings on linear mixed models. The use of LMMs for smoothing purposes
is not discussed in this chapter, but interested readers can find in Example 8.9 a brief
introduction and useful references.
Example 8.9 (Smoothing with Mixed Models). A semi-parametric regression
model incorporates both parametric and nonparametric functional relationships
between a response and a set of covariates. These models are particularly useful when
a globally linear pattern is inappropriate or parametric nonlinear curves are difficult
to determine. Such nonlinear effects frequently occur when time-related covariates
are present, such as drivers age, development lag, or years in business of the insured
company. For example, in an LM the effect of age of the insured on the number
of claims reported is often expressed with a categorical Age covariate. The analyst
splits Age into several categories and estimates a regression parameter for each one.
In a nonparametric analysis we model the effect of Age on the response with an
unknown, smooth function, in comparison with the piecewise constant assumption in
linear models.
Penalized splines (also called P-splines) are popular nonparametric tools that specify the smoothing function as a linear combination of basis functions, in which some
coefficients associated with the basis functions are constrained to avoid overfitting.
That is, they are penalized, or shrunk toward zero, reducing the effective number of
coefficients to be estimated. The widespread popularity of P-splines is largely because
they can be written in the form of mixed models (Ruppert et al. 2003; Wood 2006) so
that we can rely on software, diagnostic and inferential tools designed for mixed models directly in fitting P-splines, or can use a Bayesian implementation of the model to
make inference of the full posterior distribution. Of course, hierarchical components
can be included in addition to smoothing terms, thus often leading to models that are
both intuitively appealing and structurally flexible when studying practical problems
in predictive modeling.
For example, Figure 8.9 shows an application of the P-splines in estimating insurance loss reserves. In this example, the incremental paid insurance losses, represented
by the dots in the plot, exhibit a nonlinear dependence on the report lag (the xaxis). Standard loss reserving methods specify a model with these lags as categorical
covariates. In contrast, P-splines allow us to estimate a smooth functional relationship between paid losses and report lags. One advantage over the reserving model
with dummy variables is the reduced number of model parameters, because generally a small number of knots can capture the observed pattern sufficiently well. The
example shown here is based on a four-knot penalized spline, and Zhang and Dukic
(2012) find that the resulting model has significantly better predictive performance
than a dummy-variable-based reserving model. Another benefit is that estimates at any
time point can be produced based on interpolation or extrapolation of the estimated
References
215
20
15
10
5
0
0
6
Lag
10
Fig. 8.9. The plot of the company-level smoother (incremental losses) along with the
50% prediction interval for a loss triangle.
functional form. This can be very helpful when the goal of a reserving study is to
make forecasts for a short period ahead, say one month or a quarter.
More examples of semi-parametric models in insurance loss reserving can be found
in Antonio and Beirlant (2008) and Zhang and Dukic (2012). Multivariate extensions
of penalized splines are available for spatial regression (e.g., in postcode rating).
References
Antonio, K. and J. Beirlant (2007). Actuarial statistics with generalized linear mixed models.
Insurance: Mathematics and Economics 40(1), 5876.
Antonio, K. and J. Beirlant (2008). Issues in claims reserving and credibility: A
semiparametric approach with mixed models. Journal of Risk and Insurance 75(3),
643676.
Antonio, K., E. Frees, and E. Valdez (2010). A multilevel analysis of intercompany claim
counts. ASTIN Bulletin 40(1), 151177.
Antonio, K. and E. Valdez (2012). Statistical aspects of a priori and a posteriori risk
classification in insurance. Advances in Statistical Analysis 96(2), 187224.
Buhlmann, H. and A. Gisler (2005). A Course in Credibility Theory and Its Applications.
Springer Verlag, Berlin.
Czado, C. (2004). Linear Mixed Models. Lecture slides on GLM, TU Munchen.
Dannenburg, D., R. Kaas, and M. Goovaerts (1996). Practical Actuarial Credibility Models.
Institute of Actuarial Science and Econometrics, University of Amsterdam.
Frees, E. (2004). Longitudinal and Panel Data. Analysis and Applications in the Social
Sciences. Cambridge University Press, Cambridge.
Frees, E., V. Young, and Y. Luo (1999). A longitudinal data analysis interpretation of
credibility models. Insurance: Mathematics and Economics 24(3), 229247.
Frees, E., V. Young, and Y. Luo (2001). Case studies using panel data models. North
American Actuarial Journal 5(4), 2442.
Gelman, A. (2006). Multilevel (hierarchical) modeling: What it can and cannot do.
Technometrics 48(3), 432435.
216
Gelman, A. and J. Hill (2007). Applied Regression and Multilevel (Hierarchical) Models.
Cambridge University Press, Cambridge.
Hachemeister, C. (1975). In Credibility: Theory and Applications, Credibility for regression
models with application to trend, pp. 129163. Academic Press, New York.
Kacker, R. and D. Harville (1984). Approximations for standard errors of estimators of fixed
and random effects in mixed linear models. Journal of the American Statistical
Association 79, 853862.
Klugman, S. (1992). Bayesian Statistics in Actuarial Science with Emphasis on Credibility.
Kluwer, Boston.
Laird, N. and J. Ware (1982). Random-effects models for longitudinal data. Biometrics 38(4),
963974.
Makov, U., A. Smith, and Y. Liu (1996). Bayesian methods in actuarial science. The
Statistician 45(4), 503515.
McCulloch, C. and S. Searle (2001). Generalized, Linear and Mixed Models. Wiley Series in
Probability and Statistics, Wiley, New York.
Robinson, G. (1991). That BLUP is a good thing: The estimation of random effects.
Statistical Science 6, 1551.
Ruppert, D., M. Wand, and R. Carroll (2003). Semiparametric Regression. Cambridge
University Press, Cambridge.
Scollnik, D. (1996). An introduction to Markov Chain Monte Carlo methods and their
actuarial applications. Proceedings of the Casualty Actuarial Society Forum LXXXIII,
114165.
Searle, S., G. Casella, and C. McCulloch (2008). Variance Components. Wiley, New York.
Verbeke, G. and G. Molenberghs (2000). Linear Mixed Models for Longitudinal Data.
Springer Series in Statistics, New York.
Wood, S. (2006). Generalized Additive Models: An Introduction with R. Chapman & Hall,
CRC Texts in Statistical Science.
Zhang, Y. and V. Dukic (2012). Predicting multivariate insurance loss payments under the
Bayesian copula framework. Journal of Risk and Insurance.
Zhang, Y., V. Dukic, and J. Guszcza (2012). A Bayesian nonlinear model for forecasting
insurance loss payments. Journal of the Royal Statistical Society, Series A 175, 637656.
9
Credibility and Regression Modeling
Vytaras Brazauskas, Harald Dornheim, and Ponmalar Ratnam
Chapter Preview. This chapter introduces the reader to credibility and related regression modeling. The first section provides a brief overview of credibility theory and
regression-type credibility, and it discusses historical developments. The next section
shows how some well-known credibility models can be embedded within the linear mixed model framework. Specific procedures on how such models can be used
for prediction and standard ratemaking are given as well. Further, in Section 9.3, a
step-by-step numerical example, based on the widely studied Hachemeisters data, is
developed to illustrate the methodology. All computations are done using the statistical software package R. The fourth section identifies some practical issues with the
standard methodology, in particular, its lack of robustness against various types of
outliers. It also discusses possible solutions that have been proposed in the statistical
and actuarial literatures. Performance of the most effective proposals is illustrated on
the Hachemeisters dataset and compared to that of the standard methods. Suggestions
for further reading are made in Section 9.5.
9.1 Introduction
9.1.1 Early Developments
Credibility theory is one of the oldest but still most common premium ratemaking
techniques in insurance industry. The earliest works in credibility theory date back
to the beginning of the 20th century, when Mowbray (1914) and Whitney (1918)
laid the foundation for limited fluctuation credibility theory. It is a stability-oriented
form of credibility, the main objective of which is to incorporate into the premium
as much individual experience as possible while keeping the premium sufficiently
stable. Despite numerous attempts, this approach never arrived at a unifying principle
that covered all special cases and that opened new venues for generalization. Its range
of applications is quite limited, and thus it never became a full-fledged theory.
217
218
Instead of focusing solely on the stability of the premium, the modern and more
flexible approach to credibility theory concentrates on finding the most accurate
estimate of an insureds pure risk premium. This is accomplished by striking a balance
between the individuals risk experience and the average claim over all risk classes.
Although initial contributions to this area can be traced back to the 1920s (see Keffer
1929), it is generally agreed that the systematic development of the field of greatest
accuracy credibility started in the late 1960s with the seminal paper of Buhlmann
(1967). A few years later, Buhlmann and Straub (1970) introduced a credibility
model as a means to rate reinsurance treaties, which generalized previous results and
became the cornerstone of greatest accuracy credibility theory. The model is one of
the most frequently applied credibility models in insurance practice, and it enjoys
some desirable optimality properties. For more historical facts and further discussion
about credibility, see the classic textbook by Klugman, Panjer, and Willmot (2012,
chapters 1718).
9.1.2 Regression-Type Credibility
The first credibility model linked to regression was introduced by Hachemeister (1975)
who employed it to model U.S. automobile bodily injury claims classified by state
and with different inflation trends. Specifically, Hachemeister considered 12 periods,
from the third quarter of 1970 to the second quarter of 1973, of claim data for bodily
injury that was covered by a private passenger auto insurance. The response variable
of interest to the actuary is the severity average loss per claim, denoted by yit . It is
followed over the periods t = 1, . . . , ni for each state i = 1, . . . , m. Average losses
were reported for n1 = = nm = 12 periods and from m = 5 different states (see
Appendix, Table 9.3).
A multiple time series plot of the observed variable average loss per claim, yit , and
of the average loss per claim in logarithmic units, ln(yit ), is provided in Figure 9.1
(log-claim modeling is considered in Section 9.4). The plots indicate that states differ
with respect to variability and severity. For instance, State 1 reports the highest average
losses per claim, whereas State 4 seems to have larger variability compared to other
states. For all five states we observe a small increase of severity over time. Since
the response variable yit grows over time t and varies from one state to another,
this provides a hint about the possible structure of explanatory variables. Therefore,
Hachemeister originally suggested the use of the linear trend model a regression
model which can be viewed as a special case of the linear mixed models of Chapter 8:
yit = xit + zit ui + it ,
where xit = zit = (1, t) are known designs for the fixed effects (parameter ) and
the subject-specific random effects (parameter ui ), respectively, and it denotes
within-subject residuals.
9.1 Introduction
2600
2400
Severity
State 1
2000
State 3
1800
1600
1400
1200
1000
State 5
State 2
State 4
0 1 2 3 4 5 6 7 8 9 10 11 12
Time Period
LnSeverity
2200
7.9
7.8
7.7
7.6
7.5
7.4
7.3
7.2
7.1
7
6.9
219
State 1
State 3
State 5
State 2
State 4
0 1 2 3 4 5 6 7 8 9 10 11 12
Time Period
Fig. 9.1. Multiple time series plot of the variable average loss per claim, yit , and the
logarithmic average loss per claim, ln(yit ).
Later in this chapter, more details and examples are provided about the link between
the linear mixed models and the most popular credibility models (see Section 9.2.1).
Also, as is shown in Section 9.2.2, the general linear prediction problem for linear
mixed models is closely related to credibility ratemaking. It turns out that generalized
least squares and best linear unbiased predictors correspond to the well-known pricing
formulas of credibility theory.
220
In the remainder of the chapter, we first present the standard likelihood-based procedures for ratemaking, then provide a step-by-step numerical example, and conclude
with a brief review of robust techniques and a comparison of their performance to
that of the standard methods. All computations are done using the statistical software
package R and are based on Hachemeisters data.
221
222
credibility regression line. To overcome this drawback, Buhlmann and Gisler (1997)
suggested taking the intercept of the regression line at the center of gravity of the
time variable, instead of at the origin of the time axis. That is, choose design matrices
Xi = (xi1 , . . . , xini ) and Zi = (zi1 , . . . , zini ) with xit = zit = (1, t Ci ) , where
1
Ci = i
ni
t it
t=1
i
is the center of gravity of the time range in risk i, and i = nt=1
it . This modification
ensures that the regression line stays between the individual and collective regression
lines; the model is called the revised Hachemeister regression model.
From a practical point of view, volumes are often equal enough across periods for
a single risk to be considered constant in time, which yields similar centers of gravity
between risks. Then, it is reasonable to use the center of gravity of the collective,
m ni
m ni
1
which is defined by C =
i=1
i=1
t=1 t it , where =
t=1 it (see
Buhlmann and Gisler 2005, Section 8.3).
i = 1, . . . , m,
(9.1)
where xi,ni +1 and zi,ni +1 are known covariates of risk i in time period ni + 1, and
GLS
and
uBLUP,i are computed using equations (8.29) and (8.30), respectively (see also the
discussion following equation (8.30)).
BLUP,i is called a homogeneous estimator of Wi
In the actuarial literature, W
(Dannenburg, Kaas, and Goovaerts 1996), and it is used to predict the expected
claim size i,ni +1 = E(yi,ni +1 |ui ) of risk i for time ni + 1. This estimator is even
optimal for non-normally distributed claims (Norberg 1980).
Recall that the central objective of credibility is to price fairly heterogeneous risks
based on the overall portfolio mean, M, and the risks individual experience, Mi . This
relation can be expressed by the general credibility pricing formula:
Pi = i Mi + (1 i ) M = M + i (Mi M),
i = 1, . . . , m,
(9.2)
223
1 m
m
1 ni
2
where y =
i=1 i
i=1 i yi , yi = i
t=1 it yit , and i = (1 + /
(i u2 ))1 . To compute formulas (9.3), one needs to estimate the structural parameters
u2 and 2 . The estimators
u2 and
2 are obtained from (RE)ML (i.e., as a byproduct
from Hendersons mixed model equations) and coincide, when assuming normality,
with the following nonparametric estimators:
m ni
it (yit yi )2
2
and
= i=1mt=1
i=1 (ni 1)
$ m
%
2
i (yi y)
2 (m 1) ,
u2 = 2
2
m
i=1 i
i=1
1
where y =
m
t=1
where
i=1
ni
t=1
it yit
t=1
it yit (t Ci )
bi = A1
ni
i
224
i = diag
1
2
2
1 + / u2 ai2
is the credibility factor for risk i, Ai = diag(ai1 , ai2 ) with ai1 = i , ai2 = 3
i =
ni
1 ni
2
(t
C
)
,
and
C
=
is
the
center
of
gravity.
We
still
have
it
i
i
it
i
t=1
t=1
to estimate the process variance 2 and variances of hypothetical means u21 and u22 .
2
2 = m1 m
,i
,
It is reasonable to estimate 2 by the natural variance estimator
i=1
ni
2
1
2
(y
)
is
a
(conditionally)
unbiased
estimator
where
,i = (ni 2)
it
t=1 it it
2
of the within-risk variance ,i
, and
it is the fitted value of the ith regression line in
time t. The structural parameters u21 and u22 are estimated by
m
2
m
m
and
(bi,1 b1 )2
u21 = c1
m 1 i=1
m
2
3
m
m
u22 = c2
,
(bi,2 b2 )2
m 1 i=1 3
where
41
m
m
i
m 1 i
1
1
, b1 =
i bi,1 ,
c1 =
m
i=1
i=1
4
1
m
m
i
3
i
m1 3
1
1
, and b2 = 3
3
i bi,2 .
c2 =
m
3
i=1
i=1
9.3 Numerical Examples
In this section, we revisit Section 9.1.2 and model the Hachemeisters dataset, which,
over the years, has been extensively analyzed by a number of authors in the actuarial
literature. For example, Dannenburg et al. (1996), Buhlmann and Gisler (1997),
Frees et al. (1999), Pitselis (2008), and Dornheim and Brazauskas (2011a) used this
dataset to illustrate the effectiveness of various regression-type credibility ratemaking
techniques.
To get a feel for how things work, let us study a detailed example that shows
how to fit and make predictions based on the Hachemeister model and the revised
Hachemeister model. All computations are done using the statistical software package
R. Parts of the computer code are available in actuar, an R package for actuarial
science that is described in Dutang et al. (2008).
225
To fit the linear trend regression model of Section 9.2.1.3 to the Hachemeisters
dataset, we employ the following R-code:
>
+
+
>
>
The label hachemeister in the first line of the code reads the dataset that is
available in the actuar package. The last line produces predictions that are based
on formula (9.1). The R-code yields the following credibility-adjusted parameter
estimates and predictions for the five states:
Parameter Estimates
State
i
0 +
ui,0
1 +
ui,1
Prediction
i,12+1
1
2
3
4
5
1693.52
1373.03
1545.36
1314.55
1417.41
57.17
21.35
40.61
14.81
26.31
2436.75
1650.53
2073.30
1507.07
1759.40
In addition, the grand parameters are found by taking the average across all states; they
1 = 32.05. Also, within-state variance is
0 = 1468.77 and
2 = 49,870,187,
are
and the other estimates of variance components are
u21 = 24,154.18 and
u22 =
301.81. Note that the numerical values in the display table differ from the ones
reported by Dutang et al. (2012), who used the same R-code but applied the reversed
time variable; that is, time = 12:1 instead of time = 1:12.
Fitting and predictions based on the revised Hachemeister model (see Sections 9.2.1.4 and 9.2.2.2) are much more involved. The complete program for running
these tasks is presented in the code on the book website. Comments explaining the
formulas or a block of the program are listed between the signs # and #. To run the
program, download the R-code to your computer and use the following command
lines:
> source("HachemRevised.R")
> HachemRevised(5,12, hachemeister[,2:13], hachemeister[,14:25],
1:12,13)
226
Let us now go through the main blocks of the program and review the numerical output
of each block. The first group of commands (which is labeled Program Initialization)
defines the variables used in computations and sets their initial values, if necessary. The
second group of commands, labeled Center of Gravity & Recentered X, computes
the collective center of gravity (the outcome is C = 6.4749) and recenters the design
matrix X, which results in
1 5.4749
1 4.4749
1 3.4749
1 2.4749
1 1.4749
1 0.4749
0.5251
1
1
1.5251
1
2.5251
1
3.5251
1
4.5251
1
5.5251
The third group of commands, labeled Volume Measures & Other Constants,
computes the constants required for further calculations. The results of this step
are = 2,104,688; c1 = 1.3202; c2 = 1.3120; b1 = 1865.4040; and b2 = 44.1535.
The fourth and fifth groups of commands, labeled Variance Components and
Parameters & Prediction, respectively, yield estimates of the structural parameters
and deliver next-period predictions. The summarized results of these two program
blocks are
State
i
1
2
3
4
5
ui,1
0 +
ui,0
1 +
ui,1
2
,i
Prediction
i,12+1
Std. Error
i,12+1
385.71
157.67
127.71
283.51
72.24
27.06
12.60
6.58
2.40
18.63
2058.85
1515.48
1800.85
1389.63
1600.90
60.70
21.04
40.22
31.25
15.01
121,484,314
30,175,637
52,560,076
24,362,730
21,075,078
2847.94
1789.01
2323.70
1795.83
1796.00
110.13
123.15
195.49
242.22
76.39
0 +
Note that the grand parameters are found by taking the average across all states for
2
2
ui,1 , and
,i ; they are 0 = 1673.14, 1 = 33.64, and
= 49,931,567. In
ui,0 , 1 +
addition, the estimates of variance components are
u21 = 93,021.43 and
u22 = 665.48.
The last line of the code screen-prints the results of the program.
227
Note 2 (Potential Outliers): A careful examination of Figure 9.1 reveals that the
quarterly observations #6 in State 1, #10 in State 2, #7 in State 4, and perhaps #6 in
State 5 differ somewhat from their state-specific inflation trends. This suggests the
topic for the next section. That is, we would like to know how to identify outliers and,
if they are present, what to do about them.
9.4 Theory versus Practice
The modeling approach of Section 9.3 is well understood and widely used, but it is
not very realistic in practice. In particular, insurance loss data are often highly skewed
and, heavy-tailed, and they contain outliers. Although complete treatment of these
issues is beyond the scope of this chapter, in the following we present some insights
on how to modify and improve the standard methodology. First, in Section 9.4.1 we
formulate heavy-tailed linear mixed models. Note that heavy-tailed models are also
known as long-tailed or fat-tailed (see Chapter 10). Then, Section 9.4.2 introduces
a three-step procedure for robust-efficient fitting of such models. Robust credibility
ratemaking based on heavy-tailed linear mixed models (which are calibrated using
the robust-efficient fitting procedure of Section 9.4.2) is described in Section 9.4.3.
Finally, in Section 9.4.4, we revisit the earlier examples and illustrate performance of
the robust methods using Hachemeisters data.
9.4.1 Heavy-Tailed Linear Mixed Models
Suppose we are given a random sample (xi1 , zi1 , yi1 , i1 ), . . . , (xini , zini , yini , ini ),
where xit and zit are known p- and q-dimensional row-vectors of explanatory variables
and it > 0 is some known volume measure. Assume the claims yit follow a loglocation-scale distribution with cdf of the form:
$
%
log(yit ) it
, yit > 0, i = 1, . . . , m, t = 1, . . . , ni ,
G(yit ) = F0
1/2
it
defined for < it < , > 0, and where F0 is the standard (i.e., it = 0, =
1, it = 1) cdf of the underlying location-scale family F (it , 2 /it ). Following
regression analysis with location-scale models, we include the covariates xit and
zit only through the location parameter it . Then, the following linear mixed model
may be formulated:
log( yi ) = Xi + Zi ui + i = i + i ,
i = 1, . . . , m,
where log( yi ) = (log(yi1 ), . . . , log(yini )) and i is the ni -dimensional vector of the
within-subject locations it that consist of the population location and the subjectspecific location deviation ui . While normality of ui is still assumed, the error term
228
i = 1, . . . , m.
In the first step, the corrected reweighting mechanism automatically detects and
removes outlying events within risks whose standardized residuals computed from
initial high breakdown-point estimators exceed some adaptive cut-off value. This
threshold value is obtained by comparison of an empirical distribution with a theoretical one. Let us denote the resulting pre-cleaned random sample as
), i1
xi1 , zi1 , log(yi1
, . . . , xin , zin , log(yin ), in , i = 1, . . . , m.
i
Note that for each risk i, the new sample size is ni (ni ni ).
(2) Detection of Between-Risk Outliers
In the second step, the procedure searches the pre-cleaned sample (marked with )
and discards entire risks whose risk-specific profile expressed by the random effect
significantly deviates from the overall portfolio profile. These risks are identified
when their robustified Mahalanobis distance exceeds some adaptive cut-off point.
The process results in
x
i = 1, . . . , i ,
i1 , zi1 , log(yi1 ), i1 , . . . , xin , zin , log(yin ), in ,
i
a cleaned sample of risks. Note that the number of remaining risks is i (i m).
229
Using the described procedure, we find the shifted robust best linear unbiased
predictor for location:
EF0 (i ),
urBLUP, i +
i = X
i CATL + Zi
i = 1, . . . , m,
where
CATL and
urBLUP, i are standard likelihood-based estimators but computed on the
clean sample from Step 2. Also,
EF0 (i ) is the expectation vector of the ni -variate cdf
i ). For symmetric error distributions we obtain the special case
EF0 ( i ) = 0.
Fni (0,
9.4.3 Robust Credibility Ratemaking
The reweighted estimates for location,
i , and structural parameters,
CATL =
2
2
2
uq ,
), are used to calculate robust credibility premiums for the ordinary
(
u1 , . . . ,
but heavy-tailed claims part of the original data. The robust ordinary net premiums
=
ordinary
(
urBLUP,i ), t = 1, . . . , ni + 1, i = 1, . . . , m
ordinary
it
it
are found by computing the empirical limited expected value (LEV) of the fitted loglocation distribution of claims. The percentile levels of the lower bound ql and the
upper bound qg used in LEV computations are usually chosen to be extreme (e.g.,
0.1% for ql and 99.9% for qg ).
Then, robust regression is used to price separately identified excess claims. The
risk-specific excess claim amount of insured i at time t is defined by
ordinary
,
for yit < ql .
it
ordinary
Oit = (yit ql )
it , for ql yit < qg .
ordinary
, for yit qg .
(qg ql )
it
Further, let mt denote the number of insureds in the portfolio at time t and let N =
max1im ni , the maximum horizon among all risks. For each period t = 1, . . . , N,
mt
t = m1
we find the mean cross-sectional overshot of excess claims O
t
i=1 Oit , and
fit robustly the random effects model
t = ot + t ,
O
t = 1, . . . , N,
230
i = 1, . . . , m.
bias
(
)
i,12+1
i
i
1
= i
12
t=1
$
1
it it (yit
it )2 i
12
%2 1/2
it it (yit
it )
t=1
where
it denotes the credibility estimate obtained from the pursued regression
method, i is the total number of claims in state i, and it is the hard-rejection weight
for the observed average loss per claim yit when using the CATL procedure. For
nonrobust REML where no data points are truncated we put it = 1 as a special case.
231
Table 9.1. Individual State Predictions for Hachemeisters Bodily Injury Data Based
on Various Model-Fitting Procedures (if available, estimated standard errors are
provided in parentheses)
Prediction for State
Fitting Procedure
base
m-rc (c = 1.5)
gm-rc (k = 1)
mm-rc
reml
catl
reml
catl
2436
2437
2427
2427
2465 (109)
2471 (111)
2451 (109)
2450 (113)
1650
1650
1648
1648
1625 (122)
1545 (74)
1661 (123)
1552 (74)
2073
2073
2092
2092
2077 (193)
2065 (194)
2065 (193)
2049 (195)
1507
1507
1505
1505
1519 (248)
1447 (174)
1613 (242)
1477 (172)
5
1759
1759
1737
1737
1695 (77)
1691 (57)
1706 (78)
1693 (57)
Several conclusions emerge from Table 9.1. First, we note that the REML and
REML estimates (predictions) are based on Hendersons mixed model equations,
and thus they slightly differ from those of Section 9.3. Second, for States 1, 3, and
5, all techniques, standard and robust, yield similar predictions. For the second and
fourth states, however, CATL produces slightly lower predictions. For instance, for
State 4 it results in 1447 whereas the base model finds 1507 predictions. This can be
traced back to the truncation of the suspicious observations #6 and #10 in State 2 and
#7 in State 4. Third, CATL also identifies claim #4 in State 5 as an outlier and, as a
result, assigns a small discount of 1.47 to each risk. For the revised model (i.e., for
REML and CATL ), prediction patterns are similar.
To illustrate robustness of regression credibility methods that are based on MRC, MM-RC, and GM-RC estimators for quantifying an individuals risk experience,
Pitselis (2008) replaces the last observation of the fifth state, 1,690, by 5,000. We
follow the same contamination strategy and summarize our findings in Table 9.2.
As one can see, the choice of the model-fitting methodology has a major impact on
predictions. Indeed, in the presence of a single outlier, we find that robust procedures
provide stability and reasonable adjustment to predictions, whereas standard methods
overreact. For example, in the contaminated State 5 the REML and base credibility
estimates get inflated by the outlying observation and increase from 1,695 to 2542
and from 1759 to 2596, respectively. Further, because outliers usually distort the
estimation process of variance components and thus yield too low credibility weights,
predictions across all states increase significantly. Note also a dramatic explosion of
standard errors (e.g., the standard error of REML in State 5 jumps from 77 to 829),
which is due to increased within-risk variability that was caused by the contaminating
observation. Furthermore, not all robust credibility predictions react equally to data
232
2501
2755
2645
2649
2517 (119)
2477 (111)
2455 (110)
2459 (112)
1826
1979
1868
1870
1852 (150)
1550 (74)
1949 (166)
1559 (74)
2181
2396
2311
2315
2206 (204)
2071 (194)
2229 (204)
2057 (195)
1994
1841
1723
1724
1987 (255)
1452 (174)
2141 (275)
1484 (172)
5
2596
2121
1964
1943
2542 (829)
1689 (60)
2629 (818)
1694 (60)
9.6 Appendix
233
2800
5500
2600
5000
4500
4000
State 1
2200
2000
Severity
Severity
2400
State 3
3500
3000
2500
1800
2000
1600
1400
State 5
1500
1000
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Time Period
Time Period
Fig. 9.2. Selected point and interval predictions for contaminated Hachemeisters
data. The thin lines connecting o denote one-, two-, and three-step predictions
using REML. The corresponding CATL predictions are marked by .
correlated claims data, we recommend reading Frees et al. (1999, 2001) and Frees
(2004). Further, to gain a deeper understanding of and appreciation for robust credibility techniques, the reader should review Garrido and Pitselis (2000), Pitselis (2004,
2008, 2012), and Dornheim and Brazauskas (2007, 2011b,a). If the reader is not familiar with the philosophy and methods of robust statistics, then the book by Maronna,
Martin, and Yohai (2006) will provide a gentle introduction into the subject. Finally,
an introduction to and developments of hierarchical credibility modeling can be found
in Sundt (1979, 1980), Norberg (1986), Buhlmann and Jewell (1987), and Belhadj,
Goulet, and Ouellet (2009).
9.6 Appendix
Table 9.3. Hachemeisters Bodily Injury Dataset Comprising Average Loss Per
Claim, yit , and the Corresponding Number of Claims Per Period, it
Average Loss Per Claim in State
Period
1
2
3
4
5
6
7
8
9
10
11
12
1738
1642
1794
2051
2079
2234
2032
2035
2115
2262
2267
2517
1364
1408
1597
1444
1342
1675
1470
1448
1464
1831
1612
1471
1759
1685
1479
1763
1674
2103
1502
1622
1828
2155
2233
2059
1223
1146
1010
1257
1426
1532
1953
1123
1343
1243
1762
1306
1456
1499
1609
1741
1482
1572
1606
1735
1607
1573
1613
1690
7861
9251
8706
8575
7917
8263
9456
8003
7365
7832
7849
9077
1622
1742
1523
1515
1622
1602
1964
1515
1527
1748
1654
1861
1147
1357
1329
1204
998
1077
1277
1218
896
1003
1108
1121
407
396
348
341
315
328
352
331
287
384
321
342
2902
3172
3046
3068
2693
2910
3275
2697
2663
3017
3242
3425
234
References
Belhadj, H., V. Goulet, and T. Ouellet (2009). On parameter estimation in hierarchical
credibility. ASTIN Bulletin 39(2), 495514.
Buhlmann, H. (1967). Experience rating and credibility. ASTIN Bulletin 4, 199207.
Buhlmann, H. and A. Gisler (1997). Credibility in the regression case revisited. ASTIN
Bulletin 27, 8398.
Buhlmann, H. and A. Gisler (2005). A Course in Credibility Theory and Its Applications.
Springer, New York.
Buhlmann, H. and W. Jewell (1987). Hierarchical credibility revisited. Bulletin of the Swiss
Association of Actuaries 87, 3554.
Buhlmann, H. and E. Straub (1970). Glaubwurdigkeit fur Schadensatze. Mitteilungen der
Vereinigung Schweizerischer Versicherungsmathematiker 70, 111133.
Dannenburg, D. R., R. Kaas, and M. J. Goovaerts (1996). Practical Actuarial Credibility
Models. Institute of Actuarial Science and Economics, University of Amsterdam.
Dornheim, H. (2009). Robust-Efficient Fitting of Mixed Linear Models: Theory, Simulations,
Actuarial Extensions, and Examples. Ph.D. thesis, University of Wisconsin-Milwaukee.
Dornheim, H. and V. Brazauskas (2007). Robust-efficient methods for credibility when
claims are approximately gamma-distributed. North American Actuarial Journal 11(3),
138158.
Dornheim, H. and V. Brazauskas (2011a). Robust-efficient credibility models with
heavy-tailed claims: A mixed linear models perspective. Insurance: Mathematics and
Economics 48(1), 7284.
Dornheim, H. and V. Brazauskas (2011b). Robust-efficient fitting of mixed linear models:
Methodology and theory. Journal of Statistical Planning and Inference 141(4),
14221435.
Dutang, C., V. Goulet, X. Milhaud, and M. Pigeon (2012). Credibility theory features of
actuar. http://cran.r-project.org/web/packages/actuar/index.html.
Dutang, C., V. Goulet, and M. Pigeon (2008). Actuar: An R package for actuarial science.
Journal of Statistical Software 25(7), 137.
Frees, E. W. (2004). Longitudinal and Panel Data: Analysis and Applications in the Social
Sciences. Cambridge University Press, Cambridge.
Frees, E. W., V. R. Young, and Y. Luo (1999). A longitudinal data analysis interpretation of
credibility models. Insurance: Mathematics and Economics 24, 229247.
Frees, E. W., V. R. Young, and Y. Luo (2001). Case studies using panel data models. North
American Actuarial Journal 5(4), 2442. Supplemental material is available at:
http://research3.bus.wisc.edu/course/view.php?id=129.
Garrido, J. and G. Pitselis (2000). On robust estimation in Buhlmann-Straubs credibility
model. Journal of Statistical Research 34(2), 113132.
Goovaerts, A. S. and W. Hoogstad (1987). Credibility Theory, Surveys of Actuarial Studies.
National-Nederlanden N.V., Rotterdam.
Hachemeister, C. A. (1975). Credibility for regression models with applications to trend. In
P. M. Kahn (Ed.), Credibility: Theory and Applications. Academic Press, New York.
Kaas, R., D. Dannenburg, and M. Goovaerts (1997). Exact credibility for weighted
observations. ASTIN Bulletin 27, 287295.
Keffer, R. (1929). An experience rating formula. Transactions of the Actuarial Society of
America 30, 130139.
Klugman, S., H. Panjer, and G. Willmot (2012). Loss Models: From Data to Decisions
(3rd ed.). Wiley, New York.
Maronna, R. A., D. R. Martin, and V. J. Yohai (2006). Robust Statistics: Theory and
Methods. Wiley, New York.
References
235
10
Fat-Tailed Regression Models
Peng Shi
Chapter Preview. In the actuarial context, fat-tailed phenomena are often observed
where the probability of extreme events is higher than that implied by the normal
distribution. The traditional regression, emphasizing the center of the distribution,
might not be appropriate when dealing with data with fat-tailed properties. Overlooking the extreme values in the tail could lead to biased inference for rate-making and
valuation. In response, this chapter discusses four fat-tailed regression techniques that
fully use the information from the entire distribution: transformation, models based
on the exponential family, models based on generalized distributions, and median
regression.
10.1 Introduction
Insurance ratemaking is a classic actuarial problem in property-casualty insurance
where actuaries determine the rates or premiums for insurance products. The primary
goal in the ratemaking process is to precisely predict the expected claims cost which
serves as the basis for pure premiums calculation. Regression techniques are useful
in this process because future events are usually forecasted from past occurrence
based on the statistical relationships between outcomes and explanatory variables.
This is particularly true for personal lines of business where insurers usually possess
large amount of information on policyholders that could be valuable predictors in the
determination of mean cost.
The traditional mean regression, though focusing on the center of the distribution,
relies on the normality of the response variable. It is well known that insurance claims
data are rarely normal. In actuarial practice, fat-tailed phenomena are often observed
where the probability of extreme events is higher than that implied by the normal
distribution. Data that sometimes exhibit fat tails are also described as heavy-tailed
or long-tailed distributed. This observation motivates the discussion of regression
techniques for fat-tailed data in this chapter.
236
1.0
Pareto(1,1)
Weibull(1.5,1)
Gamma(1.5,1)
0.0
0.0
0.2
0.2
f(y)
0.4 0.6
F(y)
0.4 0.6
0.8
Pareto(1,1)
Weibull(1.5,1)
Gamma(1.5,1)
237
0.8
1.0
10.1 Introduction
3
y
In fact, fat-tailed distributions are not new to the actuarial literature and have been
employed for a variety of actuarial applications. For example, in severity modeling,
actuaries examine the distribution of individual claims using historically observed
data; in loss reserving, valuation actuaries are interested in the distribution of aggregated losses from a single line or multiple lines of business; in risk management,
actuaries use risk metrics, such as value-at-risk, of the loss distribution of an insurance
portfolio to determine the risk capital required by regulatory bodies. Historically, actuaries have been exposed to fat-tailed data modeling extensively in nonlife insurance
where heavy-tailed distributions are often used for the analysis without inclusion of
covariates (see Klugman, Panjer, and Willmot 2008). In contrast, this chapter extends
such concerns to situations in which covariates are available.
When compared with the normal distribution, tail heaviness of a certain distribution is usually measured by moment-based statistics, such as skewness and kurtosis
coefficients. However, moments only provide partial distributional information and
are not guaranteed to be finite for many random variables. We focus on another metric
of fat tails where the size is measured by distribution functions (see the Appendix
for the technical definition). Essentially, we examine how fast the survival function
decays as the variable increases. In the following, we define distributions with a
lower decay than the normal as fat-tailed distributions. The interpretation is rather
intuitive: the fat-tailed random variable has a much higher probability of exceeding
the threshold than the normal random variable. For example, one can show that both
gamma and Pareto have heavier tails than the normal distribution. To visualize the tail
heaviness, we exhibit in Figure 10.1 the density and survival function of three commonly used fat-tailed distributions in actuarial science: gamma, Weibull, and Pareto.
The first panel shows that fat tails come along with skewness, and the second panel
demonstrates that the survival function approaches zero at different rates.
This figure emphasize the right tail because extreme events, in insurance applications, often represent unusual large claims. However, a similar analysis could be
238
performed for the left tail of distributions as well. An example could be the downside
risk in financial markets.
Note that the normal distribution is not required to be the benchmark when studying tail behaviors. In fact, actuaries might have observed heavy-tailed distributions in
insurance risk models (see Panjer and Willmot 1992), where the criterion of subexponential is used to define heavy tail. For regression analysis, we use the normal
distribution as the benchmark because it is the key assumption embedded in the traditional mean regression model. Through minimizing a symmetric loss function, the
traditional regression focuses on the center of the distribution and downplays extreme
values in the tail. Thus a direct application of usual regression to fat-tailed outcomes
could lead to biased statistical inference, especially for small sample studies. For
example, the hypothesis test of statistical significance in finite sample applications
would no longer be valid because of the heavy-tailed distribution. Even in the case
of large samples, the tail heaviness could affect the normality of parameter estimates
and cause asymptotic efficiency loss.
In this chapter, we introduce four regression techniques for fat-tailed data. We start
with transformation techniques. Using transformation, one symmetrizes the distribution of heavy-tailed outcomes so as to apply the usual least squares regression. The
second approach is to use generalized linear models (GLMs). Instead of focusing
on the link function and the variance function, we view the GLMs as a parametric
model where one has the option to select the appropriate distribution for fat-tailed
data from the exponential family. Extending the idea of parametric regressions, we
further discuss fat-tailed regressions using generalized distributions. In particular, we
introduce covariates into a generalized beta distribution that nests many commonly
used fat-tailed distribution as special cases. Finally, we present a nonparametric technique quantile regression. Unlike mean regression, quantile regression looks into the
quantiles of the response rather than the mean. We highlight the median regression, a
special quantile that is relevant to insurance ratemaking.
10.2 Transformation
When modeling claim size, actuaries often use the natural log of claims rather than
the original claims directly in the ordinary regression analysis. This is because insurance claims data often present long tails and the logarithmic transformation helps
symmetrize the claims distribution. The approach has its strengths and limitations. It
is straightforward to apply, and the regression coefficients can be easily interpreted in
terms of proportional change. However, the transformation also means switching from
an additive to a multiplicative mean structure, as well as changing the variance structure. These changes might be appropriate for some datasets, but for others, actuaries
might only want to change the distribution, not the mean and variance structure.
10.2 Transformation
239
(10.1)
with a simple structure of mean E((y, )|x) = x and a normal error term
N(0, 2 ). Though transformation techniques could be applied to both response and
explanatory variables, our discussion refers to functions of dependent variables.
As foreshadowed earlier, the transformation method enjoys the advantage of simple
implementation. However, the transformation also changes the mean structure, and
thus the interpretation of regression coefficients is not straightforward. In the classical
linear regression model, coefficients are interpreted in terms of partial derivatives.
In the transformation model (10.1), one could be tempted to write E(y|x)/xj =
1 (x , )/xj , such as is the common practice in logarithmic models. However,
the difficulty of this practice is obvious because E(y) is not equivalent to (Ey) in
general circumstances.
(y 1)/, > 0
or (y, ) =
ln y,
=0
r Signed power: (y, ) = {sign(y)|y| 1}/, > 0
r Modulus: (y, ) =
sign(y){(|y| + 1) 1}/,
= 0
1 = 0
240
Modified Modulus
10
0
y
10
10
5
5
0
y
10
(y, )
10
=0.0
=0.5
=1.0
=1.5
=2.0
=0.0
=0.5
=1.0
=1.5
=2.0
10
10
=0.0
=0.5
=1.0
=1.5
=2.0
10
(y, )
0
(y, )
0
10
10
BoxCox
10
0
5
c(y 20, y)
10
Fig. 10.2. Illustration of transformations. The line type in the legend from top to
bottom corresponds to = 0, 0.5, 1, 1.5, 2.
{(y + 1) 1}/,
y
ln(y + 1),
y
r Modified modulus: (y, ) =
ln(y + 1),
y
0, = 0
0, = 0
< 0, = 2
< 0, = 2
Each transformation family has its own strength and weakness and might be suitable for different types of data. Box-Cox transformation is a popular power family
that has been extensively used in practice (Box and Cox 1964). The first form is only
valid for positive observations. The second form, handling bounded negative observations through a shift parameter, may sacrifice the asymptotic properties of maximum
likelihood method because of the unknown parameter. Many other transformations
have been considered to circumvent these difficulties: two examples are the signed
power transformation (Bickel and Doksum 1981) and the modulus transformation
(John and Draper 1980). Both families are appropriate for either positive or negative
outcomes. However, the log-likelihood is not well defined for the signed power transformation when there exist some zeros. In addition, both transformations could lead
to bimodal distributions in the case where the support consists of the whole real line,
because the Jacobian is not a monotone function when the response changes sign.
This issue is addressed by the modified modulus family proposed in Yeo and Johnson (2000), where the modulus transformation is forced to be smooth by imposing
different parameters on the positive and negative line. In addition to power functions,
variables could be transformed based on other families. The inverse hyperbolic sine
due to Burbidge, Magee, and Robb (1988) is an example. This family can be applied
to responses assuming both positive and negative values and is shown to be useful in
shrinking the extreme observations.
Figure 10.2 illustrates the difference among selected transformations. We observe
that the modulus transformation on the positive line could be obtained by shifting the
Box-Cox transformation by constant one. Replacing parameter with 2 when the
10.3 GLM
241
response changes sign, the modified modulus is continuous to the second derivative.
Similar to the modulus family, the inverse hyperbolic sine transform suffers from the
discontinuity at zero.
10.2.2 Parameter Estimation
For a given transform family, the best transformation could be derived using
likelihood-based inference. Considering model (10.1), the log-likelihood function
associated with independent original observations (y1 , . . . , yn ) could be shown as
n
n
n
1
2
((yi , ) xi ) +
J (yi , )
l(, , ) ln( ) 2
2
2 i=1
i=1
2
with J (yi , ) = |d(yi , )/dyi |. First we maximize the log-likelihood function holding fixed. Denote y = (y1 , . . . , yn ) and X = (x1 , . . . , xn ) , one could show
) = (X X)1 X(y, ) and
(
2 ( ) = (y, ) (I X(X X)1 X)(y, )/n = S(y, )/n,
) is the standard least square estimates for dependent variable (y, ) and
where (
S(y, ) is the residual sum of squares in the corresponding variance analysis. Then
242
intercept
year2
year3
year4
year5
ppsm
pci
s(pci)
pci2
logLik
AIC
Model 2
Model 3
Estimate
p-value
Estimate
p-value
Estimate
p-value
4.414
0.023
0.097
0.070
0.092
0.174
0.033
<0.001
0.667
0.066
0.184
0.082
<0.001
<0.001
3.781
0.022
0.098
0.071
0.095
0.169
<0.001
0.670
0.057
0.165
0.065
<0.001
5.823
0.019
0.102
0.075
0.100
0.174
0.167
<0.001
0.711
0.051
0.149
0.057
<0.001
0.001
2.793
<0.001
0.003
0.010
677.354
1370.707
672.347
1364.280
673.616
1365.232
Using a GLM, actuaries have the opportunity to select an entire distribution for
the claims, not just the mean and variance. For example, one could use a gamma
distribution for medium-tail claims and an inverse-Gaussian distribution for long-tail
claims. This could be particularly important for prediction purposes.
Another well-known class of statistical models based on the exponential family of
distributions is the generalized additive model (GAM). The GAM replaces the linear
predictors in GLM with additive predictors. A detailed discussion on GAMs is found
in Chapter 15.
Though both are based on the exponential family, the GLM and GAM can be
used for different analytic purposes. Using nonparametric and smoothing techniques,
a GAM provides good fits to the training data and is more suitable for detecting
the relationship between the response and covariates. In contrast, a GLM focuses
on estimation and inference with a better interpretability of results. This point is
demonstrated through the following data example.
We consider a dataset on insurance claims used in Frees and Wang (2005). It contains automobile bodily injury liability claims from a sample of 29 Massachusetts
towns over five years, 19931997. The response is the average claims per unit of
exposure (the pure premium) adjusted for the effects of inflation. Two explanatory
variables are per capita income (pci) and logarithmic population per square mile
(ppsm). The original analysis shows that the gamma regression is a reasonable candidate for the insurance claims. Thus we first fit a generalized linear model using the
gamma distribution with a logarithmic link function. Time effects are accommodated
with binary variables. Estimates are displayed as Model 1 in Table 10.1. Consistently,
we observe the significant positive relation with ppsm and negative relation with pci.
20
25
pci
30
0.4
0.6
243
0.2
Deviance Residual
15
s(pci,2.79)
80
Fig. 10.3. The left panel plots partial residuals from the GAM against covariate. The
right panel plots deviance residuals from the GLM against fitted value.
If one wants to explore the effect of pci further, one could fit a generalized additive
model with an additive component s(pci), shown as Model 2 in Table 10.1. To facilitate
a reasonable comparison, the GAM is also fitted with the gamma distribution and a log
link. The small p-value indicates statistical significance, and the degrees of freedom
value of 2.793 suggests a quadratic pattern. This relation could also be visualized
through an analysis of partial residuals that are the Pearson residuals added to the
partial prediction of the smooth term. We show in the left panel of Figure 10.3 the
partial residual against pci, where the rug plots at the bottom indicate the value of
the covariate. After uncovering the quadratic form of the dependence on pci using
a GAM, we can use a GLM to fit and assess the corresponding parametric model.
In Model 3, a quadratic term of pci is included in addition to the linear one. As
anticipated, the nonlinear effect is well captured in this formulation. The plot of
deviance residuals against fitted values is exhibited in the right panel of Figure 10.3,
and no particular patterns are found. The goodness-of-fit statistics reported at the
bottom of Table 10.1 show that the GAM and the GLM with a quadratic term provide
similar fits and both outperform the simple GLM model. An F -test could be used
to compare different models, though certain types of approximations are involved. A
p-value of 0.09 suggests that the additive structure can be rejected at 5% significance
level when comparing Model 2 and Model 3. However, a p-value less than 0.01
implies the superior fit of Model 3 to Model 1.
244
based on parametric distributions are less common. One can think of, for example,
the Pareto distribution. It is probably the most commonly used severity distribution,
along with the gamma, by actuaries. The Pareto distribution captures well the situation
where the majority of claims are small, but there is the potential for very large claims
(heavy tails). It is particularly prevalent in claims modeling for liability insurance
and excess of loss reinsurance. However, actuaries often need to split the data to
fit separate distributions, thereby allowing for the heterogeneity in the claims of
each subpopulation, for instance, male versus female. A more appealing approach
is to introduce covariates when fitting a Pareto distribution instead of splitting the
data. Relaxing the distributional assumption of an exponential family, this section
discusses regression techniques based on more generalized distributions that could
accommodate fat-tailed data.
10.4.1 The GB2 Distribution
This section discusses a four-parameter distributional family, the generalized beta
of the second kind (GB2), and introduces regression techniques for the GB2 model.
Introduced by McDonald (1984), the GB2 nests more than 20 distributions as limiting
cases, including many commonly used skewed distributions in actuarial science.
In general, one could use transformation techniques to create new distributions. A
GB2 distribution can be generated from a beta random variable B(1 , 2 ) as
ln Y = + ln
B(1 , 2 )
.
1 B(1 , 2 )
Here B(1 , 2 )/[1 B(1 , 2 )] follows the beta prime distribution also known as the
beta distribution of the second kind, and Y is known to follow a GB2 distribution
denoted by GB2(, , 1 , 2 ). This result implies that the GB2 distribution belongs
to the log-location-scale family, a parametric model commonly used for analyzing
duration data. Alternatively, using the relation between a beta distribution and an
F -distribution,
22 B(1 , 2 )
F (21 , 22 ),
21 (1 B(1 , 2 ))
a GB2 distribution can also be constructed from Y = exp()[1 /2 F (21 , 22 )] .
Based on these relations, the density of a GB2 distribution can be expressed as
f (y; , , 1 , 2 ) =
[exp(w)]1
y| |B(1 , 2 )[1 + exp(w)]1 +2
245
distribution being left skewed when 1 > 2 and right skewed when 1 < 2 . The kth
moment exists E(y k ) = exp(k)B(1 + k, 2 k )/B(1 , 2 ) for 1 < k < 2 .
Note that there are different ways to parameterize the GB2 distribution. We choose
this particular method to facilitate the regression analysis.
As pointed out already, the GB2 is a log-location-scale model in the sense that Z =
ln Y has density in the form f (z) = f0 ((z )/ )/ . Under this parametrization, we
have
f0 (z) =
[exp(z)]1
,
B(1 , 2 )[1 + exp(z)]1 +2
y 1
,
( + y )+1
f (y; , , ) =
exp sign()u exp(u)
y()
with = ||2 , u = (ln y )/. It can also be derived from a GB2 when =
ln(1 /2 ), = , 1 = , 2 +. Special cases include standard
gamma, log-normal, and Weibull distributions. The generalized gamma regression
has been used in health economics literature to handle skewed medical costs (for
example, see Manning, Basu, and Mullahy 2005), where both location parameter (x)
and scale parameter (x) could be allowed to vary with covariates.
246
n
zi
n
i=1
(1 + 2 )
ln yi n ln | | n ln B(1 , 2 )
i=1
n
i=1
i=1
1 + exp(zi )
i=1
ln L(|y)
=
zi + n[!(1 ) !(1 + 2 )]
ln[1 + exp(zi )]
1
i=1
i=1
n
ln L(|y)
= n[!(2 ) !(1 + 2 )]
ln[1 + exp(zi )]
2
i=1
n
n
n
exp(zi )
1
1 + 2
ln L(|y)
, j = 1, . . . , p
=
xij +
xij
j
i=1
1 + exp(zi )
i=1
of is found by solving
Thus, the maximum likelihood estimate
ln L(|y) 22
2j =j = 0, j = 1, . . . , p + 3
j
subject to constraints 1 > 0, 2 > 0, and 3 > 0. The asymptotic properties of maximum likelihood estimators could be used to construct corresponding t-ratios and
confidence intervals for model parameters. Specifically, the asymptotic covariance
could be derived from the Fisher information matrix I () and thus could
matrix of
= I 1 ().
The detailed calculation of the Fisher matrix is
)
be estimated by Acov(
provided in Appendix B.
247
As in the least squares regression, residual analysis could be used for model
diagnostics. One way to define residuals is ri = (ln yi xi )/ . Straightforward calculations show that ri are i.i.d EGB2(0, 1, 1 , 2 ). Thus the distributional assumption
could be validated by a QQ-plot:
F01
i 0.5
, r(i) , i = 1, . . . , n
n
248
Description
Mean
StdDev
age
female
married
hispanic
black
asian
edu
familysize
msa
Demographic characteristics
age in years
=1 if person is female
=1 if person is married
=1 if person is Hispanic
=1 if person is black
=1 if person is Asian
number of years of education
family size
=1 if metropolitan statistical area
42.49
0.58
0.66
0.16
0.14
0.07
13.84
2.98
0.87
12.33
limitation
chronic
excellent
verygood
good
fair
poor
hmo
0.18
1.60
0.28
0.35
0.27
0.08
0.02
0.45
employed
Employment status
=1 if employed
0.85
2.52
1.45
1.65
10000
20000
30000
40000
50000
Empirical Quantile
Empirical Quantile
Inverse-Gaussian
Density
Gamma
249
5000
10000
15000
Theoretical Quantile
Theoretical Quantile
Fig. 10.4. Distribution of medical care costs for office-based visits and corresponding
QQ plots for gamma and inverse-Gaussian distributions.
heavy-tailed data. Figure 10.4 shows the QQ-plot for both distributions. Apparently,
neither fits the tail well.
10.4.3.2 Results
Regression techniques based on generalized distributions that provide more flexibility
in modeling skewed and fat-tailed data could be useful in such cases. The GB2
regression model was fit for the medical care costs of office-based visits, and the
estimation results are presented in Table 10.3. The location parameter is parameterized
as a linear function of covariates. The t-ratios of regression coefficients show that
most of the explanatory variables are statistically significant. The effects of some
covariates are quite intuitive. For example, an elderly person is expected to have
higher medical care costs, and a healthier person, indicated by either health conditions
or self-assessed health, is expected to have lower medical costs. For the scale and
shape parameters, standard errors are reported to facilitate some hypothesis tests. For
instance, if one is interested in the null hypothesis 1 = 1 (i.e., whether the GB2
is significantly better than the Burr XII), a t-statistic, approximately calculated as
(5.883 1)/1.874 = 2.61, could be used to reject the null hypothesis. In general, the
estimates of the shape parameters demonstrate the value added by the extra complexity
of the GB2 in fitting heavy-tailed data.
Goodness-of-fit could be assessed by residual analysis. The left panel in Figure 10.5
presents fitted values versus residuals, where no particular pattern is identified. The
right panel in Figure 10.5 exhibits the QQ-plot of residuals. The alignment along the
45-degree line suggests that the GB2 distribution provides a favorable fit, especially for
the observations in the right tail. To compare with GLMs, we also report in Table 10.3
the fits from a gamma model and an inverse-Gaussian model. Not surprisingly, we
observe comparable estimates of regression coefficients between GLMs and the GB2.
This is because a logarithmic link function is used for both models. Consistent with
250
t-stat
Inverse Gaussian
Estimate
t-stat
GB2
Estimate
t-stat
5.302
0.007
0.415
0.171
0.208
0.227
0.340
0.063
0.025
0.009
0.406
0.104
0.127
0.221
0.411
0.892
0.035
0.043
42.182
4.857
14.226
4.779
4.864
5.271
6.041
10.344
2.204
0.201
10.299
9.168
3.471
5.451
6.819
7.797
1.222
1.061
5.237
0.008
0.459
0.175
0.216
0.211
0.391
0.066
0.017
0.108
0.455
0.134
0.078
0.171
0.468
0.827
0.013
0.042
23.794
3.080
8.567
2.606
2.866
2.733
4.246
6.337
0.834
1.255
5.026
5.415
1.222
2.348
3.631
2.451
0.244
0.554
3.794
0.008
0.494
0.177
0.119
0.264
0.297
0.068
0.064
0.182
0.364
0.130
0.133
0.187
0.369
0.749
0.011
0.126
14.592
5.176
15.597
4.646
2.605
5.682
4.920
10.164
5.263
3.864
8.429
10.968
3.358
4.257
5.647
5.964
0.354
2.848
Estimate
0.706
Std Error
0.010
Estimate
0.070
Std Error
0.001
Estimate
1.999
5.883
4.763
Std Error
0.337
1.874
1.461
56350.99
112739.97
112870.45
56418.18
112874.35
113004.83
55961.60
111965.20
112109.41
the QQ-plots, the goodness-of-fit statistics AIC or BIC support the better fit of the
GB2 model.
251
Fitted Value
1 0
1
Empirical Quantile
2 1 0
1
2
GB2
8
Residual
10
Theoretical Quantile
Fig. 10.5. Residual analysis of the GB2 regression. The left panel depicts residuals
versus fitted values, and the right panel is the QQ-plot of residuals.
the median of the response variable. The median regression is discussed from the
perspectives of both the frequentist and Bayesian approaches. This section concerns
the basics of quantile models that are linear in parameters.
(10.2)
n
|yi xi |
(10.3)
i=1
This model is known as median regression, where one estimates the conditional
median of the response, rather than the conditional mean as in the classic regression
model. Unlike least squares regression, the coefficient in the median regression is
estimated by minimizing the sum of absolute deviances of observed and fitted values,
leading to the LAD (which stands for least absolute deviation) estimator defined in
equation (10.3).
252
SpecifiThe rule of thumb for large samples applies to the regression median .
cally, is a consistent estimator and follows a normal distribution with mean and
asymptotic variance
1
1
=
2fi (0)xi xi
xi xi
2fi (0)xi xi
Avar()
.
Here fi is the density function associated with distribution Fi . The approximate
normality provides the basis for standard tests of statistical significance. When i are
= 2 (X X)1 , an analogy of the OLS (for
i.i.d. error, one could further show Avar()
ordinary least squares) estimator, where 2 = 1/(2f (0))2 and X X = xi xi . See
Bassett and Koenker (1978) for more details on the asymptotic properties of the LAD
estimator.
The median regression is by itself distribution-free, and the LAD estimator does
not require the distributional assumption on the error term. Compared with traditional
models, median regression enjoys certain advantages in the analysis of fat-tailed data.
First, the interpretation of parameters is somewhat simpler. Coefficients of mean
regression can be interpreted as partial derivatives. However, as noted in Section 10.2,
the situation is complicated in the case of transformation models. In contrast, there
is no such interpretation difficulty for any monotone transformation in a quantile
regression model. Second, the sensitivity of least square estimator to the outliers
makes it a poor estimator for a wide class of non-Gaussian, especially fat-tailed,
error models. Median regression is more robust than classical regressions in the sense
that the conditional median is not as sensitive to outliers as the conditional mean in
estimating the average location of a population.
where > 0 is the scale parameter, and < < is the location parameter. The
asymmetric Laplace density is defined on the whole real line, making it a candidate
to model both positive and negative responses.
253
n
i=1
(yi xi ),
(10.5)
254
) is called the
where (u) = u( I (u 0)) is known as the check function and (
th regression quantile. The optimization can be formulated as a nonlinear programming problem and is readily implemented using modern statistical packages.
Quantile regression is a natural extension of conditional mean regression to a basket
of conditional quantile functions. Apparently the regression coefficient depends on
the fraction , reflecting the partial effect of covariates on the th conditional quantile
of the response. Apparently equation (10.5) reduces to (10.3) when = 0.5. Thus the
regression median is a special case of regression quantiles. See Koenker (2005) for a
book-long review of quantile regression.
The Bayesian quantile regression could be implemented in a similar vein as median
regression. Instead of Laplace distribution, one could use a three-parameter asymmetric Laplace density. The relation between quantile regression and the asymmetric
Laplace distribution and a detailed discussion on Bayesian modeling are found in Yu
and Moyeed (2001).
255
intercept
age
female
married
hispanic
black
asian
edu
familysize
msa
limitation
chronic
verygood
good
fair
poor
hmo
employed
Bayesian
Est.
S.E.
Est.
LB
UB
3.938
0.010
0.568
0.152
0.150
0.307
0.358
0.079
0.056
0.204
0.345
0.145
0.086
0.150
0.288
0.726
0.011
0.116
0.159
0.002
0.038
0.048
0.048
0.063
0.075
0.007
0.015
0.057
0.056
0.015
0.046
0.054
0.080
0.186
0.038
0.053
3.841
0.011
0.572
0.154
0.151
0.304
0.394
0.081
0.052
0.230
0.332
0.144
0.097
0.154
0.308
0.760
0.017
0.107
3.681
0.008
0.517
0.068
0.227
0.401
0.489
0.070
0.073
0.131
0.245
0.122
0.031
0.075
0.196
0.565
0.078
0.188
3.999
0.014
0.629
0.218
0.098
0.197
0.294
0.091
0.032
0.314
0.421
0.168
0.164
0.220
0.425
0.946
0.038
0.028
intervals. These results show that the two approaches could be very consistent when
no a priori information is incorporated into the Bayesian inference. Another comparison is between the GB2 regression and the median regression. It is noticeable that the
regression coefficients from the two models are of similar size. This relation could be
seen from the following observation. In the former, the conditional mean function can
be shown to be E(log(yi )|xi ) = xi + constant. In the latter, the conditional median
function is Q0.5 (log(yi )|xi ) = xi (0.5). In this application, the responses in log scale
are more or less symmetric; thus its conditional median and mean are close, and so
are the slope parameters in the two regressions.
y+ Pr(Y2 > y)
y+ FY (y)
y+ fY (y)
2
2
lim
(10.6)
In the equation (10.6), FY () and fY () denote the survival function and density
function of random variable Y , respectively. The second equality is attained because
256
suggesting that the Pareto has a heavier tail than the Weibull. As another example,
the comparison between a Weibull and a gamma distribution could be performed by
examining the ratio of their density functions:
fgamma (y)
[ ()]1 y 1 exp(y/ )
exp(( ) ln y y/ + (y/) ).
=
fWeibull (y)
y 1 exp((y/) )
The above ratio shows that the shape of the Weibull distribution changes drastically
with the value of , with < 1 indicating a fatter tail and > 1 indicating a thinner
tail compared with the gamma distribution. In fact, the Weibull could have a thinner
tail even than the normal distribution, since
fWeibull (y)
exp(( 1) ln y + (y/) (y )2 /(2 )2 ) 0
fNormal (y)
as y + when > 2.
10.7 Appendix B. Information Matrix for GB2 Regression
defined in Section 10.4.1 could be derived
The asymptotic covariance matrix of
= I 1 (), where
from the Fisher information matrix Acov()
2
ln L(|y)
.
I () = E
The second-order derivatives of the log-likelihood function are
n
n
2 ln L(|y)
n
21
2(1 + 2 ) zi exp(zi )
=
+
z
i
2
2
2 i=1
2
1 + exp(zi )
i=1
n
1 + 2 zi2 exp(zi )
2
[1 + exp(zi )]2
i=1
2 ln L(|y)
= n[! (1 ) ! (1 + 2 )]
12
257
2 ln L(|y)
= n[! (2 ) ! (1 + 2 )]
22
n
2 ln L(|y)
exp(zi )
1 + 2
=
xij xik
2
j k
[1 + exp(zi )]2
i=1
n
n
2 ln L(|y)
1
1 zi exp(zi )
=
zi +
1
i=1
i=1 1 + exp(zi )
n
2 ln L(|y)
1 zi exp(zi )
=
2
i=1 1 + exp(zi )
n
n
2 ln L(|y)
exp(zi )
1
1 + 2
= 2
xij
xij
2
j
i=1
1 + exp(zi )
i=1
n
zi exp(zi )
1 + 2
xij
2
[1 + exp(zi )]2
i=1
2 ln L(|y)
= n! (1 + 2 )
1 2
n
n
exp(zi )
1
1
2 ln L(|y)
=
xij +
xij
1 j
i=1
i=1
1 + exp(zi )
n
2 ln L(|y)
exp(zi )
1
=
xij
2 j
i=1
1 + exp(zi )
n
2n1
2n(1 + 2 )2
[!(
)
!(
)]
+
[!(1 + 1) !(2 )]
1
2
2
2
1 2
n(1 + 2 )2 (1 + 2 1)
1 2 2
7
8
[!(1 + 1) !(2 + 1)]2 + ! (1 + 1) + ! (2 + 1)
I2,2 () = n[! (1 + 2 ) ! (1 )]
I3,3 () = n[! (1 + 2 ) ! (2 )]
(1 + 2 )2 (1 + 2 1)
xij xik , j, k = 1, . . . , p
I3+j,3+k () =
1 2 2
i=1
n
258
I1,2 () =
n
n(1 + 2 )
[!(1 ) !(2 )]
[!(1 + 1) !(2 )]
I1,3 () =
I1,3+j () =
n(1 + 2 )
[!(1 + 1) !(2 )]
1
n
n
1
(1 + 2 )2
(1 + 2 )2 (1 + 2 1)
x
+
x
+
ij
ij
2 i=1
1 2
1 2 2
i=1
n
xij , j = 1, . . . , p
i=1
I2,3 () = n! (1 + 2 )
n
2
xij , j = 1, . . . , p
I2,3+j () =
1 i=1
2
1
xij , j = 1, . . . , p
I3,3+j () = (1 + )
1 i=1
n
References
Bartlett, M. (1947). The use of transformations. Biometrics 3(1), 3952.
Bassett, G., Jr. and R. Koenker (1978). Asymptotic theory of least absolute error regression.
Journal of the American Statistical Association 73(363), 618622.
Beirlant, J., Y. Goegebeur, R. Verlaak, and P. Vynckier (1998). Burr regression and portfolio
segmentation. Insurance: Mathematics and Economics 23(3), 231250.
Bickel, P. and K. Doksum (1981). An analysis of transformations revisited. Journal of the
American Statistical Association 76(374), 296311.
Box, G. and D. Cox (1964). An analysis of transformations (with discussion). Journal of the
Royal Statistical Society: Series B (Methodological) 26(2), 211252.
Burbidge, J., L. Magee, and A. Robb (1988). Alternative transformations to handle extreme
values of the dependent variable. Journal of the American Statistical Association
83(401), 123127.
Frees, E., J. Gao, and M. Rosenberg (2011). Predicting the frequency and amount of health
care expenditures. North American Actuarial Journal 15(3), 377392.
Frees, E., P. Shi, and E. Valdez (2009). Actuarial applications of a hierarchical claims model.
ASTIN Bulletin 39(1), 165197.
Frees, E. and E. Valdez (2008). Hierarchical insurance claims modeling. Journal of the
American Statistical Association 103(484), 14571469.
Frees, E. and P. Wang (2005). Credibility using copulas. North American Actuarial
Journal 9(2), 3148.
John, J. and N. Draper (1980). An alternative family of transformations. Applied
Statistics 29(2), 190197.
Klugman, S., H. Panjer, and G. Willmot (2008). Loss Models: From Data to Decisions (3nd
ed.). Wiley.
Koenker, R. (2005). Quantile Regression. Cambridge University Press, Cambridge.
References
259
11
Spatial Modeling
Eike Brechmann and Claudia Czado
Chapter Preview. This chapter presents statistical models that can handle spatial
dependence among variables. Spatial dependence refers to the phenomenon that variables observed in areas close to each other are often related. Ignoring data heterogeneity due to such spatial dependence patterns may cause overdispersion and erroneous
conclusions. In an actuarial context, it is important to take spatial information into
account in many cases, such as in the insurance of buildings threatened by natural
catastrophes; in health insurance, where diseases affect specific regions; and also in
car insurance, as we discuss in an application.
In particular, we describe the most common spatial autoregressive models and show
how to build a joint model for claim severity and claim frequency of individual policies
based on generalized linear models with underlying spatial dependence. The results
show the importance of explicitly considering spatial information in the ratemaking
methodology.
11.1 Introduction
It is important to take spatial information related to insurance policies into account
when predicting claims and ratemaking. The most prominent example is the modeling
of natural catastrophes needed for the insurance of buildings. Another example is
health insurance, where spatial information is relevant for an accurate assessment
of the underlying risks, because frequencies of some diseases may vary by region.
Frequencies in neighbor regions are often expected to be more closely related than
those in regions far from each other. This phenomenon is usually referred to as spatial
dependence.
Statistically speaking, spatial dependence means that the joint behavior of variables
in different regions depends on some kind of underlying distance, the simplest example
being the Euclidean distance between two points. If variables in different regions do
not behave independently, we speak of a spatial dependence pattern. This pattern
260
11.1 Introduction
261
may be the reason for unobserved data heterogeneity, which causes overdispersion.
Overdispersion denotes the situation when a higher variance is observed than assumed
by the model. In this chapter we demonstrate how such spatial dependence patterns
of regions can be accounted for appropriately in an actuarial context. By precisely
and flexibly specifying a neighborhood structure of regions, the presented methods go
beyond the classical use of territory variables as categorical covariates in regression
models, which do not represent explicit territorial characteristics, so that, for example,
topographic effects are ignored. Detailed statistical treatments of spatial modeling can
be found in the seminal book by Cressie (1993) and also in newer volumes such as
Banerjee, Carlin, and Gelfand (2003), Gelfand et al. (2010), and Cressie and Wikle
(2011).
To make things more concrete: What exactly do we mean when we talk about
spatial data? For this, we come back to the earlier mentioned examples. First, when a
natural catastrophe such as a flood occurs, it usually does not affect isolated buildings.
Therefore, claims due to a natural catastrophe for a portfolio of building insurance
contracts are generally not independent. For example, they might depend on the
distance of the building to the next river. The statistical challenge here is in predicting
flood levels at any location in an area of interest. Then, claim frequency and size can
easily be predicted for any building in the portfolio. For obvious reasons, such data
are called point-level or point-referenced data.
Second, the exact location of the occurrence of a hurricane, a wildfire, or an
earthquake is usually unknown and can be considered as random. In this case, the set
of locations, at which events occur, is random itself. The observed data then simply
are equal to 1 for all locations in this set and are called point-process data. In addition,
the data can be supplemented by covariate information, such as the damage caused
by the events (marked point-process data).
Third, the frequency of a disease can hardly be assessed at each single location
in a region. In such cases, one therefore partitions the area of interest into a certain
number of areal locations, such as cells of a grid or geographical regions, which are
often directly available because of the type of data. As an example, one may think
of skin cancer rates observed in all 50 U.S. states. The quantity of interest is then
the sum or average of variables in each areal location. Such data are called areal or
lattice data and are the focus of this chapter. For the sake of simplicity, we refer to
areal locations as regions for the remainder of this chapter. More information on
the other common types of spatial data and statistical models for them can be found
in the earlier mentioned references (see, e.g., Cressie 1993, chapter 1, and Banerjee
et al. 2003, chapter 1).
A characteristic feature of areal data is its neighborhood structure. This means
that for each region we can identify a set of neighbor regions, where the notion of a
neighbor can have different meanings. Most often, neighbors are those regions that
262
Spatial Modeling
share a common border, but the set of neighbors may, for instance, also be defined as
the regions within a certain distance of the region of interest.
As an illustrative example, we analyze the full comprehensive car insurance portfolio of a German car insurance company in the year 2000. The dataset includes personal
information such as age and gender of the policyholders as well as policy-specific
details such as distance driven per year, deductible, and type and age of the car. As in
Gschlol and Czado (2007), we focus on a subset of the data with claims from traffic
accidents of policyholders with three models of mid-sized cars. This subset of the
data still contains more than 350,000 observations.
According to where policyholders live, they can be attributed to 440 regions in
Germany. Clearly, we expect a certain degree of heterogeneity in claim sizes and
frequency among the regions. The question that we investigate here is whether there
is a spatial pattern in this heterogeneity, which is important to take into account to
ensure accurate ratemaking. The left panel of Figure 11.1 shows the average claim size
per region. This gives a first indication for spatial dependence: in eastern Germany
claim sizes appear to be larger on average than in the middle and in the South-West
of Germany. Because some of these effects are possibly due to the population density
of the regions, we discuss ways to adjust for this factor.
The average claim size per region is analyzed in a first step in Section 11.4. In Section 11.5 we then describe a more elaborate model for the total loss per region, which
combines separate generalized linear models (GLMs) for frequency and severity and
accounts for spatial dependence in the data. The exploratory analysis of spatial data
is treated in Section 11.2, whereas while appropriate spatial dependence models are
discussed in Section 11.3.
11.2 Exploratory Analysis of Spatial Data
In this section we only briefly touch on the exploratory analysis of spatial data.
Detailed discussions can be found in any book on spatial statistics (see, in particular,
Cressie 1993, and Banerjee et al. 2003).
Let Y = (Y1 , . . . , Yn ) be a vector of random variables observed in n different
regions. One of the first steps and at the same time one of the most instructive ones
in such an exploratory analysis is the graphical illustration of the data Yi , i = 1, . . . , n,
as in the left panel of Figure 11.1. Here, observations in each regions are color coded
so that it is easy to assess whether there is a spatial dependence pattern with similar
values in close regions.
Analytically, spatial association can be measured, for example, using Morans I or
Gearys C, which can be regarded as spatial equivalents of common tools in time series
analysis (the autocorrelation coefficient and the Durbin-Watson statistic, respectively).
To define the two quantities, we need to introduce the concept of the adjacency matrix,
263
0.5
498
5
6
7
8
Population density in log form
10614
Fig. 11.1. Left panel: Average claim size per region in Germany. The minimum
average claim size is 498.17, and the maximum is 10 613.97. Right panel: Fitted
third-order cubic smoothing spline of the population density in log form as covariate
for average claim sizes also in log form. Dotted lines are pointwise estimated twicestandard-error curves (assuming independent and identically distributed data, and
hence ignoring potential spatial dependence).
1, j N (i)
0, j N(i)
(11.1)
264
Spatial Modeling
association. Morans I is, however, not restricted to [1, 1], although values usually
fall into this range.
Alternatively, Gearys C is available as
C =
n1
n n
i=1
j =1
n n
j =1 wij (Yi
n
2
i=1 (Yi Y )
i=1
wij
Yj )2
[0, 2].
265
p( y)
,
p( yi )
where yi is the vector y with the ith observation removed; that is, yi =
(y1 , . . . , yi1 , yi+1 , . . . , yn ) . We assume that the conditional distribution of Yi given
all other variables depends only on the neighbors N(i) of Yi ; that is,
p(yi | yi ) = p(yi |yj N(i)),
i = 1, . . . , n.
Especially when the number of regions n is large, this assumption significantly reduces
the complexity compared to working with the n-dimensional joint distribution directly.
Due to the Markov property known from time series analysis (future values depend
only on the present value and not on past values), which is here induced by the
neighborhood structure, the random vector Y is then called a Markov random field
(see, e.g., Rue and Held 2005, for more details).
The question is, however, whether the collection of conditional distributions
p(yi | yi ), i = 1, . . . , n, uniquely identifies the joint distribution p( y) for all y;
that is, whether the full conditionals are compatible. Although the converse is always
true, particular conditions have to be met to ensure compatibility. One such case is
discussed here.
Definition 11.1 (Conditionally autoregressive model). In the conditionally autoregressive (CAR) model by Besag (1974), the full conditional distribution of
Yi |{yj , j N (i)}, i = 1, . . . , n, is given by a normal distribution with mean
266
i +
Spatial Modeling
j ) and variance i2 ,
cij (yj j ), i2 ,
Yi |{yj , j N(i)} N i +
j N(i)
(11.2)
1 1
1
M
(I
MW
)
=
(M 1 W ).
2
2
267
where 0 = (0, . . . , 0) and 1 = (1, . . . , 1) . Hence, the joint distribution of Y is not
proper. Nevertheless, the IAR model can still be used for conditionally specified random effects in a hierarchical model specification. The model discussed in Section 11.5
provides an example of such improper prior distributions yielding a proper posterior
distribution. In general, the availability of the full conditionals makes CAR models
very attractive in the context of Markov chain Monte Carlo (MCMC) methods, as
presented in Chapter 13.
One convenient way to remedy this shortcoming of the IAR model is by adding a
parameter [0, 1] to the specification of the spatial dependence matrix C; that is,
we set C = MW (see, e.g., Sun, Tsutakawa, and Speckman, 1999). If = 1, this is
the IAR model. If, however, < 1, then (M 1 W ) is guaranteed to be nonsingular
and we obtain a proper distribution. Although the parameter should not be confused
with a correlation parameter, it controls the degree of spatial dependence between the
regions. In particular, the case = 0 implies independence among the regions.
(11.3)
This implies that the errors CAR are not independent. Expression (11.3) therefore
motivates the consideration of an alternative model, which can be considered as the
spatial equivalent of autoregressive models in time series analysis.
Definition 11.2 (Simultaneously autoregressive model). The simultaneously autoregressive (SAR) model is defined as
SAR := (I B)(Y ) Nn (0, D ),
(11.4)
268
Spatial Modeling
i = 1, . . . , n.
j N(i)
i = 1, . . . , n,
269
Table 11.1. Summary of the Main Model Components of the CAR and the SAR Model
Model
CAR
SAR
Var(Y )
Var(AR )
(I C)1 M
M(I C)
(I B)1 D (I B )1
if and only if
(I C)1 M = (I B)1 D (I B )1 .
Cressie (1993) shows that any SAR model can be represented as a CAR model,
whereas the converse is not generally true. Resuming the discussion of the proximity
matrices B and C, it is clear that B and C cannot have the same interpretation if the
CAR and the SAR model coincide. For further insights and comparisons between SAR
and CAR models, we strongly recommend the work by Wall (2004), who points out
potential non-intuitive results implied by either models using an illustrative example.
Here, we use the scaled normalized adjacency matrix (see (11.1)) as a proximity matrix
in both models. In particular, = 0 ( = 0) in the CAR (SAR) model corresponds
to spatial independence, since then the proximity matrix C (B) is a matrix of zeros.
For the mean, we investigated different specifications: first, an intercept model with
i = 0 , i = 1, . . . , n, and, second, based on the exploratory analysis in the right
panel of Figure 11.1, covariate models with intercept 0 and orthogonal polynomials
270
Spatial Modeling
Table 11.2. Estimated Parameters and Standard Errors of Linear, CAR, and SAR
Models for Average Claim Sizes with Intercept and Covariate Models for the Mean
LM
0
Inter.
8.484
(0.015)
8.484
(0.015)
8.484
(0.015)
8.484
(0.015)
1.373
(0.312)
1.373
(0.308)
1.373
(0.308)
1.011
(0.308)
1.011
(0.308)
0.125
(0.308)
CAR
0
1
2
3
Inter.
8.468
(0.021)
8.484
(0.015)
8.484
(0.015)
8.484
(0.015)
1.421
(0.311)
1.415
(0.307)
1.407
(0.307)
1.009
(0.307)
1.010
(0.307)
SAR
0
1
2
Inter.
8.479
(0.019)
8.483
(0.018)
8.483
(0.018)
8.483
(0.018)
Lin.
Quad.
Cubic
Lin.
Quad.
Cubic
Lin.
Quad.
Cubic
1
1.394
(0.327)
1.420
(0.323)
1.411
(0.323)
2
0.993
(0.323)
0.977
(0.324)
3
Log lik.
AIC
BIC
119.8
243.6
251.8
110.3
226.5
238.8
104.9
217.8
234.2
104.8
219.7
240.1
Log lik.
AIC
BIC
114.0
234.1
246.4
110.3
228.5
244.9
104.9
219.8
240.3
0.142
(0.308)
0.463
(0.126)
0.001
(0.018)
0.001
(0.006)
0.001
(0.012)
104.8
221.7
246.2
3
Log lik.
AIC
BIC
114.7
235.4
247.7
105.8
219.6
236.0
101.1
212.3
232.7
0.178
(0.312)
0.220
(0.067)
0.208
(0.068)
0.196
(0.070)
0.197
(0.070)
101.0
214.0
238.5
Note: The last columns show log likelihoods, AICs, and BICs of the estimated models.
up to order 3 of the population density in log form for each region i. Associated
regression parameters are denoted by 1 , 2 , and 3 .
The models are fitted using the function spautolm in the R-package spdep by
Bivand et al. (2013), and estimated parameters and standard errors are reported in
Table 11.2, which also shows fits of linear models (LMs) with corresponding mean
specifications.
The estimation results underline the differences between both models. Although
estimates for the intercept and for the influence of the population density are very
similar across both models, there are striking differences in the interpretation of
271
the degree of spatial dependence. In the SAR model, the inclusion of the spatial
information regarding the population density of each region leads to a decrease of up
to 11% in the estimated parameter
. This means that, according to this model, the
population density accounts for a certain amount of spatial heterogeneity, but not for
all of it. This is in line with the empirical Morans I and Gearys C of the residuals of
the GAM fit in Section 11.2.
The situation is completely different for the CAR model. Here, most of the spatial
variability can be accounted for by including the population density as a covariate. In
other words, according to the CAR model, the spatial pattern observed in the average
claim sizes can be attributed completely to differences in the population density. Note
that the parameters and cannot be compared directly because of the different
model formulations. For a more detailed discussion on this see Wall (2004).
indicates a potential misspeciThe extreme change in the parameter estimate
fication of the CAR model. The log likelihoods, AICs, and BICs confirm that the
SAR model provides a better fit for the data. The fit is also superior to standard
linear regression models with the same covariate specification, which ignore spatial
dependence. Figure 11.2 shows fitted average claim sizes per region according to the
CAR and SAR models with intercept model and with quadratic covariate model for
the mean, since the inclusion of the population density as a covariate significantly
improves the model fits. In particular, a quadratic effect of the population density is
sufficient, as also indicated in Figure 11.1.
First, looking at the CAR and the SAR model with the pure intercept model (left
column of Figure 11.2) shows that the degree of smoothness implied by the SAR
model is stronger than for the CAR model. Nevertheless, both models are able to
reproduce the main spatial characteristics observed in the data: for instance, there are
high average claim sizes in the East and lower ones in the South-West.
When covariate information regarding the population density is included in the
model (right column of Figure 11.2), the results of the CAR and the SAR model
become closer, although the CAR almost assumes spatial independence, whereas
there is still spatial dependence modeled in the SAR model. The main conclusion
from this analysis is therefore that the population density plays an important role
for explaining average claim sizes. In particular, average claim sizes in mid-sized to
larger cities appear to be smaller than those in rural regions of Germany.
Overall, however, the model fits are unsatisfactory. The empirical correlation
between observed and fitted values is 26% under spatial independence and can only be
increased to 30% by the SAR model with a covariate model for the mean. In the next
section, we therefore investigate the benefit of explicitly taking into account spatial
information in a more comprehensive modeling approach, including more relevant
covariates.
272
Spatial Modeling
3000
7000
3000
7000
3000
7000
3000
7000
Fig. 11.2. Fitted average claim sizes per region according to CAR (first row) and
SAR (second row) models with intercept model (first column) and quadratic covariate
model (second column) for the mean. Note that the color coding is different from the
left panel of Figure 11.1 in order to emphasize differences more distinctly.
273
274
Spatial Modeling
model-building strategy), where r(i) denotes the region of the ith policyholder. Using
this notation, we extend the common Poisson regression model (see Chapter 4) for
the number of claims by including spatial random effects. For this, let
Ni Poisson(N
i ),
i = 1, . . . , m,
(11.5)
Here, Ei denotes the exposure, xi = (1, xi1 , . . . , xik ) , the covariate vector associated
with policyholder i, and = (0 , . . . , k ) , the vector of unknown regression parameters. Further, 1 , . . . , R are random effects corresponding to the R regions. They
induce, through r(i) , spatial dependence among the variables N1 , . . . , Nm .
In Section 11.3 we discussed appropriate models for spatial autoregression, in
particular the CAR and the SAR model. Because of the availability of the full conditionals, we opt here for a CAR model, but with a more sophisticated specification of
the proximity matrix C as proposed by Pettitt, Weir, and Hart (2002), which yields a
proper joint distribution in contrast to the IAR model discussed in Section 11.3.1.
According to Pettitt et al. (2002), we set in the general CAR model formulation
(11.2),
cj k =
wj k
,
1 + ||nj
j, k = 1, . . . , R,
if k = j and 0 otherwise, where wj k are the elements of the adjacency matrix (as
defined in (11.1)) and R. As before, we denote by nj := |N(j )| the number of
neighbors of region j {1, . . . , R}.
Further, we define M = 2 M with mjj = (1 + ||nj )1 for j = 1, . . . , R. These
choices imply that cj k k2 = cj k j2 holds and thus Q := M 1 (I C) is symmetric.
Pettitt et al. (2002) also show that Q is positive definite with entries
Qj k =
1 + ||nj ,
j =k
wj k ,
j = k
(11.6)
which is, in contrast to the IAR specification, a proper joint distribution of , since
the inverse of a positive definite matrix is again positive definite. Note that we assume
= 0 here, because the mean is modeled explicitly through the regression parameters
in (11.5).
275
i = 1, . . . , m.
For both approaches, we choose appropriate extended Gamma regression models (see
Chapters 5 and 6) that include spatial random effects.
11.5.2.1 Individual Claim Size Model
For the individual claim sizes Sij , j = 1, . . . , Ni , we assume that they are independently Gamma distributed given the number of claims Ni ,
Sij |Ni Gamma(Si , ),
j = 1, . . . , Ni , i = 1, . . . ., m,
.
exp
f (s|, ) =
()
Then mean and variance of the individual claim size model are given as
E(Sij |Ni ) = Si ,
(Si )2
.
Similar to the model for the number of claims (11.5), a regression on the mean Si is
considered. In particular, we choose a log link and study the following specification
S
),
Si = exp(vi S + r(i)
i = 1, . . . , m,
(11.7)
276
Spatial Modeling
1, Ni = j
0, otherwise
i = 1, . . . , n, j = 2, . . . , N ,
(11.8)
2
(A
i )
,
Ni
and therefore
Si |Ni Gamma(A
i , Ni ),
i = 1, . . . ., m.
As in (11.7) we use a log link for regression on the mean and set
A
A
vi A + r(i)
),
i = exp(3
i = 1, . . . , m,
where 3
vi = (3
vi1 , . . . ,3
vip ) is a covariate vector, which is potentially different from
v, and A = (1A , . . . , pA ) are the associated unknown regression coefficients. As
earlier, the number of claims may be used as a covariate here in order to introduce
dependence among frequency and severity. For the vector of the random effects
A = (1A , . . . , RA ) , a CAR specification as in (11.6) is assumed.
11.5.3 Application
Having described the two-component hierarchical model for the total claim size, we
now apply it to the more than 350,000 policies of the German car insurance dataset. For
parameter estimation, we follow a fully Bayesian approach. Parameters are estimated
using Markov chain Monte Carlo (MCMC) techniques, as presented in Chapter 13.
277
Table 11.3. DIC and Scoring Rules for the Total Claim Size Model with and Without
Frequency-Severity Dependence and Spatial Effects
Freq.-sev.
Spatial
yes
yes
no
no
yes
no
yes
no
DIC
LS
IS0.5
IS0.1
CRPS
269092
269136
269122
269175
9.5642
9.5699
9.5669
9.5734
55760
56125
55805
56170
20412
20526
20434
20575
2471.9
2481.8
2474.2
2484.3
Explicit prior choices for the frequency and the average claim size models can be
found in Gschlol and Czado (2007).
According to a standard Poisson regression analysis ignoring spatial dependence,
we include an intercept as well as 17 other variables such as age and gender of the
policyholder and deductible as covariates for the frequency model. Most importantly,
we also include the population density as a covariate to account for the heterogeneity
identified in Section 11.4. Similarly, for the average claim size model, an intercept and
a range of 14 covariates are considered after a pre-analysis using a standard Gamma
regression model. Furthermore, we include three indicator variables for the number
of claims (as defined in (11.8)).
We found a negative relationship between the number of claims and the average
claim sizes. This means that the larger the number of claims that are observed the
lower the average claim size. A possible interpretation could be that, although more
claims occur in cities, they are usually less severe car body incidents in contrast to
accidents in more rural areas.
In terms of spatial dependence, we found significantly less accidents in the East
and the North than in the South. Conversely, average claim sizes tend to be larger
in the East and smaller in the South. Overall, the spatial dependence of the average
claim size is mostly weak given the information about the population density. This is
in line with the findings in Section 11.4.
Using different model evaluation criteria, the inclusion of an explicit spatial dependence model is in fact shown to substantially improve the model fit. In particular,
we consider the deviance information criterion (DIC) by Spiegelhalter et al. (2002)
(see Chapter 14), which also takes into account the model complexity, and scoring
rules for assessing the quality of probabilistic forecasts (see Gneiting and Raftery
2007): a scoring rule assigns a specific score to a model, which is highest for the
true model. Typical choices are the logarithmic score (LS), the interval score at level
(0, 1) (IS ), and the continuous ranked probability score (CRPS). Table 11.3,
which is reproduced from Gschlol (2006), shows the DIC, LS, IS0.5 , IS0.1 , and CRPS
of our hierarchical model for the total claim size with and without frequency-severity
278
Spatial Modeling
dependence and spatial effects. From these numbers, the dependence between claim
sizes and numbers of claims is also determined to be significant.
The benefit of explicitly modeling spatial dependence is further shown in an analysis of the predictive distribution of the total claim size per region. Ignoring spatial
information can lead to an underprediction of total claim sizes of up to 21% in certain regions and, at the same time, to an overprediction of 31% in other regions (see
Gschlol and Czado 2007, section 4.4). Thus, spatial dependence clearly needs to be
accounted for in the ratemaking process.
11.6 Discussion and Conclusion
In this chapter we presented models that can account for spatial dependence of areal
data, as it is often encountered in an actuarial context. We discussed and applied the
model classes of CAR and SAR models to the modeling of average claim sizes in a
dataset of German car insurance policies. We also used a specific CAR model as a
random effects model to induce spatial dependence into the modeling of claim size
and frequency through dependent regression models.
In a regression context, an appropriate covariate with spatial information may not
always be identifiable or available (e.g., due to privacy concerns). Then, a CAR or a
SAR model can be used as an exploratory tool to detect and analyze potential spatial
dependence in the data.
More on spatial count regression models can be found in Gschlol and Czado (2008)
and Czado, Schabenberger, and Erhardt (2013), as well as in the related R-package
spatcounts by Schabenberger (2009). The modeling of a binary response variable
with underlying spatial effects has been studied by Czado and Prokopenko (2008).
The class of so-called structured additive regression models (see, e.g., Fahrmeir et al.
2013) provides a comprehensive approach to handling spatial effects in regression
problems. Such models, which include GAMs as a special case, can be estimated
using Bayesian techniques with the software tool BayesX by Belitz et al. (2012).
References
Banerjee, S., B. P. Carlin, and A. E. Gelfand (2003). Hierarchical Modeling and Analysis for
Spatial Data. Chapman & Hall, London.
Belitz, C., A. Brezger, T. Kneib, S. Lang, and N. Umlauf (2012). BayesX Software for
Bayesian inference in structured additive regression models. Version 2.1,
http://www.stat.uni-muenchen.de/~bayesx.
Besag, J. (1974). Spatial interaction and the statistical analysis of lattice systems. Journal of
the Royal Statistical Society, Series B 36(2), 192236.
Besag, J., J. York, and A. Mollie (1991). Bayesian image restoration with two applications in
spatial statistics. Annals of the Institute of Statistical Mathematics 43(1), 159.
Bivand, R. et al. (2013). spdep: Spatial dependence: weighting schemes, statistics and
models. R package, http://CRAN.R-project.org/package=spdep.
References
279
12
Unsupervised Learning
Louise Francis
280
12.1 Introduction
0.15
281
0.05
0.05 0.10
S & P Return
proxied by the return on a widely recognized index composed of many stocks, such
as the S&P 500 (see Figure 12.1 where an individual stocks return is regressed on
the market return2 ).
Unsupervised learning, a term coined by artificial intelligence professionals, in
contrast, does not involve dependent variables and predictors.
One of the earliest uses of unsupervised learning was in the social and behavioral
sciences (for example, see Gorden 1977; Maranell 1974). These early applications
used data collected (often from questionnaires) about attitudes, beliefs, and behaviors
and constructed sociometric and psychometric scales from the data. These scales had
no dependent variable per se. A system was used to convert the answers to a set of
questions, rankings, and so on, into a single measure believed to quantify a social or
psychological concept such as occupational prestige.
Another early application of unsupervised learning in science was in the classification of plants and animals. Data on plant and animal characteristics are collected
and used to create a taxonomy. All plants and animals in a particular group, such as
species, are expected to be similar to each other and dissimilar to the members of
other groups. Many of the techniques discussed in this chapter fall into one of these
two types: (1) a scale or quantitative factor based on a weighted sum or average of
variables or (2) a classification scheme that organizes the data into (usually) distinct
groups based on features they have in common.
In insurance a common example of unsupervised learning is the construction of
territories. In certain lines of business such as automobile insurance, it is common
to include territorial geographic factors in computing premium rates (Zhang, Wang,
and Song 2006). Policyholders are typically included in territories that have some
geographic proximity to each other, though increasingly other factors, such as demographic information, are considered in constructing the territories. The key idea is that
the policyholders are grouped together based on characteristics they have in common
(i.e., similarity on geographic and perhaps other variables). These groupings then can
2
Data for Selective Insurance Company and the S&P 500 returns were downloaded from the Yahoo Finance website.
282
Unsupervised Learning
CSL
1,000,000
1,000,000
500,000
1,000,000
1,000,000
1,000,000
1,000,000
1,000,000
500,000
1,000,000
Initial Indemnity
Reserve
Initial Expense
Reserve
Initial
Reserve
1,000
150,000
7,500
5,000
10,000
17,500
65,000
75,000
5,600
15,500
1,000
35,000
2,000
185,000
7,500
5,000
20,000
21,000
65,000
100,000
5,600
15,500
10,000
3,500
25,000
be used for other purposes such as rating policies or marketing products. Other areas
that unsupervised learning has been applied in insurance include the following:
r fraud classification (Brockett et al. 2002; Polon 2008)
r compressing a set of policies to fewer cells for reserving (Freedman and Reynolds 2009)
r pricing or marketing applications, classifying patterns of persistency of life insurance
policies (Brzezinski 1981)
r employer rating of actuarial performance (Erin Research 2003)
r predicting mortgage default (Francis and Prevosto 2010)
r variable reduction for predictive modeling (Sanche and Lonergan 2006)
r analyzing the efficiency of financial performance of insurance companies (Mahmoud
(2008)
r text mining (Francis and Flynn (2010).
A key concept in unsupervised learning is dimension reduction. Table 12.1 displays
a snapshot of a typical claims database, in this case closed liability claims posted on
the Texas Department of Insurance website. Each column in the exhibit represents
a variable or feature. In a typical database used in predictive modeling, there may
be hundreds or even thousands of features. Unsupervised learning can be used to
condense the features into a much smaller set of factors or components that capture
most of the important information in the many variables. (Editors note: in this context,
factor refers to an unobserved, or latent, variable. Sometimes, a factor refers to a
categorical variable as in the statistical program R.)
Each row in the data represents a record, here a liability claim. In a large dataset
one may wish to classify similar records together. For instance, an analyst may want
to group all records that are near each other geographically into territories for
further analysis. In classical statistics, variable reduction techniques known as factor
12.2 Datasets
283
analysis or principal components analysis have been used for variable reduction,
whereas cluster analysis has been used for record reduction.
In sociology and psychology one of the earlier uses of unsupervised learning was
to develop scales. The scales were intended to measure concepts such as intelligence,
ability, and attitude that could not be directly observed or measured, but could be
considered related to observed data that could be collected. A number of these scales,
such as the Likert scale, are based on summing the responses to items on a survey or
questionnaire. For instance, the Likert scale typically requires the respondent to select
from a range of five ordered options (say from strongly agree to strongly disagree).
The answers to the questions can then be added (see Coaley 2010 for an introduction).
The items on the questionnaire are often subject to further analysis to validate that
they in fact belong in the questionnaire and are related to the concept being measured.
One approach is to look at the correlation of the responses on a given question to those
of other questions. Items with low correlations may then be rejected. Unidimensional
methods such as the Likert scale attempt to measure a single dimension or concept.
A similarity between psychometric and sociometric scaling and some of the variable
reduction methods in this chapter such as factor analysis and principal components
analysis is the use of scores based on summed (possibly weighted) values of the items
contributing to the scores. In fact, Spearmans development of factor analysis was
largely due to his work on intelligence scales (p. 150 of Coaley 2010). According
to Coaley, Spearman believed that intelligence was attributable to a general factor
that he labeled g that enables people to solve problems in a variety of settings.
However, others proposed multiple kinds of intelligence. Factor analysis has been
used to identify multiple components of a more general concept such as intelligence.
Another similarity is the use of correlation matrices in the construction of the
scales. This is pursued more thoroughly in the next section on factor analysis and
principal components.
The following sections present the data for this chapters illustrated examples and
then the major unsupervised learning approaches. The chapter provides an introduction to the major kinds of dimension reduction (in the column dimension/reduction
of variables and in the row dimension/reduction of records). More advanced forms of
unsupervised learning, such as k-nearest neighbor, Kohonen neural networks, association rules, multidimensional scaling, and PRIDIT analysis, are not covered.
12.2 Datasets
Example 12.1 (California Auto Assigned Risk (CAARP) Data). This assignedrisk automobile data were made available to researchers in 2005 for the purpose of
studying the effect of constraining the weight of territorial variables on California personal auto premiums. The data contain exposure information (car counts, premium)
284
Unsupervised Learning
and claim and loss information (bodily injury (BI) counts, BI ultimate losses,
property damage (PD) claim counts, PD ultimate losses). From these data we have
constructed additional variables such as frequency (claim counts/car counts) and
severity (ultimate losses/claim counts). The data also contain fields computed by a
consultant that capture the effect of proposed regulatory changes (Newpremium1,
Newpremium2, etc.). Each record in the data represents a zip code. To create examples
that are simple and somewhat exaggerated (compared to the original data), we have
also simulated some data roughly based on relationships observed in the original
data, but with some parameters exaggerated/ altered to better illustrate our concepts.
The code used to create the simulated data is provided on the web site for this book.
Example 12.2 (Texas Closed Claim Data). Download from: http://www.tdi.texas.
gov/reports/report4.html.
These data are collected annually by the Texas Department of Insurance on closed
liability claims that exceed a threshold (i.e., 10,000). The liability claims are from a
number of different casualty lines, such as general liability, professional liability, and
so on. A key that is downloaded with the data provides information on the variables
included in the data. The data include information on the characteristics of the claim,
such as report lag, injury type, and cause of loss, as well as data on various financial
values such as economic loss, legal expense, and primary insurers indemnity. Due
to the truncation models, fit to the data cannot be considered representative of all
ground-up losses. However, these data are useful for illustrating and testing various
predictive modeling techniques.
Example 12.3 (Simulated Automobile PIP Questionable Claims Data). Francis
(2003, 2006) used a Massachusetts Automobile Insurers Bureau research dataset
collected in 1993 for research purposes to illustrate data mining techniques, including
clustering and their application to modeling questionable claims3 for further action,
such as referral to a special investigation unit. We use the description of the data as well
as some of the published statistical features to simulate automobile PIP fraud data.
These data contain numeric (number of providers, number of treatments, report lag)
and categorical variables (injury type, provider type, whether an attorney is involved,
whether a police report was filed, whether claimant was treated in the emergency
room).
Example 12.4 (Insurance and Economic Indices). These data contain various
insurance and economic inflation indices, including workers compensation severity,
3
Only a small percentage of claims suspected of not being legitimate meet the legal definition of fraud, so such
claims are typically not labeled or referred to as fraud.
285
health insurance, the consumer price index, and various medical-related components
of the producer price index. The indices tend to be highly correlated and are useful
illustrations of variable reduction methods.
12.3 Factor and Principal Components Analysis
12.3.1 R Princomp and Factanal Functions
This section uses the R princomp and factanal functions. Venables, Ripley,
and Venables (1999) describe the application of princomp to principal components
analysis, whereas Crawley (2007) describes the application of factanal to factor
analysis.
Smith (2002) offers an introductory tutorial to principal components analysis
(PCA). This tutorial contains a very simple example, including a basic introduction to some needed linear algebra. It then shows an application to computer vision.
Kim and Mueller (1978) provide an excellent introduction to factor analysis aimed
at social scientists. This introduction covers additional details (such as the topic of
factor rotation, which is mentioned briefly in the discussion of factor analysis) that
because of the need for brevity, cannot be covered here.
Factor analysis allows us to do variable reduction along the variable dimension.
Like the related method, principal components, it can be used to combine many
variables into one or perhaps several variables; thus hundreds of variables can be
transformed into a handful of predictors. Sometimes the end result is the factors or
components themselves, but often they become inputs to a prediction model.
The conceptual model for factor analysis with one factor is4
xi = bF + ui ,
where x is a variable, b a loading, F a factor, and u a unique component. Because
underlying the observed variables is a single unobserved factor, the observed variables
are expected to be correlated due to their correlation with the common factor, and the
correlation matrix typically plays a role in estimating the factors.
Factor analysis is an unsupervised learning technique that analyzes relationships
among a set of variables by reducing the original (manifest) variables to a smaller set of
latent (hypothetical) variables. For instance, it is commonly believed in the insurance
industry that annual trend costs for many property and casualty lines of business, such
as workers compensation, are higher (and sometimes lower) than overall economic
inflation rates observed in the economy, such as payroll inflation and medical cost
inflation. This excess cost is often referred to as social inflation. Social inflation is
more of a concept than an actual empirical item that can be measured. Suppose we
4
286
Unsupervised Learning
One-Factor
Model
Social Inflation
Litigation
Rates
U1
Size of Jury
Awards
U2
Index of State
Litigation
Environment
U3
consider social inflation to be the unobserved or latent variable and wish to measure
it through factors that can be observed (as in Figure 12.2). In this simple illustration
we assume three observed variables that can be measured: litigation rates, an index
of the legal environment in the state, and the size of jury awards (see Francis 2001,
for a fuller development of this particular example).
The factor analysis procedure is applied to data collected on the observed variables
to infer values for the unobserved variable.
PCA is a variable reduction technique that does not require the assumption that
a latent factor causes the observed results. It simply assumes that two or more variables can be decomposed into components that capture the key features of the data.
Figure 12.3 shows an example based on two different medical cost indices. (The two
indices, general medical and surgical costs and physician services costs, were downloaded from the Bureau of Labor Statistics website, www.bls.gov, from the producer
price section.) The arrow drawn through the scatterplot can be thought of as a component that captures the aspects of the indices that are highly correlated with each other.
We show later how the two indices can be replaced by their principal component, thus
reducing the two variables to one. Figure 12.4 illustrates this in three dimensions.
The plot shows three medical indices along with two principal components; it adds
an index of pharmaceutical manufacturing costs to indices in Figure 12.3. The three
variables can be reduced to the two top components.
287
4.7
4.8
4.9
5.0
5.1
5.2
5.3
PCA-related methods include singular value decomposition (SVD), another dimensionality reduction method that is frequently included as a multivariate technique in
statistical and data mining packages. SVD can be thought of as allowing decomposition of a dataset along the column dimension as with PCA, as well as the row
dimension as with cluster analysis. Independent component analysis (ICA) extends
PCA by allowing for nonlinear correlations.
Principal component analysis and factor analysis (FA) are related but different
methods. PCA decomposes total variance in a set of variables, whereas FA measures
shared variance or correlation. In PCA, a principal component is a linear combination
of the manifest variables. In FA, a manifest variable is a linear combination of the
latent variables. PCA is often viewed as purely exploratory, whereas FA can be
both exploratory and confirmatory. In FA, rotations are often needed to help create
interpretable factors. In PCA, no interpretations are done on the principal components,
so such rotations are not necessary for PCA.
Pharmaceutical
5.4
5.2
5.0
4.8
4.6
4.4
4.2
4.6 4.74.74.7
General Medical
Fig. 12.4. Plot of three medical indices and two principal components.
288
Unsupervised Learning
1.000
0.980
0.988
0.994
0.990
0.972
0.952
1.000
0.986
0.968
0.993
0.988
0.958
1.000
0.984
0.990
0.980
0.977
1.000
0.985
0.973
0.962
1.000
0.993
0.963
1.000
0.966
1.000
Shared/common
variance
Table 12.2 presents a correlation matrix for several (mostly health-related) price
indices and workers compensation (WC) severity. All variables seem to be relatively
highly correlated with each other, though the severity variable, which is more volatile
from year to year, tends to show a lower correlation with the other variables.
Lets do a simple example of applying principal components to the two variables.
PCA uses either the covariance matrix or correlation matrix. Table 12.3 shows the
covariance matrix for the general medical and physicians indices.
The covariance matrix can be decomposed into eigenvalues and eigenvectors:
= C T C,
= eigenvalues, C = eigenvectors5
(12.1)
289
GenMedical
Physicians
GenMedical
Physicians
0.031
0.018
0.011
MedIndices2<-data.frame(Indices$LnGeneralMed,Indices$LnPhysicians)
Simple.Princomp<-princomp(MedIndices2,scores=TRUE)
The graph of eigenvalues, produced using princomps plot function, shows how
important each component is. From this graph we can tell that the first component
vastly dominates and that component 2 can be discarded without significant loss of the
pattern captured by the data. Table 12.4 shows the values of the principal component,
produced as a linear sum of the original two variables.
Table 12.4. Original Cost Indices and their Principal Component.
Component1 = 0.475 Gen Med Index + 0.88 Physicians Index
Year
IndicesGeneral Medical
IndicesPhysicians
Component 1
1992
1993
1994
1995
1996
4.6
4.7
4.7
4.7
4.7
4.6
4.6
4.7
4.7
4.7
0.347
0.259
0.211
0.211
0.211
290
Unsupervised Learning
0 03
0 02
0 00
0 01
Var ances
0 04
Simple.Princomp
Comp.1
Comp.2
0.20
0.10
0.00
Variances
0.30
Indices.Princomp
Comp.1
Comp.3
Comp.5
Comp.7
291
From the loadings we see that all variables load negatively on the first component,
and the workers compensation severity and the health insurance index have the
highest (absolute value) loadings.
Loadings:
LnGeneralMed
LnPhysicians
LnPharma
LnHealthInsurance
LnCPI
LnCompensation
LnSeverity
Comp.6 Comp.7
-0.617 0.185
-0.165 0.125
-0.106
0.105 -0.918
0.750 0.308
To illustrate factor analysis we apply factor analysis to the cost indices using the
factanal function and ask for two factors.
>Indices.Factanal<-factanal(Indices2,factors=2,scores="regression")
It appears that the indices load almost equally on the first factor. Also, unlike
with principal components analysis, the first factor is positively correlated with the
indices. These loadings also assist our interpretation of the factors. For instance,
based on the loadings we can think of the first factor as a general inflation factor
and the second factor as specific to certain components of medical inflation. Only
292
Unsupervised Learning
physicians and health insurance load on the second factor, and only physicians load
on the third. The communality is the proportion of variance explained by the factors.
A high communality (> 0.5) indicates that the number of extracted factors adequately
explain the variances in the variables. A low communality (< 0.5) indicates that more
factors may be needed or that this dataset is not appropriate for factor analysis.
The communalities produced by factanal (shown next) indicate that the first
factor accounts for more than 95% of the variance.
Factor1 Factor2 Factor3
6.727
0.090
0.036
0.961
0.013
0.005
0.961
0.974
0.979
SS loadings
Proportion Var
Cumulative Var
The factors derived from factor analysis are not guaranteed to be interpretable. A
procedure called factor rotation is used to create new factors and to improve their
interpretability. Commonly used methods include orthogonal and oblique methods.
Orthogonal rotations are where the axes are kept at a 90-degree angle and include
methods referred to as Varimax, Equimax, and Quartimax. Oblique rotations are where
X and Y axes are not 90 degrees apart and include the method Promax. Although
factor rotation is not covered in this chapter, R offers the varimax and promax
functions for rotating the factor loading outputs of factanal.6
A factor component score can be assigned to each record. The score is in the form
of a standardized value assigned to individual observations. This is useful if follow-up
studies are performed. The score may itself become an input to a predictive model.
Weaknesses of factor analysis include the following:
r The following aspects of FA are subjective: rotation method, number of factors, extraction method, and interpretation of factors. As a result, a wide variety of outcomes could
result from performing FA.
r It only measures linear relations.
r Outliers may adversely affect the results.
Its strengths include that factor analysis is a well-known and widely used method,
especially in the social science fields, and it is Nonparametric.
Factor analysis (and principal components as implemented in the R princomp procedure) uses metric/numeric variables. Although categorical variables could
be represented by dummy variables, it is good practice to apply FA and PCA
to numeric/metric values. Except for the maximum likelihood estimation rotation
6
293
Scree Plot
Eigenvalue
0
1
3
4
5
Component Number
method, normality is generally not required. However, the results are enhanced if
normality is kept.
Although many software packages (e.g., R, SAS, etc.) offer the option of executing
PCA within factor analysis functions (e.g., Proc Factor), to avoid confusion it is
recommended that PCA-specific functions be used (e.g., Proc Princomp). The two
methods are so closely related that some software packages (i.e., SPSS7 ) offer only
one of the options. Thus, unless one has a specific purpose, such as identifying latent
factors or concepts underlying the variables, it may not matter much which approach
is used for variable reduction.
294
Unsupervised Learning
295
implementation. Most syntax for S-PLUS will work in R. The functions used in this
chapter match those described in Struyf et al. (1997) and have been run in R. Additional help with the R cluster functions can be obtained from Rs help capability. The
user must access the help function from the R command line. For instance to get help
with the function daisy, type:
?daisy
Available Packages
Currently, the CRAN package repository features 4677 av:
CRAN
Mirrors
Whats new?
Task Views
Search
About R
R Homepage
The R Journal
Software
R Sources
R Binaries
Packages
Other
Documentation
Unsupervised Learning
BIFrequency
8
12
296
7000
8000
9000
11000
BISeverity
13000
297
r Members of different groups tend to be dissimilar. That is, their values on the analysis
variables tend to be different.
Thus, to characterize the different groups, measures of similarity and dissimilarity
are used. Similarity statistics measure how alike two records are. The Pearson correlation coefficient is a kind of measure of similarity (between two variables, rather
than two records). Dissimilarity statistics measure how far apart two records are.
In general, there is a complementarity between measures of similarity and dissimilarity. The two most common measures of dissimilarity are Euclidean distance and
Manhattan distance.9
Euclidean Distance: The statistic is based on the record-by-record squared difference between the value of each of the variables for a record and the values for
the record it is being compared to. The distance d measured on p numeric variables
between two records i and j is
;
< p
<
2
(xik xj k )2 .
(12.2)
dij = =
i=1
Manhattan Distance: The statistic is based on the record-by-record absolute difference between the values for a record and for the record it is being compared to.
The distance d measured on p numeric variables between two records i and j is
p
2
2
2xik xj k 2.
dij =
(12.3)
k=1
In R, the daisy procedure in the cluster library can be used to compute dissimilarity matrices. If no method is selected for computing dissimilarities, the Euclidean
method is used as the default.
The R code to compute a dissimilarity matrix for our automobile data (contained
in dataframe ClusterDat1) is:
>distance<-daisy(ClusterDat1)
Table 12.5 displays the Euclidian dissimilarity measure for selected records from
the simulated auto data. Table 12.1 shows dissimilarity measures for the first two
records of our data. The dissimilarity output is typically an n n matrix, where n is
the number of records in the data.
See Aldenderfer and Blashfield (1984) for a fuller development of the formulas for dissimilarity.
298
Unsupervised Learning
1
2
3
4
5
1,400
2,671
4,488
410
1,400
1,271
3,088
1,810
2,671
1,271
1,817
3,081
4,488
3,088
1,817
4,898
410
1,810
3,081
4,898
in the table, the number of variables that the two claimants are zero on is denoted b, and
the number of variables that the two have a 1 on is denoted c. The off-diagonal counts
are a and d. The sample matching method computes the proportion of agreements:
Sample matching:
dij =
b+c
.
a+b+c+d
Jaccards coefficient:
dij = 1 Sin ,
Sij =
a
S = similarity.
a+b+c
2(b + c)
a + d + 2(b + c)
The Rogers and Tainimoto statistic gives more weight to the count of disagreements
between records. The Jaccard coefficient does not count the negatives (zero values for
binary variables).
Our example of categorical dissimilarity uses the simulated fraud data (see Table
12.6). For this illustration the sample contains only two binary categorical variables:
299
1
2
3
4
5
0
0
0.5
1
0.5
0
0
0.5
1
0.5
0.5
0.5
0
0.5
1
1
1
0.5
0
0.5
0.5
0.5
1
0.5
0
whether the claimant was represented and whether the injury was a soft tissue injury.
To obtain this dissimilarity matrix in R, use:
metric= "gower" ) )
where ClustrerDat2 is the simple dataset with two variables from the
fraud data.
Since our binary variables are coded as zero for a no on a given category and one
for a yes (i.e., if represented by an attorney, the litigation variable is zero; otherwise
it is one), Euclidean distances can be computed. For comparison, we display the
dissimilarities computed using Euclidean distances (see Table 12.7). Note that Zhang
et al. (2006) point out a number of problems with treating categorical variables as
scalar and computing Euclidean or Manhattan distances. For instance the measure
is not independent of the scale chosen (i.e., coding nos as 1 and yess as 3, rather
than using zeros and ones). For categorical variables with many categories, many
dummy binary variables would be required as separate variables for the dissimilarity
matrix.
Table 12.7. Euclidean Dissimilarity Matrix for Fraud Data
(two-variable sample of binary variables)
1
2
3
4
5
0
0
1
1.414214
1
0
0
1
1.414214
1
1
1
0
1
1.414214
1.414214
1.414214
1
0
1
5
1
1
1.414214
1
0
300
Unsupervised Learning
1
2
3
4
5
0
0.4
0.511
0
0.067
0.4
0
0.911
0.4
0.467
0.511
0.911
0
0.511
0.444
0
0.4
0.511
0
0.067
5
0.067
0.467
0.444
0.067
0
metric= "gower" ))
12.4.5 Standardization
Scale affects the results of clustering. Therefore the analyst often prefers to standardize
variables before performing clustering (i.e., for each variable x, subtract its mean
and divide by its standard deviation or absolute deviation). The dissimilarity function
daisy allows for scaling. The daisy function uses the absolute deviation rather than
the standard deviation when scaling because the absolute deviation is more robust to
outliers. To scale the auto BI data, we can standardize the frequency data with the code:
>distance2<-as.matrix(daisy(ClusterDat1,stand=TRUE))
301
1
2
3
4
5
0
2.987
2.141
2.176
2.466
2.987
0
2.383
2.393
3.51
2.141
2.383
0
0.123
1.511
2.176
2.393
0.123
0
1.606
5
2.466
3.51
1.511
1.606
0
302
Unsupervised Learning
Next we see that the output displays the medoids (or by variable means) for each
cluster, thereby identifying a low-frequency/low-severity and high-frequency/highseverity cluster. The low-frequency cluster contains 63% of the records.
> BICluster1
Call: clara(x = ClusterDat1, k = 2, metric = "euclidean")
Medoids:
BIFrequency BISeverity
[1,]
11.39769
8202.802
[2,]
13.28089 10749.593
Objective function: 577.0351
Clustering vector:
int [1:100] 1 1 2 1 1 1 1 1 1 1 1 2 1 2 2 2
1 1 ...
Cluster sizes:
63 37
Next we standardize the data used for clustering. The two clusters contain 64% and
37% of the data, about the same as for the unstandardized data. However, the procedure
now separates these data into a low-frequency/high-severity and high-frequency lowseverity group, similar to the actual data.
BICluster2<-clara(ClusterDat1,2,metric="euclidean",stand=TRUE)
Call: clara(x = ClusterDat1, k = 2, metric = "euclidean",
stand = TRUE)
Medoids:
BIFrequency BISeverity
[1,]
13.040838
8043.759
[2,]
7.570882 11019.007
Objective function: 0.625121
Clustering vector:
int [1:100] 1 1 1 1 1 1 1 1 1 1 1 2 1 2 2 2
1 1 ...
Cluster sizes:
64 36
It is customary to plot clustered data using principal components as the X and Y axes
of the graph (see Figure 12.9). Principal components are discussed in Section 12.2.
0.5
0.5
303
1.5
Component 2
1
Component 1
Fig. 12.9. Cluster plot of auto data. These two components explain 100% of the point
variability.
In the case of our simple example using simulated data, the two components correspond closely to the two variables that we are clustering on.
>plot(BICluster2)
The silhouette plot provides information on the goodness of fit of the clusters to the
data (see Figure 12.10). The silhouette statistic measures the similarity of members of
a cluster relative to other clusters (i.e., the next closest cluster in the case of more than
two clusters). The closer the statistic is to 1, the more homogeneous the partitions are.
The silhouette plot is a bar plot of the silhouettes of the records for each cluster.
To illustrate clustering on mixed numeric and categorical variables, we apply the
clara function to the simulated auto questionable claims data. To utilize categorical
data we first create a dissimilarity matrix (using the daisy function) using the gower
n = 44
2 clusters Cj
j: nj aveicj Si
1: 31 0.68
2: 13 0.66
0.0
0.2
0.4
0.6
0.8
Silhouette width Si
Average silhouette width: 0.67
1.0
304
Unsupervised Learning
dissimilarity measure. The matrix then becomes the input to the pam function (note
the clara function does not accept dissimilarity matrices). Figure 12.11 shows the
silhouette plot for this data.10 Because we cannot create a cluster plot for categorical data, we display the confusion matrix or the cross-tabulation of actual versus
predicted categories.
As can be seen from the cross-tabulation in Table 12.10, the k-means cluster
correctly predicts 92% of the smaller (questionable claims) cluster and 96% of the
larger (legitimate claims) cluster.
We ran the k-means clustering on the actual auto BI CAARP data and note that
the results are much noisier, as can be seen from the cluster plot in Figure 12.12.
We clustered on the following variables: BI frequency, PD frequency, log of the BI
severity, and log of the PD severity. Because of the skewness of the severity data, we
used the log transformation. We ran the procedure with k = 2, 3, and 4. The average
silhouette was highest for k = 2, suggesting that it is the best partition. The k = 2
average silhouette was 0.3 as opposed to 0.23 or less for k = 3 and 4. The results are
shown in the box. Note that the PD severity appears to have no meaningful effect on
the clustering. The BI and PD frequencies separate into a high-frequency and lowfrequency group. The low-frequency group has a higher BI severity (on a log scale).
Call: clara(x = AutoBIVars, k = 2, metric = "manhattan",
stand = TRUE, keep.data = TRUE)
Medoids:
AutoBI.FREQBI AutoBI.FREQPD AutoBI.LnBiSev AutoBI.LnPdSev
[1,]
11.953501
38.04287
9.080598
7.748904
[2,]
8.347081
25.45903
9.328678
7.763548
Objective function: 3.243358
Clustering vector:
int [1:1459] 1 2 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 ...
Cluster sizes:
780 679
Best sample:
Note that for this plot the sample size was reduced to 100.
305
659
97%
26
8%
685
2
Grand Total
Grand Total
22
3%
293
92%
315
681
319
1000
2 clusters Cj
n = 100
j: nj aveicj Si
1: 70 0.82
2: 30 0.65
0.0
0.2
0.4
0.6
0.8
1.0
Silhouette width Si
Average silhouette width: 0.77
5
0
5
Component 2
10
Component 1
Fig. 12.12. Cluster plot of actual auto data. These two components explain 75.68%
of the point variability.
Unsupervised Learning
Component 2
1
1
306
Component 1
Fig. 12.13. Fuzzy clustering plot of fraud data. These two components explain
40.44% of the point variability.
two or more groups, rather than assigning it to only one group. The procedure that
does this is called fuzzy clustering. The R function for fuzzy clustering is fanny.
Kaufman and Rousseeuw (1990) and Maechler (2012) state that fanny minimizes
a weighted dissimilarity function where the weights are the fuzzy membership
function denoting the proportion of its membership in a given cluster. For any record,
the membership functions sum to 1. At the limit, the membership functions equal 1
for only one cluster and 0 for all others, yielding the traditional hard clusters:
$ k n n
%
i=1 j =1 uriv urj v d(i, j )
n
min
,
(12.5)
r
u
j
=1
j
v
v=1
over 0 < u < 1 and u = 1, r 1. Here, n is the number of records, k is the number
of clusters, and d(i, j ) is the dissimilarity between i and j .
An iterative procedure is used to solve for the clusters and for the membership
functions of the clusters. The user must specify the number of clusters or a range of
clusters.
To illustrate, we apply fanny to our automobile fraud data. This dataset contains
a number of binary variables: attorney represented, soft tissue injury, type of medical
provider, filing of a police report, use of an emergency room, and whether the claimant
had a prior claim. Because fanny does not use the gower dissimilarity measure,
we treated these variables as numeric. The data also show several numeric variables
(number of providers, number of treatments and lag from accident date to report date).
All variables were used in the clustering. We show the code next, and Figure 12.13
diplays the cluster plot from fanny. Table 12.11 shows that fanny output assigns a
membership value to each record for each cluster.
>QuesClaimsDataCluster2<-fanny(ClusterDatFraud,2,metric="euclidean",
keep.data=TRUE,memb.exp=1.5)
>plot(QuesClaimsDataCluster2)
1
3982
1423
9
5
1
8
70
1213
7
2
6
1351
8
4
3
1157
3
1204
986
12
3
1
2
1436
8
38
7
1291
13
5
36
3
7
1454
3
1360
381
1235
4
129
1247
3
578
14
0
1129
3
5
127
1270
1
87
1374
38
1222
715
363
1444
057
32
20
1317
1273
1435
2
1256
13
3
3
9
0
1014
78
9
35
5
1030
1
6
1100
556
9
1
22
5
1357
691
040
938
75
864
839
7
132
35
13
1384
80
1
1368
0
11
113
7
1193
7
9
3
973
1
251
878
681
5
7
11
660
1
77
7
9
42
1007
2
8
1372
1449
7
143
1225
1
5
13
9
418
786
610
0
1405
3
10
1306
735
833
8
1236
14
9
28
1
8
2
1240
1322
1371
86
39
1441
13
1459
37
2
77
14
4
1409
1277
1111
1432
1036
4
0
7
1
42
1301
9
3322
1150
33
372
1
58
6
8
4
866
015
1206
1251
1348
116
37
139
1218
3
9
1418
145
0
7
6
1286
24
14
1425
0
1335
4
5
997
1115
7
383
9
1299
5
3
1365
13
2
1827
97
14
09
1445
1065
985
19
3
8
1344
1330
13
2
453
262
1430
1008
1901
1452
8084
1385
2
1457
132
5
11042
4
9
6
28
7
33
3
9
85
4
5
136
2
5
65
86
3
8
162
99
51
1
7
75
81
7
187
126
6
3
1174
8
165
280
8
66
38
9
4
2
7
6
157
47
0
15
7
53
0
37
13
93
1345
665
4
6
1
5
82
45
0
177
63
141
50
146
9
129
1
53
761
2
77
8
196
3
0
84
5
1
01
26
27
35
55
60
2
7
158
29
1
57
79
8
27
26
198
78
235
3
147
2
5
4
336
70
5
7
7
203
368
128
210
4
6
221
166
0
7
173
20
7
3
5
355
22
1
64
9
8
227
11
105
2
3
305
08
216
4
36
1
9
167
65
8
17
21
116
55
134
190
5
7
234
8
3
292
239
2
3
2
5
247
308
3
28
5
880
35
4
1
71
77
23
389
5
7
339
148
557
8
4
259
34
5
6
25
407
6
25
70
383
4
241
373
0
172
7
415
306
3
5
59
78
4
20
9
5
278
119
65
5
0
3
313
3
759
2
1
664
33
47
603
5
8
287
5
50
7
889
311
73
4
562
5
3
102
537
6
76
8
9
73
483
10
2
8
4
1
37
543
82
5
396
326
297
318
79
2
6
434
8
322
322
2
913
309
263
223
7
5
1
87
8
3
7
283
143
400
4
88
275
410
67
5
7
28
163
8
7
9
254
273
325
458
37
432
155
7
26
97
532
120
272
605
479
83
347
475
384
15
3
6
7
77
386
78
8
5
0
8
482
23
7
582
1051
3
57
4
9
934
72
8
76
848
7
3
5
5
4
963
1
113
3
9
742
01
55
7
1055
758
3
13
4
620
175
738
824
27
808
25
358
682
2
9
7
380
1421
1020
450
320
0
97
078
1162
5
9
7
872
1
94
5
7
53
7
53
7
74
351
4
9
73
1220
3
1252
813
1077
045
7
345
732
7
1329
264
520
769
80
578
1
59
641
783
5
70
357
17
5
3
592
663
777
6
4
7
782
75
335
7
564
8
1
8
2
5
801
882
18
49
615
9
2
397
246
357
84
760
653
429
8
33
9
505
5
17
6
236
528
5
7
828
2
5
87
461
0
775
430
1
26
5
898
7
5
832
8
75
2
796
8
3
57
631
575
92
9
1123
5
7
1
7
63
838
128
8
948
27
73
9
945
33
9
072
1044
938
75
1028
5
7
5
58
1200
77
1772
867
8
1
8
85
1093
3
3
27
7
0
768
2
776
23
703
7
89
11
206
3
1
03
331
7
9
636
224
228
255
2
3
29
5
328
792
6
4
382
26
449
2
9
8
770
471
6
3
1
7
580
88
766
1141
788
802
646
139
3
1095
8
1
9
555
5
382
8
984
0
7
2
1081
1052
1
30
07
1120
3
9
481
79
9
875
204
62
48
787
5
7
6
3
1179
37
8
1
87
1
539
54
4
63
88
2
71
8
1016
7
7
517
1046
1048
98
77
257
1011
67
1153
789
850
1285
492
8
1
5
5
9
07
496
8
7
1
0
9
1158
12
2
77
1089
1245
4
1053
7
5
75
83
1231
67
75
3
31
1448
7
1
83
9
5
12
1041
68
80
8
8
1079
3
0
08
145
0
8
6
987
446
5
1
58
89
1377
3
5
0
1274
75
8
2
536
1
7
1035
3
755
72
98
133
815
876
1433
6
18
1428
127
3
852
8
6
9
1
58
437
52
854
1210
7
5
1307
7
567
990
595
1
1403
2
436
58
38
9
8
774
5
3
1054
83
1353
717
1208
36
9
8
547
5
58
84
798
30
2
67
1312
5
1124
988
1
440
6
7
73
881
208
5
15
8
685
553
1
69
2
392
62
13
6
99
1382
552
1031
2
1038
7
3
564
3
312
8
85
8
1033
904
53
1227
19
713
76
573
7
1
4
1
6
843
0
870
1387
97
1117
7
5
238
501
9
1132
02
811
3
3
7
1
32
1416
13
977
11
67
597
678
49
2
7
1378
9
326
1276
1
7
3
39
967
05
58
54
996
7
8
7
48
13
1101
12
121
645
1107
468
1
42
17
392
1411
902
573
33
1440
8
857
545
609
77
29
10
1396
7
199
1837
9
1
8
5
1182
08
13
7
4
1086
10
9
7
6
578
37
8
0
96
529
3737
00
4
781
7
7
791
1238
8
9
29
1
9
588
735
1269
941
5
1379
2
78
8
7
9
1058
1
3
747
7
778
1022
72
5
70
955
3
73
370
1221
3
7
1
3
756
859
869
795
1230
845
53
495
8
9
994
5
13
5
1323
725
7
1908
551
6
885
4
1313
696
767
1249
0
8
1
8
577
964
5
3
12
1133
55
856
3
13
632
182
757
7
1
25
02
710
1183
7
1148
1209
7
1354
1338
2
119
1349
797
914
642
936
25
1049
1153
586
1308
20
Height
50
Record
1
2
3
4
5
307
0.812
0.094
0.232
0.885
0.585
0.188
0.906
0.768
0.115
0.415
1
2
2
1
1
Unlike k-means clustering, hierarchical clustering does not create a specific number
of clusters. Instead it sequentially partitions the data. At each step additional clusters
are created under divisive clustering (see Figure 12.14), or clusters are merged under
agglomerative clustering. The results are presented in the form of a tree. Like k-means
clustering, hierarchical clustering uses a dissimilarity matrix to split data into groups.
Under divisive clustering, the procedure starts with a cluster of size one, containing
all the data. It then splits the data into two groups. These two groups are then split into
more groups. Ultimately the data can be partitioned into as many clusters as there are
records (i.e., each record is its own cluster and no grouping is performed). To perform
divisive clustering we use the diana function:
>BICluster7<-diana(AutoBIVars,metric="manhattan",keep.data=TRUE,
stand=TRUE)
>plot(BICluster7)
AutoBIVars
Divisive Coefficient = 0.98
11
10
10
1042
2
3
73
581
4
8
0113
15
1206
1251
1385
1324
1161
901
21457
115
72 884
2
96
33
83
22
44
5
215
46
7
3
72
7
75
81
2
3
8
5
41
89
162
99
39
0
50
5
160
87
10
51
1
5
177
62
76
7
24
219
68
109
2
7
1
8
271
198
276
29
0
106
27
142
3
5
5
134
7
535
190
72
4
0
321
3
5
9
78
220
2
4
3
292
239
3
305
29
225
451
3
5
2
8
7
152
37
7
90
77
16
203
368
25
108
3
543
361
58
07
73
200
8
22
9
4
649
1282
6
1
2
2
355
199
150
98
183
7
5
33
53
122
154
128
19
605
55
226
532
974
65
1
1
2
27
98
127
5
249
373
1020
202
826
3
1
1329
645
89
1211
1
4
409
7
7
8
3
87
5
3
582
0
28
283
9
254
186
3
7
6
57
443
10
3
514
8
472
631
57
68
268
75
48
769
293
15
6
9
377
37
505
378
3
1379
652
72
510
295
3
10
7
537
836
189
491
965
455
924
43
2
796
73
8
7
353
84
359
53
338
565
76
8
8
751
4
286
32
3
6
404
4
79
578
469
75
529
323
0
1358
03
33
719
572
6
4
825
78
899
362
776
41
584
7
6
33
289
570
164
1
306
7
782
252
345
799
80
270
1012
304
57
5
6
63
777
534
4
0
143
32
168
2
7
8
27
4
257
407
349
6
7
5
7
28
29
5
3
2
53
2
557
822
201
65
247
30
413
35
6
485
7
5
574
889
5
56
313
525
33
3
8
476
6
82
2
654
5
62
10
130
50
290
9
2
7
80
318
434
7
4
66
263
490
73
341
55
83
8
54
996
7
507
6
7
8
528
120
358
682
49
447
181
291
3
5
5
32
692
5
7
72
1078
11124
62
50
8
4
971
2
957
872
6
7
2
2
926
5
358
720
75
209
3
711
07
956
772
8
7
966
988
1118
204
1623
53
448
7
5
5
6
37
1
87
6
7
388
3
488
3
0
601
708
86
1080
08
7
2
452
625
782
896
1
2
8
765
883
8
72
497
0
3
7
8
3
502
100
775
520
85
903
1575
95
4
545
7
9
78
1159
377
44
583
831
1231
754
553
0
7
8
1
68
696
8
4
37
0
1001
7
3
1003
8
504
68
1019
0
85
2
127
917
35
1307
31
1448
5
7
283
1
9
5
1396
2
8
4
1050
87
1037
1048
1189
37
8
9
1138
2
838
1145
1
8
3
85
6
8
912
10
5
1085
7
73
109
945
33
11123
072
0
2
558
779
897
259
9
5
1295
03
766
1
1
7
0
4
771
1923
740
88
8
5
80
900
80
788
674
11166
53
789
024
10
55
170
2
79
71
337
415
15
7
3
7
423
38
58
339
2
8
325
192
3
3
28
255
28
552
464
7
508
34
768
2
89
60
1031
2
6
71
522
7
6
6
555
911
8
110
8
1117
47
261
4
9
703
6
77
80
6
164
84
101
5
1
33
272
76
310
322
690
737
1
845
1107
300
3
1238
816
929
3
5
1
7
671
82
26
13
0
632
1253
757
7
3
8
7
1215
5
137
1192
326
127
1
93
1375
394
73
967
905
5
1083
588
0
1230
495
7
8
778
1022
743
1
6
834
5
2
78
817
58
12
931
614
790
55
7
1055
8
824
634
1
1
34
859
8
9
78
795
62
913
7
35
890
755
863
870
22
13
7
688
8
5
8
1367
039
8
108
11377
333
59
727
10
4
673
1
20
38
1421
1252
67
26
145
78
5
9
5
207
436
7
21
56
446
5
8
9
0
1358
9
208
5
8
685
55
4
5
1292
499
13
6
1203
909
1
22
0
1169
1111
1240
5
3
76
122
611
1382
455
26
730
7
1773
3
1155
8
7
330
360
1
8
3
92
1035
75
8
2
536
1
70
35
500
6
1105
5
7
8
930
04
8
0
1274
887
3
1
353
380
9
1774
523
3
5
205
1082
1
8
45
1009
2
33
1386
05
10
735
1263
83
1243
38
17
3
972
1181
14
6
1
1
1371
7
7
17
1102
1320
1
9
395
285
1096
1
58
677
1043
18
7
0
5
1232
12
2
55
712
1
8
4
5
9
907
8
3
84
744
6
1356
1423
995
143
4
1185
177
745
1
3
2
88
033
1177
5
0
27
868
1285
272
46
7
1046
62
718
8
0
1
7
33
755
88
133
815
87
144
598
78
336
3
7
1454
1333
3
13
5
999
7
38
1
1089
1228
1322
52
1360
8
8
40
1202
2
1436
1233
1372
7
0
1318
1409
38
297
14
0
1164
1
35
2
129
493
9
2
124
551
862
767
35
1040
3
187
0
1133
577
96
8
0
695
87
13
3
1374
853
1099
1160
6
1323
453
1196
1
8
7
1338
794
4
86
1354
57
1399
3
1337
7
1328
1212
13
9
6
1351
1
278
1389
265
57
1332
1361
1201
90
1277
1363
1359
1
35
3
9
481
0
9
1
30
875
725
76
8
4
950
1
908
468
949
0
7
2
1270
1392
1
11
1437
11
60
6705
1279
13
66997
3
857
1218
0
1335
144
1450
73
33
14
1406
5
9
95
1237
0
7
1403
99
1113
877
1
29
960
9
1281
383
1427
9
1395
82
1
9
26
7
1456
14
8
23
1365
1459
2
1397
13
31
9
742
736
95
738
238
501
373
8
1106
962
1
277
1036
251
8
68
1
73
74
892
98
1
1407
786
80
1148
1447
1368
878
9
2
1439
1
73
52
970
1163
04
7
225
1451
13
9727
7
1
5
1067
37
13
8
6
938
1135
1005
10
5678
3
943
1120
1
55
1309
1
1384
58
1357
691
1280
8
1183
1021
32
1376
1306
1
25
7
5
1393
1114
1030
126
1300
1310
1
00
3
1222
1167
1261
8
255
8
1369
1420
14
4460
24
1
973
11301
54
393
70
690
881
11
977
1047
1142
140
2
1065
1
9
134
797
914
0
14
5
6
115
1187
138
1262
937
418
1453
3
1342
1434
1008
1
1430
1452
324
25
1049
117
6
8
82
2
157
74
6
6
6
38
923
1174
5
1
2
1
28
7
3
37
10
15
17
33
14
93
345
88
280
1077
813
586
1308
1348
617
9165
1150
307
732
34
665
455
Height
20
308
Unsupervised Learning
AutoBIVars
Agglomerative Coefficient = 0.96
the first step, the two records with the smallest dissimilarities are combined. In the
next step two more records are combined, or a record is joined to the cluster grouping
created in step 1. In R the function agnes is used to perform agglomerative clustering.
The statistic agglomerative coefficient is supplied with the agnes output, and the
divisive coefficient11 is supplied with the Diana output. According to Kaufman
and Rousseeuw (1990), the closer to zero these coefficients are, the less is the natural
grouping structure in the data; the closer to one they are, the higher the degree of
structure. Higher coefficients are observed for larger datasets, so the coefficients
cannot be used to compare data of different sizes.
The graphical output from a hierarchical clustering can be used to assess how many
clusters to keep. Note that when your data contain many records, as in this example,
the details are blurred at deeper levels of the tree, since ultimately each record is a point
on the plot. The length of each branch of the tree (sometimes called an icycle graph)
is proportional to the distance between the groups. The smaller the length, the less
dissimilar the clusters are. Note that under both divisive and agglomerative clustering
the first (next to last last) partition has greater depth than subsequent clustering. Under
divisive clustering, the second through fourth partitions (i.e., six groups) also seem to
be meaningful. Under agglomerative clustering the two-way split of the data from the
next to last clustering seems much more significant than further partitions of the data
(see Figure 12.15).
See pages 210218 of Kaufman and Rousseeuw (1990) for a more thorough discussion of the construction of the
coefficients.
12.5 Exercises
309
a model of potential fraud and abuse claims, it typically does not have an indicator
in its data as to whether a particular claim was considered legitimate or possibly
suspicious. That is, there is no dependent variable related to suspicion of a staged
accident, exaggeration of damages, or other abuses. Instead, the company can use
the data it does have to group the records into those that are similar to each other
on a number of claimant, accident, provider, policyholder, geographic, and other
variables. After records have been clustered by one of the procedures illustrated in
this text, one can then examine the characteristic of the clusters. In the questionable
claims example, it turns out that one of the clusters has a much higher proportion
of records with attorney representation, soft tissue injury, use of a chiropractor or
physical therapist, and history of a prior claim, whereas the other cluster has a much
higher percentage of claimants who used the emergency room and filed a police report.
As a result, we might label the first cluster as the suspicious cluster and the second
cluster as probably legitimate.
Our examples with the California automobile insurance data also illustrate another
way in which unsupervised learning can be used: to develop geographic groupings
that can be based on important features of the data other than geographical proximity.
Although our simple examples did not show it, we could have used geographic
and demographic data in addition to loss data in the construction of geographic
groupings. Christopherson and Werland (1996) and Weibel and Walsh (2008) provide
more detailed illustrations. Weibel uses clustering along with mixed models in the
development of territories. In our illustrations we ignored credibility considerations,
and Weibel shows how to take them into account.
12.5 Exercises
Exercise 12.1. Apply both principal components and factor analysis to the simulated
auto data in Example 12.3. Keep the first component and the first factor.
Exercise 12.2. Apply factor analysis to the economic and insurance indices data in
Example 12.4. How many factors would you keep? How would you use these factors
in an insurance pricing analysis?
Exercise 12.3. Using the primary paid loss and total allocated loss
adjustment expense variables from the Texas data in Example 12.2, compute
dissimilarity matrices using
(a) the Euclidean measure,
(b) the Manhattan distance,
(c) standardized variables with the Manhattan distance.
310
Unsupervised Learning
Exercise 12.4. From the Texas data in Example 12.2, using the Euclidean dissimilarity measure and the Manhattan dissimilarity measure, standardized, choose
k = 2, 3, 4, 5. Run cluster and silhouette plots. Which k would you select. Why?
Exercise 12.5. Using the injury and cause variables from the Texas data in Example
12.2, compute a dissimilarity matrix and perform k-means clustering using the gower
measure. Make sure to convert the injury and cause variables to factors in R, using the
as.factor function before clustering. Choose k = 2, 3, 4, 5. Which k would you
choose?
Exercise 12.6. Use the R mona function to cluster the categorical variables in the
simulated questionable claims data in Example 12.3. Output the clusters to a text file
and compare actual to predicted clusters.
Exercise 12.7. Apply hierarchical clustering to the simulated questionable claims
data in Example 12.3. Where do you think the best split is?
Exercise 12.8. Apply hierarchical clustering to the CAARP data in Example 12.1
using BI frequency PD frequency, BI severity, and PD severity. Where do you think
the best split is?
Exercise 12.9. Use loss variables along with geographic variables (latitude, longitude,
and elevation) to cluster the CAARP data in Example 12.1.
Exercise 12.10. Use the scores from Exercise 12.1 to predict cluster membership.
Exercise 12.11. Apply fanny to the questionable claims data in Example 12.3. Test
2, 3, and 4 clusters. Output the membership functions and examine them.
References
Aldenderfer, M. S. and R. K. Blashfield (1984). Cluster Analysis. Sage Publications.
Brockett, P. L., R. A. Derrig, L. L. Golden, A. Levine, and M. Alpert (2002). Fraud
classification using principal component analysis of RIDITs. Journal of Risk and
Insurance 69(3), 341371.
Brzezinski, J. R. (1981). Patterns in persistency. Transactions 33, 203.
Christopherson, S. and D. Werland (1996). Using a geographic information system to
identify territory boundaries. Casualty Actuarial Society Forum.
Coaley, K. (2010). An Introduction to Psychological Assessment and Psychometrics. Sage
Publications.
Crawley, M. (2007). The R book.
Erin Research (2003). Comparison analysis implications report of employer and member
research. Prepared for the Society of Actuaries.
Everitt, B., S. Landan, M. Leese, and D. Stahl (2011). Cluster Analysis. Wiley, New York.
References
311
Part III
Bayesian and Mixed Modeling
13
Bayesian Computational Methods
Brian Hartman
316
the estimates of the parameters are correct. Bayesian methods provide an entire distribution for each parameter and probabilities for the various model choices, allowing
explicit modeling of both the model and parameter uncertainty. Those uncertainties
can have a large impact on the results of the model, especially on the risk measures.
For a more thorough discussion of the problem and a few examples of the large effect,
see Hartman and Heaton (2011).
Actuaries, like other professionals, use their expert judgment to both assess the
validity of model results and adjust those results based on personal knowledge not
explicitly incorporated in the model. These adjustments are often performed on an
ad hoc basis. Under the Bayesian paradigm, the prior distribution provides a structure
to incorporate those expert beliefs efficiently and consistently (see Section 13.5 for
further discussion).
Finally, Bayes framework is readily applicable to and could be beneficial for all
the other methods in this book. See Section 13.7 for a list of references describing
how to apply the methods in this book in a Bayesian setting.
13.2 Personal Automobile Claims Modeling
The methods presented throughout this chapter are applied to a dataset originally
presented in De Jong and Heller (2008). This dataset contains one-year vehicle insurance policies taken out in 2004 or 2005. There are 67,856 total policies in the set,
but this chapter focuses on the 4,624 policies (6.8%) that had at least one claim, and
therefore a positive total claim amount. The dataset contains many covariates that
could easily augment the models in this chapter using the methods in the next chapter.
These policies often had exposures smaller than one, so the annual claim amount is
the claim cost for the policy divided by the exposure. Figure 13.1 presents the kernel
density estimates (KDEs) of the annual claim amount both in its original scale and
after applying a logarithmic transformation. Notice that the original data are extremely
right-skewed. Even after the log transformation some skewness remains. Additionally,
there may be more than one mode in the transformed data. Basic summary statistics
for both the original and transformed data are presented in Table 13.1.
13.3 Basics of Bayesian Statistics
In classical, or frequentist, statistics the parameters are assumed to be unknown
but fixed. Hypothesis tests and confidence intervals are generated by answering the
following question, Assuming that the parameter value is x, what is the probability of
obtaining sample statistics as extreme as the current sample? Essentially, the result is
compared to the distribution of all possible samples that could have been taken under
certain assumptions.
317
Table 13.1. Summary Statistics for the Scaled Total Claims Data
Data
Min.
Q1
Median
Mean
Q3
Max.
St. Dev.
Original
Transformed
200
5.299
575
6.354
1,415
7.255
11,080
7.456
4,071
8.312
5.24 M
15.47
119 K
1.417
( |y) = >
(13.1)
The updated distribution, ( |y), is known as the posterior distribution. Often, especially in high-dimensional cases, the denominator in Bayes rule has no closed-form
expression. It is important to note that, with respect to , the denominator is simply a
KDE of log(Annual Claim Amount)
0.15
0.00
0e+00
0.05
1e04
0.10
Density
Density
2e04
0.20
3e04
0.25
4e04
0e+00
2e+06
4e+06
Annual Claim Amount
6
8
10 12 14 16
log(Annual Claim Amount)
318
constant. Therefore,
( |y) = >
f (y| ) ()
f (y| )().
f (y|)( )d
(13.2)
All that is missing is a normalization constant (to make the density integrate to one).
If f (y| )() is proportional to a known distribution, then the normalization will
come from that known distribution. A simple example is presented here to illustrate
the basic process.
Example 13.1. Assume that claims follow an exponential distribution with mean 1/:
f (y|) =
n
"
i=1
n
4
(13.3)
yi .
i=1
(13.4)
n
4
yi exp {}
(13.5)
i=1
n exp
$ n
%4
yi +
(13.6)
i=1
n
n
(n+1)1 exp
n+1
1/
1/ i=1 yi +
(n + 1)
i=1 yi +
1
1
|y n + 1, n
i=1 yi +
4
(13.7)
.
(13.8)
Therefore, only the sum and size of the sample are needed to update the prior distribution to the posterior distribution in this case.
When the prior and posterior distributions are of the same family, they are known
as conjugate priors. A list of conjugate priors and their accompanying parameters is
available on Wikipedia (2012).
319
30000
20000
0
10000
Density
40000
50000
0.00047
0.00048
0.00049
0.00050
0.00051
0.00052
0.00053
Theta
A simple BUGS script with the following model statement can apply the above
conjugate model to the total claims data:
model {
Notice that the posterior distribution is not included in the code. BUGS, or its
Windows-based version WinBUGS, will take the prior(s) and likelihood and design
a sampler to obtain draws from the posterior distribution automatically. Figure 13.2
shows a kernel density estimate of the draws from the posterior distribution generated
by the above code and the true posterior distribution calculated above. As expected,
the two curves are very similar.
R and WinBUGS code for all the analyses and figures in this chapter is available in
the online supplement. The package R2WinBUGS (Sturtz, Ligges, and Gelman 2005)
provides the ability to run WinBUGS from R and is used extensively in the code for
this chapter. For more information on Bayesian modeling using WinBUGS, see Lunn
320
estimated. The class of numerical methods that use Markov chains to generate draws
from a posterior distribution is collectively known as Markov chain Monte Carlo
(MCMC) methods. The two most common methods are the Gibbs sampler (Gelfand
and Smith 1990) and the Metropolis-Hastings sampler (Hastings 1970; Metropolis
et al. 1953).
13.4.1 Gibbs Sampler
When numerically approximating the normalization constant, often the integral is
of high dimension. Gelfand and Smith (1990) showed that, under mild regularity
conditions (Besag 1974), the joint estimates of the parameters can be obtained by
sequentially estimating each of the parameters, holding all others constant. The distribution of a parameter, holding all other parameters constant, is known as a full conditional distribution. If it is possible to sample from full conditional distribution, the
parameters can be drawn directly. Otherwise the parameters can be drawn indirectly
using adaptive rejection sampling (Gilks and Wild 1992) or the Metropolis-Hastings
sampler described in the next subsection.
Specifically, if the model contains k parameters (1 , 2 , . . . , k ), start with arbitrary
starting values (2(0) , 3(0) , . . . , k(0) ), and for t = 1, . . . , T perform the following steps:
Step 1: Draw 1(t) from 1 |2(t1) , 3(t1) , . . . , k(t1) , y
Step 2: Draw 2(t) from 2 |1(t) , 3(t1) , . . . , k(t1) , y
Step 3: Draw 3(t) from 3 |1(t) , 2(t) , 4(t1) , . . . , k(t1) , y
..
.
(t)
Step k: Draw k(t) from k |1(t) , 2(t) , . . . , k1
,y .
The sampler will take some time to move from the starting values to the higher
density areas of the posterior distribution and then will take more time to fully explore
that distribution. The draws are independent of the starting values when the first n1
iterations, known as the burn-in period, are discarded. The burn-in period allows the
sampler enough time to move from the starting values and start exploring the posterior
distribution. The next n2 draws are used to estimate the posterior distribution and to
calculate any quantities of interest. Considerations for deciding the sizes of n1 and n2
are discussed later in this section.
13.4.2 Metropolis-Hastings Sampler
Even after holding the other parameters constant, the full conditional distribution
may not be of a common form. As such, samples from the distribution cannot be
321
directly sampled. Luckily, the Gibbs sampler does not require that the full conditional
distribution is known, but only that draws from that distribution can be obtained. The
Metropolis-Hastings sampler (Hastings 1970; Metropolis et al. 1953) provides those
draws.
Assume a current value of the parameter (1(t1) ) and a function (proportional to
the full conditional distribution)
h(1 ) = (1 |2 , . . . , k )(y|1 , . . . , k ) (1 |y, 2 , . . . , k ).
(13.9)
Propose a new value, 1 , drawn from the candidate density q(1 |1(t1) ). Next, calculate
the acceptance ratio r using
h 1 q 1(t1) |1
r=
(13.10)
.
h 1(t1) q 1 | (t1)
With probability r, accept the proposed value (1(t) = 1 ), and with probability 1 r
reject the proposed value and use the current value (1(t) = 1(t1) ). If the proposal
distribution is symmetric, then q(1(t1) |1 ) = q(1 | (t1) ), and those terms cancel
out, leaving
h 1
.
r=
(13.11)
h 1(t1)
This is the Metropolis algorithm, of which the Metropolis-Hastings algorithm is an
extension. Most commonly, the proposal distribution is a normal distribution centered
at the previous value
(13.12)
q | (t1) = N | (t1) , 2 .
In that common case, the only parameter to select is the 2 . It may seem desirable
to have a value of 2 that will make the acceptance probability close to one. In that
case, very few draws are wasted, but that is not the most effective selection. Taken
to the extreme and trivial case, if 2 = 0 then = (t1) and h( ) = h( (t1) ) so
the acceptance probability will always equal 1, but all of the draws will be the same.
With small nonzero values of 2 , you may have high acceptance probabilities, but the
chain will travel slowly around the posterior distribution and the draws will be highly
autocorrelated. In contrast, if 2 is too large, then the acceptance probabilities will
be very low, and the chain will tend to stick (a long series of rejected proposals) and
again produce autocorrelated draws. For certain models, the optimal acceptance rates
are available (e.g., Gelman, Meng, and Stern 1996), but a general rule of thumb is
that the acceptance rate should be somewhere between 20% and 50%. In addition to
tracking the acceptance rate, a plot of the value of the parameter against its index in
the chain, called a trace plot, can show how well the chain is exploring the posterior
322
Parameter
2 0 2
200
400
600
800
1000
Index
Parameter
3 0 2
Proposal Variance: 5
200
400
600
800
1000
Index
Parameter
2 0 2
200
400
600
800
1000
Index
distribution. Figure 13.3 shows three trace plots at different values of 2 the first
too small, the last too large, and the middle near the theoretical optimum.
In modern applications, the proposal variance is chosen and updated adaptively. For
example, monitor the acceptance rate, and if the rate is too low decrease the proposal
variance. If the rate is too high, increase the variance. These adaptive proposals can
affect the convergence of the sampler. The simplest way to avoid that concern is
to adapt the proposal variance only during the burn-in period. WinBUGS does that
automatically during the burn-in period for those parameters that require a MetropolisHastings sampler.
323
and the authors give some suggestions on how to define those starting points. The
statistic determines whether the variance within the chains for (the parameter of
interest) is approximately equal to the variance across the chains for the last N
iterations. The within-chain variance is
2
1
ij j ,
W =
m(N 1) j =1 i=1
m
(13.13)
where ij is the i th draw of in chain j and j is the mean of from chain j . The
between-chain variance is
2
N
j ,
m 1 j =1
m
B=
(13.14)
where is the overall mean of from all the chains. The Gelman-Rubin statistic is
?
df
N 1 m+1 B
R =
+
,
(13.15)
N
mN W df 2
where df is the degrees of freedom from a t-distribution
that approximates the
324
from different locations. It is hoped that, they will all migrate to the same area and
bounce around there.
r Cross-correlations: Calculate the correlations between the different parameters. A chain
with highly correlated variables may be improved by proposing and sampling the
correlated variables together.
Finally, remember to still be an analyst. Using the mean of the posterior distributions
for the parameters, compare the density of the fitted model to the original data. Be
sure the results make sense. When using a poorly fitting model, the chain may still
converge and all the above metrics could give satisfactory results. Obviously, that
model should not be used.
(13.16)
y 1 eyi
() i
(13.17)
E[Y ] = /
(13.18)
Var[Y ] = / 2
(13.19)
(13.20)
(13.21)
The choice of prior distributions is explained in the next section. As a pilot run, five
simultaneous chains are initialized from various points and updated 20,000 times. The
Gelman-Rubin statistic is calculated for both and , using only the iterations from
one to k. The statistic is then plotted against k (see Figure 13.4) along with an upper
limit to determine the length of the burn-in period.
In properly designed MCMC chains, the Gelman-Rubin statistic will converge to
1 as k increases to infinity. Practically speaking, as long as the chain appears to be
converging to 1, the chains can be assumed to have reached the posterior distribution
when the Gelman-Rubin statistic is less than about 1.1.
In the annual claims data, the statistic drops down close to one around 10,000
iterations. That seems to be a good burn-in length to use. In addition, the trace plots
(see Figure 13.5) show that the chains move very quickly (within a few hundred
observations) from the initial values to the range of the posterior distribution, so
10,000 iterations should be more than enough.
325
1.5
1.4
shrink factor
1.2
1.3
1.1
1.0
1.0
1.1
shrink factor
1.2
1.3
1.4
1.5
Beta
35
25
Alpha
Fig. 13.4. Gelman-Rubin plots for and . The solid line is the estimate, and the
dashed line is an upper confidence limit. Note: When k is small the statistic is near 15
(far above the limits of this plot). Stretching the limits to display those points would
make it hard to distinguish the details of the rest of the plot.
1000
2000
3000
4000
5000
3000
4000
5000
Beta
Iteration
1000
2000
Iteration
326
0
1
5
10
50
250
500
1.000
0.9828
0.9162
0.8378
0.4034
0.0256
0.0104
1.000
0.9828
0.9163
0.8377
0.4033
0.0257
0.0102
Notice also that both plots for the two parameters are very similar, due to the high
cross-correlation between the two parameters. The autocorrelations for various lags
in the pilot run are in Table 13.2.
The autocorrelation is very high. The value for lag-250 is acceptable, so the sampler
will record only every 250th iteration. This is unusually high. The autocorrelation in
most chains is much smaller, but luckily this is a relatively simple model, allowing
the computational time to still be palatable even with the excessive thinning. The
chosen sampler contains five chains, each with a burn-in of 10,000 iterations and
150,000 samples (recording every 250th). That will provide 3,000 total draws from
the posterior distribution. The trace plots (see Figure 13.5) for the pilot run are satisfactory. Summary statistics for the posterior distributions of and are available in
Table 13.3.
Q1
Median
Mean
Q3
St. Dev.
29.30
3.929
29.71
3.984
29.71
3.984
30.12
4.039
0.607
0.082
327
the expert, such as, What is the chance that the mean claim amount is between $100
and $1,000? The resulting prior distribution can be as specific as desired, based on
the number of questions asked.
It is important to note that the range of the posterior distribution is limited by the
range of the prior distribution. If there is no probability assigned to a certain region
in the prior distribution, that region will also have no probability in the posterior
distribution, no matter what the data indicate. Alternatively, a standard distribution
with the same range as the parameter of interest could be used instead of a histogram.
Unfortunately, it may be difficult to translate the expert opinion into the distribution. It
is rather unproductive to ask, What parameters of a gamma distribution best describe
your current feelings about the mean loss? One option is to ask for summary statistics
such as the mean and variance and then translate those answers into parameters of a
distribution.
When dealing with human opinion, there is also a natural tendency to overstate
confidence in a conclusion. This tendency is known as the overconfidence effect
(Pallier et al. 2002). In a well-known example (Adams and Adams 1960), participants
were asked to spell difficult words and then state how confident they were in the
spelling. When they said they were 100% confident they misspelled the word 20%
of the time. Therefore, when experts are defining prior probabilities, they will tend to
put too much probability in the center of the distribution and not enough in the tails.
Even after they are warned and educated about the bias, overconfidence still exists
(Alpert and Raiffa, 1982). Therefore, to properly include expert opinion in the prior,
density may need to be taken from the center of the distribution and moved to the
tails.
13.5.2 Noninformative Priors
Strong priors are not always preferred. There may not be any prior information about
a particular parameter. Alternatively, with all the inherent concerns about overconfidence, a more objective approach may be preferred even when expert opinion is
available. Unfortunately, an objective or noninformative prior can be hard to define.
With a finite sample space, a natural choice is to divide the prior density evenly
throughout the space. For example, if the parameter is only defined on the integers
between 1 and N, then a prior of the form
1
{1, 2, . . . , N}
(13.22)
N
would be noninformative. Also, if the parameter is continuous over a bounded interval,
say [a, b], then
() =
( ) =
1
ba
[a, b]
(13.23)
328
log
f
(y|)
.
I () = E
2
(13.24)
(13.25)
The Jeffreys prior is a strong default choice when no prior information is known
about the parameters. Additionally, this prior is invariant to transformation. For a
more comprehensive review of constructing priors, please Berger (1985, chapter 3).
1 yi
.
y e
() i
(13.26)
329
Q1
Median
Mean
Q3
St. Dev.
29.30
29.27
30.26
29.71
29.68
30.58
29.71
29.69
30.58
30.12
30.08
30.91
0.607
0.614
0.477
(noninformative)
(mildly informative)
(strongly informative)
Notice that both the second and third priors have the same mean (E() = 32), but
wildly different variances (Var() = 1024 for prior 2 and Var() = 0.64 for prior 3).
Plots of the resulting posterior distributions for are presented in Figure 13.6
and summary statistics are available in Table 13.4. The posterior distributions for
Density
28
29
30
31
32
33
34
35
33
34
35
34
35
Alpha
Density
28
29
30
31
32
Alpha
Density
28
29
30
31
32
33
Alpha
Fig. 13.6. Prior (dashed line) and posterior (solid line) densities for the three different
prior specifications.
330
the mildly informative and non-informative priors are very similar, but the strongly
informative prior has a significant effect on the posterior. The value of Q1 for prior
three is greater than the values of Q3 for priors one and two. If the prior is strong
enough, it can even overwhelm large amounts of data (remember n = 4, 624 in the
auto claims data).
13.6 Conclusion
With the rapid increase in computational power, Bayesian methods are far more
accessible to actuaries. The methods provide a structure to incorporate prior expert
knowledge and properly account for model and parameter uncertainty. The methods
are also very natural for actuaries who already use credibility theory and are widely
applicable to essentially any model.
This chapter serves as an introduction to the basics of Bayesian statistics, but
is limited in scope. There are a number of well-written books devoted entirely to
Bayesian statistics that are available for a more thorough introduction (e.g., Carlin
and Louis 2009; Congdon 2003; Gelman et al. 2004).
13.7 Further Reading
As mentioned in the introduction, all of the models in this book can be formulated
and estimated from a Bayesian perspective. Although there is not sufficient space to
elaborate on each model individually, the following list provides several references
on how to apply the other models in a Bayesian setting:
r Multiple linear regression Chapter 14 in this text
r Regression with categorical dependent variables Chen, Ibrahim, and Yiannoutsos
(1999), Genkin, Lewis, and Madigan (2007)
r Regression with count dependent variables Clyde (2000); El-Sayyad (1973)
r Generalized linear models Dey, Ghosh, and Mallick (2000); Zeger and Karim (1991)
r Frequency/severity models Huang, Chin, and Haque (2008); Ma, Kockelman, and
Damien (2008); Neil, Fenton, and Tailor (2005)
r Mixed models Littell (2006); Ronquist and Huelsenbeck (2003)
r Generalized additive models Hastie and Tibshirani (1990); Wood (2006)
r Fat-tail regression models Fernandez and Steel (1998); Jacquier, Polson, and Rossi
(2004)
r Spatial statistics Eberly et al. (2000); Song et al. (2006)
r Machine learning Andrieu et al. (2003); Friedman et al. (2000)
r Time series, including Lee-Carter Czado, Delwarde, and Denuit (2005); Pedroza
(2006)
r Longitudinal and panel data models Cowles, Carlin, and Connett (1996); Daniels and
Pourahmadi (2002); Laird and Ware (1982)
References
331
r Credibility Klugman (1992); Makov, Smith, and Liu (1996); Young (1998)
r Survival models Ibrahim, Chen, and Sinha (2005); Kottas (2006); Mallick, Denison,
and Smith (1999)
r Claims triangles De Alba (2006); Zhang, Dukic, and Guszcza (2012)
Acknowledgments
I would like to thank the three editors and four anonymous referees for their thoughtful
and insightful comments. I am also grateful for discussions with and suggestions
from Brad Barney, Jenn Hartman, Matt Heaton, and Shujuan Huang. Their work and
suggestions greatly improved the final product.
References
Adams, P. and J. Adams (1960). Confidence in the recognition and reproduction of words
difficult to spell. American Journal of Psychology 73(4), 544552.
Albert, J. (2009). Bayesian Computation with R. Springer, New York.
Alpert, M. and H. Raiffa (1982). A progress report on the training of probability assessors. In
A. T. Daniel Kahneman and Paul Slovic (ed.), Judgment under Uncertainty: Heuristics
and Biases, pp. 294305. Cambridge University Press, Cambridge.
Andrieu, C., N. De Freitas, A. Doucet, and M. Jordan (2003). An introduction to MCMC for
machine learning. Machine Learning 50(1), 543.
Bayes, T. and M. Price (1763). An essay towards solving a problem in the Doctrine of
Chances. By the late Rev. Mr. Bayes, FRS communicated by Mr. Price, in a letter to
John Canton, AMFRS. Philosophical Transactions (16831775), 370418.
Berger, J. (1985). Statistical Decision Theory and Bayesian Analysis. Springer, New York.
Besag, J. (1974). Spatial interaction and the statistical analysis of lattice systems. Journal of
the Royal Statistical Society. Series B (Methodological), 192236.
Carlin, B. and T. Louis (2009). Bayesian Methods for Data Analysis, Volume 78. Chapman
& Hall/CRC, Boca Raton, FL.
Chen, M., J. Ibrahim, and C. Yiannoutsos (1999). Prior elicitation, variable selection and
Bayesian computation for logistic regression models. Journal of the Royal Statistical
Society: Series B (Statistical Methodology) 61(1), 223242.
Clyde, M. (2000). Model uncertainty and health effect studies for particulate matter.
Environmetrics 11(6), 745763.
Congdon, P. (2003). Applied Bayesian Modelling, Volume 394. Wiley, New York.
Cowles, M., B. Carlin, and J. Connett (1996). Bayesian Tobit modeling of longitudinal
ordinal clinical trial compliance data with nonignorable missingness. Journal of the
American Statistical Association, 8698.
Czado, C., A. Delwarde, and M. Denuit (2005). Bayesian Poisson log-bilinear mortality
projections. Insurance: Mathematics and Economics 36(3), 260284.
Daniels, M. and M. Pourahmadi (2002). Bayesian analysis of covariance matrices and
dynamic models for longitudinal data. Biometrika 89(3), 553566.
De Alba, E. (2006). Claims reserving when there are negative values in the runoff triangle:
Bayesian analysis using the three-parameter log-normal distribution. North American
Actuarial Journal 10(3), 45.
332
De Jong, P. and G. Heller (2008). Generalized Linear Models for Insurance Data. Cambridge
University Press, Cambridge.
Dey, D., S. Ghosh, and B. Mallick (2000). Generalized Linear Models: A Bayesian
Perspective, Volume 5. CRC, Boca Raton, FL.
Eberly, L., B. Carlin, et al. (2000). Identifiability and convergence issues for Markov chain
Monte Carlo fitting of spatial models. Statistics in Medicine 19(1718), 22792294.
El-Sayyad, G. (1973). Bayesian and classical analysis of Poisson regression. Journal of the
Royal Statistical Society. Series B (Methodological), 445451.
Fernandez, C. and M. Steel (1998). On Bayesian modeling of fat tails and skewness. Journal
of the American Statistical Association, 359371.
Friedman, N., M. Linial, I. Nachman, and D. Peer (2000). Using Bayesian networks to
analyze expression data. Journal of Computational Biology 7(34), 601620.
Gelfand, A. E. and A. F. M. Smith (1990). Sampling-based approaches to calculating
marginal densities. Journal of the American Statistical Association 85(410), 398409.
Gelman, A., J. B. Carlin, H. S. Stern, and D. B. Rubin (2004). Bayesian Data Analysis.
Chapman and Hall/CRC, Boca Raton, FL.
Gelman, A., X. Meng, and H. Stern (1996). Posterior predictive assessment of model fitness
via realized discrepancies. Statistica Sinica 6, 733759.
Gelman, A. and D. Rubin (1992). Inference from iterative simulation using multiple
sequences. Statistical Science 7(4), 457472.
Genkin, A., D. Lewis, and D. Madigan (2007). Large-scale Bayesian logistic regression for
text categorization. Technometrics 49(3), 291304.
Gilks, W. and P. Wild (1992). Adaptive rejection sampling for Gibbs sampling. Journal of the
Royal Statistical Society. Series C (Applied Statistics) 41(2), 337348.
Hartman, B. M. and M. J. Heaton (2011). Accounting for regime and parameter uncertainty
in regime-switching models. Insurance: Mathematics and Economics 49(3), 429437.
Hastie, T. and R. Tibshirani (1990). Generalized Additive Models. Chapman & Hall/CRC,
Boca Raton, FL.
Hastings, W. K. (1970, April). Monte Carlo methods using Markov chains and their
applications. Biometrika 57(1), 97109.
Huang, H., H. Chin, and M. Haque (2008). Severity of driver injury and vehicle damage in
traffic crashes at intersections: A Bayesian hierarchical analysis. Accident Analysis &
Prevention 40(1), 4554.
Ibrahim, J., M. Chen, and D. Sinha (2005). Bayesian Survival Analysis. Wiley Online Library.
Jacquier, E., N. Polson, and P. Rossi (2004). Bayesian analysis of stochastic volatility models
with fat-tails and correlated errors. Journal of Econometrics 122(1), 185212.
Jeffreys, H. (1961). Theory of Probability. Oxford University Press.
Klugman, S. (1992). Bayesian Statistics in Actuarial Science: With Emphasis on Credibility,
Volume 15. Springer, New York.
Kottas, A. (2006). Nonparametric Bayesian survival analysis using mixtures of Weibull
distributions. Journal of Statistical Planning and Inference 136(3), 578596.
Laird, N. and J. Ware (1982). Random-effects models for longitudinal data. Biometrics,
963974.
Littell, R. (2006). SAS for Mixed Models. SAS Publishing, Cary, NC.
Lunn, D., A. Thomas, N. Best, and D. Spiegelhalter (2000). WinbugsA Bayesian modelling
framework: Concepts, structure, and extensibility. Statistics and Computing 10(4),
325337.
Ma, J., K. Kockelman, and P. Damien (2008). A multivariate Poisson-lognormal regression
model for prediction of crash counts by severity, using Bayesian methods. Accident
Analysis & Prevention 40(3), 964975.
References
333
Makov, U., A. Smith, and Y. Liu (1996). Bayesian methods in actuarial science. The
Statistician, 503515.
Mallick, B., D. Denison, and A. Smith (1999). Bayesian survival analysis using a MARS
model. Biometrics 55(4), 10711077.
Metropolis, N., A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller (1953).
Equations of state calculations by fast computing machines. Journal of Chemical
Physics 21, 10871091.
Neil, M., N. Fenton, and M. Tailor (2005). Using Bayesian networks to model expected and
unexpected operational losses. Risk Analysis 25(4), 963972.
Ntzoufras, I. (2011). Bayesian Modeling using WinBUGS, Volume 698. Wiley, New York.
Pallier, G., R. Wilkinson, V. Danthiir, S. Kleitman, G. Knezevic, L. Stankov, and R. Roberts
(2002). The role of individual differences in the accuracy of confidence judgments.
Journal of General Psychology 129(3), 257299.
Pedroza, C. (2006). A Bayesian forecasting model: Predicting US male mortality.
Biostatistics 7(4), 530550.
Ronquist, F. and J. Huelsenbeck (2003). Mrbayes 3: Bayesian phylogenetic inference under
mixed models. Bioinformatics 19(12), 15721574.
Scollnik, D. (2001). Actuarial modeling with MCMC and BUGS. Actuarial Research
Clearing House: ARCH (2), 433.
Song, J., M. Ghosh, S. Miaou, and B. Mallick (2006). Bayesian multivariate spatial models
for roadway traffic crash mapping. Journal of Multivariate Analysis 97(1), 246273.
Sturtz, S., U. Ligges, and A. Gelman (2005). R2winbugs: A package for running WinBUGS
from R. Journal of Statistical Software 12(3), 116.
Whitney, A. W. (1918). The theory of experience rating. Proceedings of the Casualty
Actuarial Society 4, 274292.
Wikipedia (2012). Conjugate prior Wikipedia, the free encyclopedia. Accessed May 19,
2012.
Wood, S. (2006). Generalized Additive Models: An Introduction with R, Volume 66. CRC
Press, Boca Raton, FL.
Young, V. (1998). Robust Bayesian credibility using semiparametric models. ASTIN
Bulletin 28, 187204.
Zeger, S. and M. Karim (1991). Generalized linear models with random effects; a Gibbs
sampling approach. Journal of the American Statistical Association, 7986.
Zhang, Y., V. Dukic, and J. Guszcza (2012). A Bayesian non-linear model for forecasting
insurance loss payments. Journal of the Royal Statistical Society: Series A (Statistics in
Society).
14
Bayesian Regression Models
Luis E. Nieto-Barajas and Enrique de Alba
Chapter Preview. In this chapter we approach many of the topics of the previous chapters, but from a Bayesian viewpoint. Initially we cover the foundations of Bayesian
inference. We then describe the Bayesian linear and generalized regression models.
We concentrate on the regression models with zero-one and count response and illustrate the models with real datasets. We also cover hierarchical prior specifications
in the context of mixed models. We finish with a description of a semi-parametric
linear regression model with a nonparametric specification of the error term. We
also illustrate its advantage with respect to the fully parametric setting using a real
dataset.
14.1 Introduction
The use of Bayesian concepts and techniques in actuarial science dates back to
Whitney (1918) who laid the foundations for what is now called empirical Bayes
credibility. He noted that the solution of the problem depends upon the use of inverse
probabilities. This is the term used by T. Bayes in his original paper (e.g., Bellhouse
2004). However, Ove Lundberg was apparently the first one to realize the importance
of Bayesian procedures (Lundberg 1940). In addition, Bailey (1950) put forth a clear
and strong argument in favor of using Bayesian methods in actuarial science. To date,
the Bayesian methodology has been used in various areas within actuarial science;
see, for example, Klugman (1992), Makov (2001), Makov, Smith, and Liu (1996), and
Scollnik (2001). A brief look at recent issues of the main journals in the field shows that
Bayesian applications appear regularly and cover a broad range of actuarial topics; for
example, mortality modeling (Cairns et al. 2011); extreme observations (Cabras and
Castellanos 2011); mixture models (Bernardi, Maruotti, and Petrella 2012); premium
policies (Landriault, Lemieux, and Willmot 2012); and loss reserving (Shi, Basu, and
Meyers 2012).
334
335
Bayesian methods have several advantages that make them appealing for use in
actuarial science. First, they allow the actuary to formally incorporate expert or
existing prior information. This prior information can be in the form of global or
industry-wide information (experience) or in the form of tables. In this respect it
is indeed surprising that Bayesian methods are not used more extensively, because
there is a wealth of objective prior information available to the actuary. In fact, the
structure distribution frequently used in credibility was originally formulated in a
Bayesian framework (Buhlmann 1967).
The second advantage of Bayesian methods is that the analysis is always done by
means of the complete probability distribution for the quantities of interest, either the
parameters or the future values of a random variable. Actuarial science is a field where
adequate understanding and knowledge of the complete distribution are essential.
In addition to expected values we are usually looking at certain characteristics of
probability distributions (e.g., ruin probability, extreme values, value-at-risk (VaR),
and so on).
From a theoretical point of view, Bayesian methods have an axiomatic foundation
and are derived from first principles (Bernardo and Smith 2000). From a practical
perspective, Bayesian inference is the process of fitting a probability model to a set of
data and summarizing the uncertainty by a probability distribution on the parameters
of the model and on unobserved quantities, such as predictions for new observations.
A fundamental feature of Bayesian inference is the direct quantification of uncertainty.
To carry it out, the actuary must set up a full probability model (a joint probability
distribution) for all observable and unobservable quantities in a given problem. This
model should be consistent with knowledge about the process being analyzed. Then,
Bayesian inference about the parameters in the model or about unobserved data are
made in terms of probability statements that are conditional on the observed data
(posterior distributions). Hence, these methods provide a full distributional profile for
the parameters, or other quantities of interest, so that the features of their distribution are readily apparent; for example, non-normality, skewness, or tail behavior. To
obtain these posterior distributions, Bayesian methods combine the prior available
information, no matter how limited, with the theoretical models for the variables of
interest. Therefore Bayesian models automatically account for all the uncertainty in
the parameters.
336
analyst about the occurrence of a given event, and the criterion of maximizing expected
utility is seen to be the only criterion compatible with the axiomatic system. Statistical inference is viewed as a particular decision problem, and whether estimation or
prediction, it must follow the laws of probability. As a result, the uncertainty of all
unknown quantities is described in terms of probability distributions, which implies
that these quantities are treated as random variables. The fact that parameters have a
distribution function allows the application of Bayes theorem to combine information
coming from the data with prior information about the parameters. For a comprehensive exposition on the foundations of Bayesian theory, see Bernardo and Smith (2000)
and references therein.
The ensuing methodology establishes how to formally combine an initial (prior)
degree of belief of a researcher with currently measured, observed data, in such a
way that it updates the initial degree of belief. The result is called posterior belief.
This process is called Bayesian inference because the updating process is carried out
through the application of Bayes theorem. The posterior belief is proportional to the
product of the two types of information: the prior information about the parameters
in the model and the information provided by the data. This second part is usually
thought of as the objective portion of the posterior belief. We explain this process as
follows:
Let y = (y1 , y2 , . . . , yn ) be independent random variables, each of them coming
from a probability model with density function f (yi |), where is a parameter
9
vector that characterizes the form of the density. Then f (y| ) = ni=1 f (yi |) is the
joint probability density of y given , which is usually referred to as the likelihood
function. Prior available information on the parameter is described through a prior
distribution ( ) that must be specified or modeled by the actuary. Then, from a
purely probabilistic point of view, it follows that
( | y) =
f (y | )()
f (y)
>
where f (y) is the marginal joint density of y defined as f (y) = f (y|) () d if
is continuous, and as f (y) = f (y|)() if is discrete. This is Bayes theorem,
which rules the updating of the information. Considering that f (y) is just a constant
for , then the updating mechanism can be simply written as (|y) f (y| )(),
where indicates proportionality. In other words, the posterior distribution of the
parameters, conditional on the observed data, is proportional to the product of the
likelihood function and the prior degree of belief. Any inference on the parameters is
now carried out using the posterior distribution ( |y).
As was mentioned earlier, the only criterion for optimal decision making, consistent
with the axiomatic system, is the maximization of the expected utility. Alternatively,
this criterion is equivalently replaced by the minimization of a loss function. Therefore,
337
There is a vast literature on how to specify a prior distribution. One of the most
common approaches is to use the family of natural conjugate priors. A prior distribution ( ) is said to be a natural conjugate for if, when combining it with the sample
information, ( ) and the resulting posterior (|y) belong to the same family. These
priors can be used to produce vague or diffuse priors, which reflect knowing little or
having no prior information about the parameter, or to produce informative priors that
338
reflect the prior knowledge of the actuary. In either case this is achieved by setting
the parameters of the prior to an appropriate value. In particular, vague priors are
obtained by letting the prior variance be large, but in fact these priors are just an
approximation of what are called noninformative (or objective) priors (e.g., Berger
2006). More details about Bayesian thinking can be found in Chapter 13.
(14.2)
where is interpreted as a measurement error and is assumed to have zero mean and
constant precision (reciprocal of the variance). If we further assume that the error
comes from a normal distribution, then N(0, ).
The normal assumption in the error term implies that the response variable y,
conditional on x, also follows a normal model and thus can take any possible value on
the real line. Generalized linear models extend this assumption to response variables
with positive, bounded, or discrete outcomes. For example, if one wants to describe
the behavior of the amounts in insurance claims or the number of claims in a certain
period of time, the normal assumption would not be adequate in either case because
claims cannot be negative and the number of claims is positive and discrete.
The role played by the explanatory variables x in the linear normal (linear regression) model is to help in understanding the average or mean behavior of the response
variable y. This is why the (conditional) expected value E(y|x) is equal to a linear combination of the explanatory variables x . This property justifies the name
regression to the mean of the linear regression model (14.2).
The linear regression model is a particular case of the larger class of generalized
linear models. We discuss its properties and Bayesian inference in the following
sections.
339
can also be considered regression models to the mean (of y), but in a nonlinear form
since the parameter space of E(y|x) is not necessarily the whole real line. Let us
start by recalling the form of a generalized linear model. To account for all possible
kinds of response variables (positive, bounded, discrete, etc.), the model describes
the probabilistic behavior of the responses with a member of the exponential family.
Then, for a sample of independent random variables y1 , y2 , . . . , yn , each of them
comes from the model
f (yi | i , i ) = b(yi , i ) exp[i {yi i a(i )}],
(14.3)
where a() and b() are two monotonic functions. The parameters i and i are known
as natural and dispersion parameters, respectively. It is not difficult to show that the
mean and variance of yi can be expressed in terms of derivatives of function a() as
follows:
i = E(yi ) = a (i ) and i2 = Var(yi ) =
a (i )
.
i
Here prime and double prime denote the first and second derivative, respectively. Note
that generalized linear models have also been described in Chapter 5 with a slightly
different parameterization of the exponential family.
Each individual i has its own set of explanatory variables xi , i = 1, . . . , n.
These will be combined in a single value through a linear combination with coefficients = (0 , 1 , . . . , p1 ), forming what is called the linear predictor i =
xi = 0 + 1 xi1 + + p1 xi,p1 . We note that linearity means linear in the coefficients because the linear predictor could well be a polynomial of order p 1 of the
p1
form 0 + 1 xi + + p1 xi
with a single explanatory variable for individual
i, xi .
The idea of the generalized linear models is to model the mean of the response
variable, i = E(yi ), in terms of the explanatory variables via the linear predictor
i and an appropriate transformation g(); that is, i = g(i ). The function g() is
called the link function because it links the explanatory variables with the response.
At the same time, the link function adjusts the parameter space of i to correspond
to the values of the predictor i , which is typically the real line. This can be seen as
i = g 1 (i ). A particular choice for the link function g() is to take g 1 () = a ().
In this case the linear predictor i becomes equal to the natural parameter i and g()
is called the canonical link function. Other options for the link function are available,
as long as the domain of the function g() corresponds to the parameter space of i
and the image to the real numbers. Let us consider two examples to illustrate these
ideas.
It can be shown that the normal linear regression model is also a generalized linear
model. To see this we take yi N(i , i ), parameterized in terms of mean i and
340
a(i ) =
i2
2
In this case a (i ) = i and thus the canonical link is g(i ) = i . Therefore the
mean i is modeled directly, with the linear predictor i obtaining the linear model
i = xi .
A second example, suitable for response variables in the positive real line (as
is the case for claim amounts in insurance companies), is to consider a response
with gamma distribution. Using a mean parameterization of the gamma that is,
yi gamma(i , i /i ) such that E(yi ) = i then the density function is of the
form
i
1
i
iy
yii 1 e i i ,
f (yi | i , i ) =
i
(i )
for yi > 0 and i , i > 0. Writing this gamma density as in (14.3) we get
i = i ,
i = 1i ,
b(yi , i ) =
i i
1
y i .
(i ) i
a(i ) = log 1i .
341
these parameters are available only in very special cases. In general, posterior distributions are not available in closed forms so the choice of the prior has to do with
the simplicity to accommodate prior beliefs. In this context a normal prior for the
vector has been the common choice, together with a gamma prior for the precision typically assuming independence a priori among the j elements of and
between and . The normal and gamma are well-known models that allow the
user to specify prior beliefs in terms of mean and precision (reciprocal of variance).
That is, we can take a priori j N(b0 , t0 ), where b0 is the prior mean and t0 the
prior precision for j . In the case of little or no information about j , we can set
b0 = 0 together with t0 close to zero, say 0.1, 0.01 or 0.001, for j = 1, . . . , p. For the
sampling precision parameter , if and 2 represent the prior mean and variance,
then we can take gamma(a0 , a1 ) with a0 = 2 /2 and a1 = /2 . Again, in
the case of little or no prior information about , we can set a0 = a1 equal to a small
value, say 0.1, 0.01 or 0.001, in such a way that has prior mean one and large/small
prior variance/precision. Alternatively, more diffuse priors are also considered for the
coefficients j ; for instance, a student-t prior or even a cauchy prior (Gelman et al.
2008).
Posterior inference of the parameters (, ) requires combining the information
provided by the data, summarized in the likelihood function, and the prior distributions. The likelihood function is constructed by the product of the density (14.3) of the
response variables as a function of the explanatory variables and the parameters; that
9
is, lik(, ) = ni=1 f (yi | i , ). Remember that the explanatory variables enter the
model via the natural parameter i , which in the case of using the canonical link
i = xi ; otherwise i is replaced with an appropriate function of the linear predictor
i . Finally, the posterior distribution (, | data) is proportional to the product of
this likelihood function lik(, ) and the prior distributions (, ). Point estimates
and credible intervals are obtained as summaries from this posterior distribution. A
Bayesian credible interval is also known as a posterior probability interval, which is
not to be confused with a frequentist confidence interval. For example, a 95% posterior
interval for a parameter is an interval that contains exactly 95% of that parameters
posterior probability. More details on this can be found in Gelman et al. (2008).
Posterior summaries are obtained numerically via a Markov Chain Monte Carlo
(MCMC) sampling algorithm or via Estimation Maximization (EM) techniques.
The former can be implemented in OpenBugs (http://www.openbugs.info/) within
R through the library R2OpenBUGS. The latter is implemented in the R command bayesglm from the package arm (data analysis using regression and multilevel/hierarchical models). Both are available in the Comprehensive R Archive Network (CRAN) at http://www.r-project.org/. All the examples presented in this chapter
were run in OpenBugs within R, and the corresponding code is given in each case.
In all cases the Markov chains were run for 20,000 iterations with a burn-in period
342
of 5,000 and keeping one of every 5th iterations for computing the estimates. See
Chapter 13 for the meaning of these numbers.
Another aspect of interest when using generalized regression models is prediction
of future outcomes. This inference problem is addressed naturally in the Bayesian
approach by computing the predictive distribution for a future observation yF . If a
new individual has explanatory variables xF , and assuming the canonical link, then
F = xF and the predictive distribution will be the weighted average of the density
f (yF | F , ) with respect to the posterior distribution (, | data) as in (14.1). Point
or interval predictions are produced using summaries from this predictive distribution.
This is usually done numerically.
Traditionally, the goodness-of-fit measure used to compare generalized linear
models is the deviance (see, Chapter 5). However, in a Bayesian context, model
comparison is typically made by using the deviance information criterion (DIC)
(Spiegelhalter et al. 2002), which is based on the deviance but includes a penalization for the number of parameters used in the model. Let D() = 2 log f (y| );
with = E( |y). Smaller values of DIC indicate a
then DIC = 2 E |y (D) D(),
better fit.
Example 14.1. The insurance market in Mexico operates in different sectors. Seven of
these are accident and sickness (ACC), agriculture and livestock (AGR), automobiles
(AUT), major medical expenses (MED), fire (FIR), liability and professional risks
(LIA), and health (HEA). It is of interest to the insurance companies to predict claim
amounts yi in terms of the premiums written xi , both measured in millions of Mexican
pesos. The insurance industry regulator in Mexico gathers the information from all
different insurance companies every year and makes the information available on its
web page http://www.cnsf.gob.mx/. The information is available for all 32 Mexican
states and in some cases from abroad (in the data 4 sectors showed information from
abroad). In total, for the year 2010, we have i = 1, . . . , n with n = 228 observations
classified by insurance sector. The dataset can be found in the Web Appendix of the
book. A dispersion diagram of the 228 observations in logarithmic scale is presented
in Figure 14.1. From the graph we can see that all sectors together follow a common
pattern and a single line could potentially serve for fitting the data. In fact, the least
square estimates are 0.0008 for the intercept and 0.85 for the slope.
Let us assume that the logarithm of the claim amounts log(yi ) follows a normal
distribution with mean i and constant precision ; that is, log(yi ) N(i , ). We
model the mean level i in terms of a linear combination of the premiums written
in log scale log(xi ) and class indicators zj i , for j = 2, . . . , 7, where for example
z2i takes the value of one if observation i belongs to sector 2 (AGR), and so on,
following the order of the sectors in the previous paragraph. These sector indicators will serve to determine possible differences in the intercepts and the slopes
343
log(y)
10
log(x)
344
}
#Priors
for (j in 1:7){
a[j] ~ dnorm(0,0.001)
b[j] ~ dnorm(0,0.001)
}
tau ~ dgamma(0.001,0.001)
}
Semi-Parametric
Coef.
Mean
95% CI
Mean
95% CI
1
2
3
4
5
6
7
1
2
3
4
5
6
7
1.77
1.29
2.21
2.92
0.96
2.49
1.57
0.26
0.39
0.76
0.84
0.41
0.58
0.49
(1.34, 2.20)
(2.24, 0.32)
(4.22, 0.15)
(4.01, 1.72)
(2.11, 0.15)
(3.48, 1.54)
(2.11, 1.06)
(0.14, 0.38)
(0.13, 0.64)
(0.45, 1.06)
(0.61, 1.04)
(0.17, 0.67)
(0.35, 0.84)
(0.33, 0.65)
0.48
1.00
0.45
1.79
0.66
1.47
0.11
0.64
0.31
0.32
0.48
0.24
0.25
0.13
(0.10, 1.63)
(2.20, 0.23)
(2.01, 0.63)
(2.99, 1.01)
(2.30, 0.54)
(2.46, 0.63)
(1.19, 0.57)
(0.41, 0.80)
(0.12, 0.57)
(0.13, 0.62)
(0.27, 0.73)
(0.01, 0.57)
(0.01, 0.45)
(0.04, 0.31)
345
log(y)
10
log(x)
346
the real line. Since the parameter space of i is the interval (0, 1), any function that
transforms the (0, 1) interval into real numbers is a suitable link function. In our basic
probability courses we learned that cumulative distribution functions (c.d.f.) for continuous random variables are functions with real domain and (0, 1) image. Therefore,
the inverse of any continuous c.d.f. F () can be a link function; that is, g() = F 1 ().
In particular, if F = , the c.d.f. of a standard normal, then g() = 1 () produces
the probit regression model. In fact, the inverse of the canonical link corresponds
to the c.d.f. of a logistic random variable; thus the name of logistic regression. Two
other common link functions are the log-log link g(i ) = log{ log(i )} and complementary log-log link g(i ) = log{ log(1 i )}. This latter link corresponds to
the inverse of the c.d.f. of a extreme value distribution. Whatever the link function we
choose, for a given vector of explanatory variables of individual i, the probability of
success is expressed in terms of the explanatory variables as i = g 1 (i ) = F (xi ).
Sometimes several individuals share the same value of the explanatory variables,
or it is also possible that available information is grouped and the covariate information is only available at the group level. This is the case, for example, in insurance
groups where it is assumed that all individuals in the same group show similar risk
characteristics and the number of claims yi out of ni members in group i is reported.
In such a case it is of interest to estimate the probability of presenting a claim i
for group i with characteristics xi . These kind of data can also be modeled with a
generalized linear model by assuming yi binomial(ni , i ) and i = F (xi ) with a
specific choice of continuous c.d.f. F 1 () as the link function.
For both models, Bernoulli and binomial, the precision parameter i is equal to one,
and for the grouped data, the number of individual in the group ni is assumed known.
This leaves us with one set of unknown parameters . For each j , j = 1, . . . , p, we
assign normal and student-t prior distributions as suggested earlier.
Example 14.2. The Mexican Central Bank is responsible for issuing the required
number of bills for the effective functioning of the economy. Table 14.3 contains
information on the number of bills in circulation and the number of fake bills, both in
million pieces, for different denominations ($20, $50, $100, $200, and $500 Mexican
pesos). This information is available annually from the year 2000 to 2011. Let us
disregard temporal dependence and assume that the number of fake bills yi follows
a binomial distribution with parameters ni , the number of circulating bills, and i ,
the proportion of fake bills with respect to the real bills circulating; that is, yi
binomial(ni , i ), for i = 1, . . . , n with n = 60 observations. To help understand the
information contained in Table 14.3, we present boxplots of the crude proportion of
fake bills for every thousand circulating bills; that is, = yi /ni 1000. Figure 14.3
shows these fake proportions across the different bill denominations: the $20 bill is
347
Table 14.3. Number of Bills in Circulation (C) and Number of Fake Bills (F), in
Million Pieces, for Different Bill Denominations for Years 2000 to 2011
Year
C20
F20
C50
F50
C100
F100
C200
F200
C500
F500
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2182.1
2092.6
2182.4
2449.1
2545.8
2707.8
2877.4
2959.8
3360.9
3578.6
3707.6
3858.8
14.8
13.1
18.1
9.4
1.5
1.0
0.7
0.6
1.0
3.2
2.7
1.3
3141.4
2900.9
3026.5
4245.0
4031.8
3420.2
3615.2
3847.5
3892.8
4129.0
4197.3
4375.1
179.7
150.5
109.7
140.9
149.2
249.3
215.1
122.5
59.4
28.3
67.9
208.9
2779.4
2795.0
3155.5
4455.4
4951.7
4411.0
4625.9
4768.0
4830.2
4872.5
5210.0
5416.0
178.5
136.8
64.2
60.1
117.8
142.9
106.5
77.1
87.6
81.0
101.2
88.7
4163.4
4745.5
5192.1
4870.4
5087.4
5422.1
5935.6
6358.0
6850.7
7314.7
7505.1
7528.1
83.9
64.4
97.3
77.6
80.5
117.8
88.8
78.6
97.7
136.3
139.7
120.1
1100.7
1335.0
1802.1
2352.4
3028.0
3522.5
4190.9
4889.7
5682.5
6934.4
7799.3
8907.4
26.6
20.9
35.7
42.9
34.0
43.6
70.9
90.5
91.7
91.2
96.4
89.7
20
40
60
the least falsified bill with low dispersion, and the $50 bill is the most falsified with a
large dispersion across the years. Figure 14.4 presents the fake proportions across the
years, showing a decreasing path in time, both in location and dispersion.
To identify the bill denomination we construct auxiliary dummy variables, say
xj i , j = 1, . . . , 5 for each of the five bill denominations $20, $50, $100, $200, and
$500, respectively, such that xj i takes the value of one if observation i corresponds to
Bill20
Bill50
Bill100
Bill200
Bill500
Fig. 14.3. Boxplots of crude fake proportions multiplied by 1,000 for the different
bill denominations in Table 14.3.
20
40
60
348
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
Fig. 14.4. Boxplots of crude fake proportions multiplied by 1,000 for the different
years in Table 14.3.
the bill denomination j and zero otherwise. We then define a first model (Model A)
with a linear predictor of the form i = + 5j =1 j xj i . By considering a different
parameter j for each denomination plus an intercept, Model A is overparametrized
and the j s will not be estimable. To identify the parameters we include the constraint
j j = 0. We then compare the logistic and standard normal links. These two links
imply i = ei /(1 + ei ) and i = (i ), respectively. For the prior distributions on
and the j s, we consider two alternatives N(0, 0.001) and St(0, 0.001, 3). Note that
the third parameter in the student-t distribution corresponds to the degrees of freedom,
which have to be greater than 2. We chose 3 to avoid numerical problems. This model
is translated into R (BUGS) code as shown in Table 14.4. Note that the last two
rows of the code describe the inclusion of the constraint on the parameters.
To compare the different models, we compute the deviance information criterion
(DIC). This value is easily obtained from the library R2OpenBUGS. Table 14.5 shows
the DIC values for the competing models. As can be seen from the table, the logit
link is preferred to the probit link for this particular dataset, regardless of the prior
distribution used. When comparing the two priors, the fit is practically the same, with
a negligible advantage for the normal model. For producing inference of the model
parameters, we select the logit-normal model since it achieved the smallest DIC
value.
We define the rate of fake bills for every thousand circulating bills as pj =
e+j /(1 + e+j ) 1,000, for each of the five bill denominations j = 1, . . . , 5.
Table 14.6 contains posterior point estimates (posterior means) as well as 95% credible intervals. Interpreting the estimated rates, the $50 pesos bill has the largest fake
rate, with 37.5 fake bills for every 1,000 circulating bills. In contrast, the $20 pesos
349
}
#Priors
a ~ dnorm(0,0.001)
for (j in 1:5){
b[j] ~ dnorm(0,0.001)
#b[j] ~ dt(0,0.001,3)}
#Estimable parameters
a.adj<-a+mean(b[])
for (j in 1:5){b.adj[j]<-b[j]-mean(b[])}
bill shows the smallest rate, with almost 2 fake bills for every 1,000 circulating.
According to Figure 14.3, the distribution of crude proportions for the $200 and $500
denominations seems to overlap. However, inference derived from the binomiallogistic-normal model shows that these rates have 95% credible intervals that do not
intersect (13.23, 15.27) for the $500 bill and (15.73, 17.63) for the $20 bill. In fact,
the posterior probability of p4 > p5 is 0.9995.
As suggested by Gelman et al. (2004), an important role of Bayesian inference is to
assess model fitting. Apart from the numerical measures, such as the DIC, it is possible
to obtain posterior predictive draws from the fitted model and compare them to the
actual observed data. This is known as posterior predictive checking. We sampled
Table 14.5. Deviance Information Criterion (DIC) for Four Binomial
Generalized Models Fitted to the Mexican Central Bank Data
Model A
Prior
Normal
Student-t
Model B
Logit
Probit
Logit
Probit
1295.87
1295.98
1297.45
1297.49
901.98
902.02
919.38
919.44
350
Table 14.6. Posterior Inference for the Rate of Fake Bills for Every Thousand
Circulating Bills, pi , i = 1, . . . , 5, Using the Mexican Central Bank Data, Posterior
Mean and 95% Credible Intervals
Model A
$
20
50
100
200
500
Model B
Rate
Mean
2.5%
97.5%
Mean
2.5%
97.5%
p1
p2
p3
p4
p5
1.95
37.50
23.77
16.66
14.24
1.53
35.75
22.49
15.73
13.23
2.45
39.28
25.10
17.63
15.27
1.95
37.10
23.87
16.65
15.48
1.51
35.33
22.56
15.72
14.37
2.44
38.90
25.22
17.63
16.63
p2
10 20 30 40 50 60 70
p1
351
Obs A1 A2 A3 A4 A5 B1 B2 B3 B4 B5
Obs A1 A2 A3 A4 A5 B1 B2 B3 B4 B5
p4
10
20
15
30
20
40
25
50
30
60
p3
Obs A1 A2 A3 A4 A5 B1 B2 B3 B4 B5
Obs A1 A2 A3 A4 A5 B1 B2 B3 B4 B5
10
15
20
25
p5
Obs A1 A2 A3 A4 A5 B1 B2 B3 B4 B5
Fig. 14.5. Boxplots of posterior predictive checks for the proportion of fake bills for
every one thousand circulating bills, pi , i = 1, . . . , 5. Reported are observed rates
(Obs), five replicates from Model A (A1 , . . . , A5 ), and five replicates from Model B
(B1 , . . . , B5 ).
Figure 14.5 correspond to five replicates from Model B. As is clear from the figure,
Model B better captures the variability of the data and the inferences obtained from it
are more reliable. Although Model B has improved on Model A, the model is a little
deficient in capturing the variability in the $20 pesos bills p1 .
352
i i
,
yi !
b(yi , i ) =
1
.
yi !
a(i ) = ei .
Thus the canonical link obtained as the inverse of the derivative of function a() is
g(i ) = log(i ). Therefore, the mean of the response variable is modeled as i = exi .
This model is also known as Poisson regression model (see Chapter 4 for regression
models with count dependent variables).
Sometimes, instead of modeling the mean i of a Poisson response variable, it is
of interest to model the rate of occurrence of events i relative to a known number
exposed or at risk ei , such that i = ei i . In this case the rate is modeled through the
explanatory variables as i = exi . For instance, in mortality studies, the maternity
mortality ratio is defined as the rate of maternity deaths for every 100,000 births.
In such studies the response variable yi is the number of maternity deaths, ei is the
number of births (in 100,000s), and thus i becomes the maternity mortality ratio.
As mentioned before, in regression models with count dependent variables, the
Poisson model is the common assumption; however, this model assumes that the mean
and variance of the responses are the same; that is, E(yi ) = Var(yi ) = i . In practice,
this assumption is not always satisfied by the data, due to an effect of overdispersion
(Var(yi ) > E(yi )). To account for overdispersion in a dataset, a different model for
the responses has to be used. The negative binomial is the typical alternative for
modeling counting data in the presence of overdispersion (i.e., yi NB(ri , i )).
To give the parameters of the negative binomial the same interpretation as in the
Poisson model, the integer parameter ri has to coincide with the number of exposed
ei and the probability of success i with 1/(1 + i ). This implies that E(yi ) = ei i
and Var(yi ) = ei i (1 + i ). The quantity 1 + i = Var(yi )/E(yi ) is a measure of the
amount of overdispersion present in the data. Finally, the rate i is modeled in terms
of the explanatory variables xi as in the Poisson model.
In the Bayesian literature (e.g., Bernardo and Smith 2000), it is well known that
a negative binomial distribution is a particular case of a Poisson-gamma distribution.
The latter gets its name because it can be obtained as a mixture of a Poisson distribution
353
10
15
2000
2002
2004
2006
Time
2008
2010
Fig. 14.6. Number of fake $20 peso bills from 2000 to 2011. Empty dots (linked
with solid black lines) represent observed values (F20). Red straight lines correspond
to rate estimates (solid line) and 95% credible intervals (dotted lines). Vertical grey
dashed-dotted line corresponds to year change estimate.
354
prior for on the set {2000, 2001, . . . , 2011}. The corresponding R (Bugs) code is
presented in Table 14.7.
The red lines, solid and dotted in Figure 14.6, correspond to the rate i point
estimates and 95% credible intervals, respectively, for all years. Years 2000 to 2003
(inclusive) have a common rate of 13.71 million fake pieces per year with a 95%
credible interval of (10.17, 17.52), whereas from 2004 to 2011 the rate drops to 1.28
million pieces with a 95% credible interval of (0.69, 2.12). The posterior distribution
for the change year, , concentrates its probability in only two values: 2003 with a
probability of 0.0006, and 2004 with a probability of 0.9994. The estimated change
year is denoted with a light grey (dotted-dashed) vertical line in Figure 14.6. Although
the new $20 pesos bills were introduced at the beginning of 2003, the impact on the
counterfeit rate was reflected from 2004 onward. The number of fake bills in 2003
was more similar to the previous years than the number in the following years. This
was captured by the model.
355
assumes two sets of explanatory variables xi and zi such that the former is associated
to fixed coefficients and the latter to random coefficients i . Thus, in the context
of a generalized linear model, the mixed model has a linear predictor of the form
i = xi + i zi . This is again linked to the response variable through an appropriate
link function such that g(i ) = i .
To better understand how dependence is introduced in a mixed model, let us
consider a nested structure for the observations, say yij , where j = 1, . . . , ni and
i = 1, . . . , n. For example, individual i could file a total of ni number of claims
during a year, with j denoting the specific claim. In this case, a mixed model for
the claim amounts yij , in a generalized linear setting, would have linear predictor
ij = xij + i . Note that the parameter is the fixed effect component common to
all individuals, whereas i is a random effect common to all claims j made by the
same individual i and thus introducing a dependence in those claims made by the
same individual.
Specifications for the random effects i s may vary according to the application.
iid
They could simply be i N(0, ), which is the typical specification in repeated
measurements, clustered observations, and longitudinal models. Alternative specifications include spatial effects (1 , . . . , n ) CAR(, ), where CAR stands for
a conditionally autoregressive model with association parameter and precision .
This model is a multivariate normal whose precision matrix is based on the spatial
neighborhood structure. We refer the reader to Chapter 11 or alternatively to Banerjee, Carlin, and Gelfand (2004) for details. These specifications also include temporal
iid
effects i = i1 + i , with i N(0, ), following a dynamic equation to account
for dependence over time. We refer the reader to West and Harrison (1997) for details.
In the following subsection we describe an alternative specification for the random
effects that is based on the idea of exchangeability. See Chapter 8 for more details on
mixed models.
14.4.2 Hierarchical Models
According to Gelman et al. (2004) hierarchical models are the most powerful tool
for data analysis. Hierarchical specifications are usually helpful in specifying a joint
prior distribution for a set of parameters. However, they are also useful for specifying
the distribution of random effects in a mixed model.
To describe the construction of a hierarchical model, let us consider a simple
scenario with response variables yi , for i = 1, . . . , n, where the distribution of each
yi depends on a parameter ; that is, f (yi |). This scenario assumes that there is a
unique parameter common to all individuals. So inference on will be based on a
prior distribution ( ) and all observations yi s (as in a traditional Bayesian analysis
356
i | (i | ), i = 1, . . . , n
().
The parameter is called the hyper-parameter and plays the role of an anchor of the
i s. Conditional on , the i s are independent,
> 9n and when is marginalized the i s
become dependent (i.e., (i , . . . , n ) =
i=1 (i |)()d). The hierarchical
model is completed by specifying the distribution of the data, which in general would
be yi f (yi |i ) independently for i = 1, . . . , n. This specification is analogous to the
so-called structure distribution that is frequently used in actuarial science, specifically
in credibility theory.
Hierarchical models are particularly useful for meta-analysis, where information
coming from different studies yi is linked via a hierarchical prior distribution on the
different parameters (1 , . . . , n ). Global or population inference from all studies is
usually summarized in terms of the hyper-parameter .
Example 14.4. Regarding the bills dataset of Table 14.3, consider now that the individuals i are the different bill denominations for i = 1, . . . , n with n = 5. For each
bill denomination i we have ni = 12 observations j = 1, . . . , ni corresponding to the
12 years. For each observed number of fake bills yij we assume a Poisson model
of the form yij Poisson(i ) with log(i ) = i . Here we have two options: take
independent priors for each i , say i N(0, 0.001) for i = 1, . . . , 5, or take an
exchangeable prior for the vector (1 , . . . , 5 ) with hierarchical representation given
by j |0 , N(0 , ) with 0 N(0, 0.001) and gamma(10, 1). Here the crucial parameter is . Since is a precision parameter for the i s, a small value would
imply a large uncertainty, allowing a broad combination of information across different
individuals i, whereas a large value reduces the uncertainty around 0 and constrains
357
}
#Priors
for (i in 1:5){
b[i] ~ dnorm(0,0.001)
#b[i] ~ dnorm(b0,tau)
}
#b0 ~ dnorm(0,0.001)
#tau ~ dgamma(10,1)
}
the sharing of information across different is. The prior we took for , gamma(10, 1)
is a slightly informative prior that allows a moderate sharing of information. The R
(BUGS) code of this model is presented in Table 14.8.
Posterior estimates for the two prior choices, independent and hierarchical, are
reported in Table 14.9 and presented in Figure 14.7. Numerical values are very
similar when using both priors. If we concentrate in the point estimates of the i s,
we can see that for those denominations with the smallest rates (F20 and F500)
their point estimates increase when using a hierarchical prior with respect to those
with independent priors. In contrast, for those denominations with the largest rates
(F50 and F100) their point estimates decrease. These effects are the result of sharing
information between models; the estimates tend to compromise among all pieces
Table 14.9. Posterior Estimates of Fake Rates for Different Bill Denominations
Under Distinct Scenarios
Independent
Hierarchical
Coef.
Variable
Mean
95% CI
Mean
95% CI
1
2
3
4
5
F20
F50
F100
F200
F500
F
5.62
140.16
103.57
98.58
61.17
81.76
(4.39, 7.06)
(133.65, 147.10)
(97.79, 109.40)
(93.03, 104.05)
(56.83, 65.68)
(79.54, 84.12)
5.74
140.15
103.53
98.54
61.22
58.05
(4.47, 7.20)
(133.70, 147.10)
(97.85, 109.30)
(92.96, 104.20)
(56.81, 65.73)
(32.04, 96.06)
20
40
60
80
358
F20
F50
F100
F200
F500
Fig. 14.7. Posterior estimates of fake rates for different bill denominations. Vertical
lines correspond to 95% credible intervals and big dots to posterior means. Black (left)
lines are obtained with independent priors and red (right) lines with the hierarchical
prior.
of information, but at the same time respect the differences. An advantage of using
a hierarchical prior is that the mean parameter of the coefficients, 0 , concentrates
the population information coming from all parameters i . The posterior estimate of
= e0 , the population counterfeit rate, is reported in the last row of Table 14.9.
This estimate is compared with that obtained from considering that all observations
come from the same model; that is, yij Poisson() with log() = and prior
N(0, 0.001). The estimate of from this latter model is also included in the last
row in Table 14.9 under the column of Independent (prior). These two estimates show
great differences. The model that assumes that all observations come from the same
model with a single rate produces an interval estimate that is very narrow showing
an enormous precision, whereas the interval estimate obtained with the hierarchical
model acknowledges the uncertainty coming from the different denomination rates
i producing an overall counterfeit rate estimate for that is more realistic with a
lot less precision. This effect can better be appreciated in the last pair of intervals in
Figure 14.7.
14.5 Nonparametric Regression
14.5.1 Bayesian Nonparametric Ideas
The concepts of parametric and nonparametric statistics refer to assumptions that
are placed on the distribution of the available observations. One might assume that
a particular dataset was generated from a normal distribution with unknown mean
and precision. This is a parametric assumption since the N(, ) defines a parametric
family. In general, a parametric assumption would mean that the dataset is assumed to
359
0.0
0.2
Density
0.4
0.6
0.8
1.0
360
0
y
Fig. 14.8. Five realizations (black thin lines) of a finite PT with M = 2 levels and
centered on a N(0, 1) density (red thick line), with a = 1.
i | f f , f PT (
, A),
(14.4)
with f0 = N(0, ), mj = am2 , and a > 0 (see Appendix 14.6 for details). A further adjustment needs to be made. Since E(f ) = f0 this implies that i N(0, )
marginally (on average). That is, E(i ) = 0 not always, but only on average. We
can force the Polya tree to be centered at zero always (with probability 1) by fixing the first partition of the tree to be B11 = (, 0] and B12 = (0, ), and taking
11 = 12 = 1/2 with probability 1. This implies that the median of the random density f of each i is zero. This is verified by noting that the median of i is zero
iff P(i 0) = P(i B11 ) = 11 = 1/2. Therefore, the semi-parametric regression
model (14.4) is not a mean regression model but a median regression model since xi
becomes the median of the response variable yi .
Model (14.4) has two unknown quantities: and f . Our prior knowledge on f has
been placed through the Polya tree, so we also require a prior distribution for . The
common assumption is to take j N(b0 , t0 ), independently for j = 1, . . . , p 1,
as in most (generalized) regression models. The likelihood function for this semiparametric model is a function of the prior parameters and as follows:
lik(, | y, x) =
n
"
i=1
f (yi |xi ) =
n
"
i=1
f (yi xi ) =
n "
"
i=1 m
m,ji ,
361
with i = yi xi . Posterior distributions for (, f ) will be characterized conditionally. f |, y is another Polya tree with the distribution of the parameters updated with
i s as observations. |f , y is just proportional to the product of the likelihood and
the prior, and Metropolis-Hastings steps will be required for sampling from it. More
details on this semi-parametric model can be found in Walker and Mallick (1999).
Fortunately, posterior inference with this semi-parametric model is implemented in
the function PTlm from the R library DPpackage.
Example 14.5. Consider the insurance dataset described in Example 14.1. This
dataset consisted of claim amounts yi , premiums written xi , and class indicators
zj i , for j = 1, . . . , 7 sectors and i = 1, . . . , 228 observations. In Example 14.1 a
linear model in the logarithmic scale was suggested to describe severity amounts in
terms of the premiums written by sector. Let us consider the same linear predictor, but
instead of assuming normality of the errors (responses) we will consider nonparamet
ric errors with a Polya tree prior. The new model becomes yi = 1 + 7j =2 j zj i +
1 log(xi ) + 7j =2 j log(xj ) zj i + i , with i |f f and f PT (
, A).
The Polya tree is centered on f0 = N(0, ), as in the parametric model, with
mj = am2 . We took a = 1 and assigned a prior to the error precision
gamma(0.01, 0.01). It is worth mentioning that this latter prior induces a different
specification of the Polya tree partitions
= {Bmj } for every value of , producing a
mixing over the partitions and thus implying smoothed paths of the tree. We specify a
finite tree with a number of partition levels M = 6. For the model coefficients we use
the same priors as in Example 14.1, that is, j N(0, 0.001) and j N(0, 0.001),
for j = 1, . . . , 7. The specifications for implementing this model in R with the use of
the DPpackage library are presented in Table 14.10.
Remember that the semi-parametric regression model just defined is a median
regression model with an enhanced flexibility in the specification of the errors. Posterior estimates of model coefficients are included in the last two columns of Table 14.2.
The point estimates of all coefficients are numerically different from those obtained
with the parametric model, but only few are statistically different. Estimates of 1
and 7 present intervals that do not intercept between the parametric and the nonparametric scenarios, thus implying a difference in the slope relationships between
log(yi ) and log(xi ) for insurance sectors ACC and HEA comparing parametric and
nonparametric fit. Figure 14.9 compares the parametric and the nonparametric fittings
for the seven sectors in different panels. The major differences among the two fittings
are in the first two sectors ACC and AGR. In both cases there is an extreme observation that pulls the parametric fit, whereas the nonparametric model is less sensitive
to extreme observations, producing a more realistic fit consistent with nonextreme
observations.
362
0 1 2 3 4 5
4 2 0 2 4 6 8
MED
8
6
log(y)
4
2
log(y)
AUT
4 5 6 7 8 9
log(y)
AGR
1 0 1 2 3 4
log(y)
0 1 2 3 4 5 6
ACC
log(x)
log(x)
FIR
LIA
HEA
3 4 5 6 7 8 9
log(x)
log(x)
log(x)
log(y)
2 3 4 5 6 7 8
2 0
log(y)
2 3 4 5 6 7 8
0 1 2 3 4 5 6
log(y)
1 2 3 4 5 6 7
log(x)
4 2 0
log(x)
363
IR
B11
B12
B21
B31
B22
B32
B33
B23
B34 B35
B24
B36
B37
B38
At level two (m = 2) we have a partition of 22 = 4 elements B21 , B22 , B23 , and B24
such that (B21 , B22 ) are a partition of B11 and (B23 , B24 ) are a partition of B12 . In
general, at level m, Bmj is partitioned into (Bm+1,2j 1 , Bm+1,2j ) at level m + 1 with
Bm+1,2j 1 Bm+1,2j = . Figure 14.10 presents a diagram of these nested partitions
for levels m = 1, 2, 3.
Let = {mj } be a set of parameters such that each mj is associated with the set
Bmj . The parameter mj determines the conditional probability of a random variable
Y being in the set Bmj given that it belongs to the father, Bm,(j +1)/2 if j is odd, or
Bm,j/2 if j is even. For example, 21 = P(Y B21 |Y B11 ). Since the two subsets of
a father set form a partition of the set, the conditional probabilities must sum to one.
In the example, 21 + 22 = 1, where 22 = P(Y B22 |Y B11 ). In general m,2j =
1 m,2j 1 for j = 1, . . . , 2m1 . Therefore, for the sets at level m, the probability of
Y belonging to the set Bmj is just the product of all conditional probabilities mj , one
for each level, to where the set Bmj belongs. In notation,
P(Y Bmj ) =
m
"
mk+1,r(mk+1) ,
k=1
364
References
365
References
Bailey, A. (1950). Credibility procedures: Laplaces generalization of Bayes rule and the
combination of collateral knowledge with observed data. Proceedings of the Casualty
Actuarial Society 37, 723.
Banerjee, S., B. Carlin, and A. Gelfand (2004). Hierarchical Modeling and Analysis for
Spatial Data. Chapman and Hall, Boca Raton.
Bellhouse, D. (2004). The Reverend Thomas Bayes, frs: A biography to celebrate the
tercentenary of his birth. Statistical Science 19(1), 343.
Berger, J. (2006). The case for objective Bayesian analysis. Bayesian Analysis 1,
385402.
Bernardi, M., A. Maruotti, and L. Petrella (2012). Skew mixture models for loss distributions:
A Bayesian approach. Insurance: Mathematics and Economics 51(3), 617623.
Bernardo, J. and A. Smith (2000). Bayesian Theory. Wiley, New York.
Buhlmann, H. (1967). Experience rating and probability. ASTIN Bulletin 4, 199207.
Cabras, S. and M. E. Castellanos (2011). A Bayesian approach for estimating extreme
quantiles under a semiparametric mixture model. ASTIN Bulletin 41(1), 87106.
Cairns, A. J., D. D. K. Blake, G. D. Coughlan, and M. Khalaf-Allah (2011). Bayesian
stochastic mortality modelling for two populations. ASTIN Bulletin 41(1), 2959.
Ferguson, T. (1973). A Bayesian analysis of some nonparametric problems. Annals of
Statistics 1, 209230.
Ferguson, T. (1974). Prior distributions on spaces of probability measures. Annals of
Statistics 2, 615629.
Gelman, A., J. Carlin, H. Stern, and D. Rubin (2004). Bayesian Data Analysis. Chapman and
Hall, Boca Raton.
Gelman, A., A. Jakulin, M. Grazia-Pittau, and Y.-S. Su (2008). A weakly informative default
prior distribution for logistic and other regression models. Annals of Applied
Statistics 2, 13601383.
Hanson, T. and W. Johnson (2002). Modeling regression error with a mixture of Polya trees.
Journal of the American Statistical Association 97, 10201033.
Klugman, S. (1992). Bayesian Statistics in Actuarial Science: With Emphasis on Credibility.
Huebner International Series on Risk, Insurance, and Economic Security. Kluwer
Academic Publishers, Boston.
Landriault, D., C. Lemieux, and G. Willmot (2012). An adaptive premium policy with a
Bayesian motivation in the classical risk model. Insurance: Mathematics and
Economics 51(2), 370378.
Lavine, M. (1992). Some aspects of Polya tree distributions for statistical modelling. Annals
of Statistics 20, 12221235.
Lundberg, O. (1940). On Random Processes and Their Application to Sickness and Accident
Statistics. Almqvist and Wiksells, Uppsala.
Makov, U. E. (2001). Principal applications of Bayesian methods in actuarial science: A
perspective. North American Actuarial Journal 5(4), 5373.
Makov, U. E., A. F. M. Smith, and Y.-H. Liu (1996). Bayesian methods in actuarial science.
Journal of the Royal Statistical Society. Series D (The Statistician) 45(4), 503515.
Scollnik, D. (2001). Actuarial modeling with MCMC and BUGS. North American Actuarial
Journal 5(4), 96125.
Shi, P., S. Basu, and G. Meyers (2012). A Bayesian log-normal model for multivariate loss
reserving. North American Actuarial Journal 16(1), 2951.
Spiegelhalter, D., N. Best, B. Carlin, and A. van der Linde (2002). Bayesian measures of
model complexity and fit (with discussion). Journal of the Royal Statistical Society,
Series B 64, 583639.
366
15
Generalized Additive Models and
Nonparametric Regression
Patrick L. Brockett, Shuo-Li Chuang, and Utai Pitaktong
Chapter Preview. Generalized additive models (GAMs) provide a further generalization of both linear regression and generalized linear models (GLM) by allowing the
relationship between the response variable y and the individual predictor variables xj
to be an additive but not necessarily a monomial function of the predictor variables xj .
Also, as with the GLM, a nonlinear link function can connect the additive concatenation of the nonlinear functions of the predictors to the mean of the response variable,
giving flexibility in distribution form, as discussed in Chapter 5. The key factors
in creating the GAM are the determination and construction of the functions of the
predictor variables (called smoothers). Different methods of fit and functional forms
for the smoothers are discussed. The GAM can be considered as more data driven (to
determine the smoothers) than model driven (the additive monomial functional form
assumption in linear regression and GLM).
367
368
Thus, the i th observation (yi , xi1 , xi2 , . . . , xik ) is assumed to follow the parametric
relationship
yi = 0 + 1 xi1 + 2 xi2 + + k xik + i ,
where the error terms i are independent and identically distributed (i.i.d.) with E(i ) =
0 and Var(i ) = 2 . Accordingly, the mean value of y satisfies y = E[y|x1 , . . . , xk ] =
0 + 1 x1 + 2 x2 + + k xk . Estimation of the parameters is performed by least
squares. For inferential purposes, the i variates are usually assumed to be i.i.d. normal
random variables, so y also becomes normal (conditional on the observed values of
the xi s). Normality is not needed for prediction, but distributions are needed for
inference.
Linear regression is widely used because it provides a simple, easy linear way to
explain and interpret the relationship, and because the contribution of the predictor
variable is crystallized into one number, namely the predictor variables coefficient. Under the usual normality assumption, the sampling distribution of estimates
of the coefficient can be determined, which makes inference concerning the significance of the predictor variable easy to assess. Moreover, the relative predictive
strength among the k predictors can be examined and compared by simply looking
at the standardized magnitudes of their coefficients j . Using such an equation for
prediction is also simple. If we are given a new data point with predictor variables
{x1,new , x2,new , . . . , xk,new }, the prediction of ynew is simply the expected value or mean
of y from the above equation, namely ynew = y = 0 + 1 x1,new + 2 x2,new + +
k xk,new.
That simple linear model can be relaxed in several ways. First one can preserve the additive linear parametric monomial structure for the contribution of the
xi s, but instead of predicting the mean response y of the dependent variable
y, one predicts
a function of the mean of y in a linear and additive fashion (i.e.,
g(E[y]) = g y = 0 + 1 x1 + 2 x2 + + k xk ). Chapter 5 on generalized linear models (GLMs) addressed those problems wherein the response variable y, does
not necessarily have normal distribution and
mean of y is not a simple lin
the
ear function of the x variables, but rather g y = 0 + 1 x1 + 2 x2 + + k xk
or, stated differently, y = g 1 ( 0 + 1 x1 + 2 x2 + + k xk ). The function g in
this relationship is called the link function. Consequently, in the GLM model, the
mean or expected value of y is related to the xs only as a function g 1 () where
(x) = 0 + 1 xi1 + 2 xi2 + + k xik (here (x) is called the systematic component). It follows that the impact of x exhibits itself through the linear concatenation
, and inference can be pursued under the statistical assumption that y belongs to an
exponential family of distributions. The process is still parametric, but is more flexible
15.1 Motivation for Generalized Additive Models and Nonparametric Regression 369
than simple linear regression. Instead of least squares estimation, maximum likelihood
estimation is used once the exponential family has been specified.
A second way to relax the assumptions in the simple linear model is to relax
the parametric structure on the right hand side of the regression equation (i.e.,
replace the systematic component (x) by a more general function of the xj s).
This leads to nonparametric regression. Nonparametric regression poses the model
y = F (x1 , . . . , xk ) + where the problem is to find a multivariate function F that
appropriately fits the collection of given observations on y and x1 , . . . , xk , neither
overfitting nor oversmoothing the relationship. In its generality, this model can be
very difficult to address for more than a single predictor variable, and it requires
much larger data samples as the number of predictor variables increases (the curse of
dimensionality). Hence, more structure is usually assumed. A compromise between
the fully general nonparametric regression model y = F (x1 , . . . , xk ) + and the linear parametric model y = 0 + 1 x1 + 2 x2 + + k xk + is to fit a model that
retains the additive structure but does not impose the strict monomial linear structure
j xj term accounting for the contribution of the j th variable. The impact of xj on y
is now a nonparametric function of xj . Replacing j xj by a more general function
fj (xj ), which is then summed over the predictor variables, provides a more structured
form of nonparametric regression that is referred to as the additive model (discussed
subsequently)
y=+
k
fj xj + ,
(15.1)
j =1
where the error is independent of the xj s, E() = 0, Var() = 2 , and the functions
fj are constructed to appropriately fit the relationship. For example, the impact of
driver age x on accident claim size y in automobile insurance is not linear, but rather
is a more U-shaped function f (x), with higher claims at very young and at very old
ages, and smaller claims at intermediate ages. Additive models seek to determine the
appropriate functions fj from the data.1 To avoid identifiability problems in (15.1)
associated with constant terms in the fj confounding the estimation of , we center
the fj and assume E[fj (Xj )] = 0 so E[y] = .
The final generalization we pursue is to introduce a link function into (15.1), and
we then obtain what is known as the generalized additive model (GAM). The GAM
has the same relationship to the additive model as the generalized linear model (GLM)
has to the simple linear model. As does the GLM, it allows additional flexibility in the
1
These structure additive models can be extended to additively incorporate additional functions fij (xi , xj ) into the
regression equation (15.1) if there is significant joint influence of (xi , xj ) on y. We do not pursue this extension
here.
370
choice of y (such as a binary variable) via the link function as described in Chapter 5,
as well as flexibility to have highly nonlinear relationships between the predictor
variables and the dependent variable as in additive models.
A brief synopsis of this chapter is as follows. In the next section we introduce the
additive model as a structured form of nonparametric regression and discuss methods
for determining the functions fj . This is done first for a single x and y (k = 1 in (15.1),
or general univariate nonparametric regression); subsequently we discuss the more
general situation of k > 1 in (15.1). The functions in (15.1) are called smoothers, and
several different possible types are discussed.
Next we discuss algorithms for fitting the model and the trade-off between fit
and smoothness. After the discussion of the additive model, we introduce a link
function analogous to the GLM link function so as to obtain the GAM formulation.
Fit, smoothness, and inference are discussed. We illustrate the techniques along the
way with an example involving the prediction of coronary heart disease, an important
consideration in life insurance.
15.2 Additive Models for Nonparametric Regression
The additive nonparametric regression model (or simply additive model) is defined
to be
y= +
k
fj xj + ,
(15.2)
j =1
where is random error and the functions are not determined ahead of time, but
rather are flexible and are sought to be defined to fit the relationship. For simplicity,
usually each smoother fj is selected as a univariate function.2 The individual functions
fj in the additive model can be thought of as analogous to the coefficients in simple
regression, and one encounters all the interpretation pitfalls found in linear regression;
sometimes the interpretation difficulties are more severe (Hastie and Tibshirani 1999).
The smoothers are concatenated in an additive manner in describing the overall
relationship of y to the xs in the additive model. However, actually writing down a
closed-form mathematical formula for the relationship between x and y is infeasible
because it may be defined by segment, so graphical representations are often used for
displaying the estimated fj s.
The individual functions fj in (15.2) are the building blocks that attempt to
smoothly capture (possibly nonlinear) trends and patterns between the response
2
Nonparametric regression (curve fitting of the response surface) can require substantial data and is hard to interpret
and to present graphically in higher dimensions. The additive model is a compromise between relaxing the linearity
assumption in multivariate regression and unrestricted curve fitting in general high-dimensional nonparametric
regression.
371
variable y and the predictor variables xj . If the fj s are globally polynomials, then
(15.2) is polynomial regression, and the problem becomes a parametric optimization
in the polynomial coefficients. In general, however, the structural relationship between
y and fj (xj ) may change as the values of xj change over its range (i.e., the same
polynomial or Box-Cox transformation may not fit the relationship over the entire
range of x values). Therefore, in the spline regression approach to the additive model
the estimate of fj is defined in a piecewise fashion in local neighborhoods of the
xj values. Since the relationship between y and fj (xj ) is assumed to be smooth and
continuous, the functions fj are called smoothers.
Three main decisions involving smoothers are the choice of the type or class
of smoothers, the choice of the size of the local neighborhoods within which the
relationship is fit, and the selection of how smooth the globally piecewise patched
together function will be. Smoothers basically differ on the way they summarize the
trend or average the response y within the area of fit. The size of the neighborhood
regions used for fitting involves a fundamental trade-off between the bias and variance.
Large neighborhoods tend to mitigate the variance (be smoother), whereas small
regions keep the bias low (more closely reproduce the observed data).
15.2.1 Elementary Smoothers: Univariate Nonparametric Regression Case
The additive model does not start with a specific global parametric form for variables
over the entire estimation interval (as in linear regression), but instead it utilizes some
smooth functions to form the estimation. We observe the data and let the data show
what smooth functions we should use, since no parametric forms will be postulated
a priori. For exposition purposes we discuss the process of finding the appropriate
function f in the univariate case first (i.e., one y and one x with y = + f (x) +
where is the error term) and then discuss multivariate extensions.
A scatterplot is a simple but powerful diagrammatic tool used to demonstrate the
relationship between two variables. After observing the data, however, we may want
to develop a line or curve, f , through the scatterplot that parsimoniously illustrates
the pattern between the variables represented in the scatterplot. Intuitively, scatterplot
smoothers (such as the running average and the running line smoother, to be discussed
subsequently) are candidates for numerically obtaining an estimate of this f .
The simplest smoother when x is a categorical variable is to take an average of y
for each distinct value of x. If we apply this type of smoother to a predictor x, the
smoother would be the average value of y for each category of x, and the plot of x
versus y would look piecewise constant (like a bar chart). A bin smoother extends
this idea to noncategorical predictors by partitioning the values of x into a finite
number of bins or neighborhoods, and then assigning the average value of y to each
bin. The resulting function again looks like piecewise constant, and the function f so
372
developed does not provide a very smooth estimator because there will be jumps at
each cut point (bin boundary).
Running average smoothers (also called moving mean smoothers) attempt to create
a smoother estimation by fading out the abrupt transition at the cut points between
bins by utilizing overlapping regions in the estimation. Suppose all the values of x are
sorted and there are no replications in the data. One method for determining a region
for estimating y at the predictor point x0 is to define a neighborhood that covers the
target data point x0 together with a chosen number k of points to the left of x0 and k
points to the right of x0 . The value of this smoothing function is the mean of y over
the neighborhood region that centers at x0 . The choice of the number of points to the
right and left of x0 controls the responsiveness of the running average smoother to the
data. If k = 0 points are to the right and left, you have a very rough estimator since in
this case you are only considering the single-point neighborhood (y0 , x0 ). You then
obtain a connect-the-dots estimator, which perfectly reproduces the observed data,
but is also perfectly useless since it is reproducing the noise as well as the relationship
in the data that you wish to uncover. If you take k = n/2 points and include all the
data in the estimate (making the neighborhood consist of the entire range of x values),
then you have smoothed the response so much that a single mean is produced that may
not reflect any actual local variation present in the response of y to x. A compromise
is to define the neighborhoods by taking a proportion h (between zero and one) of
the n data points for each neighborhood. The number h is called the span (it is also
called the bandwidth), and it controls the smoothness. The neighborhood is defined
data points to the left and right of each predictor point. A
by including k = [hn]1
2
larger h will introduce more data for each neighborhood, and therefore the resulting
estimation will be smoother.
There are two problems with the running average smoother: first, at the endpoints
of the interval, the neighborhoods are not symmetrical so there is estimation bias
at the end points, and second, in fact, this smoother is often not very smooth (see
Figure 15.1). Thus, we are led to another neighborhood based smoother around x0
called the running line smoother. The running line is a variation of the previous
approach, but instead of averaging the response variable y over the local interval as in
the running average, it fits a regression line over the neighborhood region centered at
xo and then uses this regression line to predict a smoother value for that y observation
corresponding to x0 .
One of the reasons for the lack of smoothness in the running average smoother
is the all or nothing contribution of the various observations (xi , yi ), with each yi
corresponding to a point in the neighborhood of x0 contributing equally to the mean
estimate of y at x0 . If yi is from a point in the neighborhood (regardless of if its xi
is very close to x0 or is at the end of the neighborhood far from x0 ), it contributes
to the average, whereas it contributes nothing if it is outside the neighborhood. To
further smooth running average predictions, weights can be applied to the adjacent
373
data points. The highest weight is given to x0 itself, and, because we think points that
are close together are similar, the weights are then defined in a decreasing and smooth
fashion as they move farther away from x0 toward the boundaries or breakpoints
of the neighborhood that define the local bin. This approach is along the line of a
popular smoother called locally weighted running-line smoother, LOcal regrESSion,
or LOESS, implemented by Cleveland (1979).
A weight function is used in a Loess smoother to address the issue that, if we
calculate locally, those data that are distant from the designed target point x0 at which
we want to estimate the response y should make only a relatively small contribution
to the estimate (i.e., correct the all-or-nothing problem referred to earlier for running
average smoothers). The weight function in a Loess is designed to weight each
datum within the neighborhood according to the distance between the datum and
the designated point with higher weights for the points near the designated point
and lower weights for more distant points. Compared to scatterplot smoothers, a
Loess can provide a more general and smoother pattern. However, a Loess could
estimate an incorrect pattern due to the influential outliers. Accordingly, Cleveland
(1981) proposes a more sophisticated weight function to reduce the impacts from
the outliers, which he calls the Lowess smoother. A Lowess is a robust version of a
Loess, and it can be iteratively calculated in a short time, which is one reason that it is
popular. The Lowess iteratively reweights the Loess estimate to improve performance.
Lowess is now available for use in statistical packages (e.g., SAS and mgcv available
in R have these smoothers as options when performing additive modeling).
Example 15.1 (Predicting Coronary Heart Disease:). Examples of Running
Average, Running Line, Loess, and Lowess Estimators Coronary heart disease
and cancer combined constitute more than half of all deaths among people over the
age of 40 (Johnson et al. 2005, p. 109), and hence life insurers and annuity providers
are clearly interested in noninvasive methods of detecting coronary heart disease in
applicants for insurance. Accordingly, for a numerical illustration of the techniques
described earlier we use a heart disease dataset from the University of California at
Irvine Machine Learning Repository. The data are available at http://archive.
ics.uci.edu/ml/datasets/Heart+Disease. The variables we use are
shown in Table 15.1. In this example we examine the different types of smoothers
using the continuous variable predictors, with diagnosis of heart disease as the
response variable (the incorporation of the categorical predictors being considered
later).
Figure 15.1 shows the running average and running line smoother while Figure 15.2
shows the Loess and Lowess smoother for the continuous predictor variables. The
span h is set as 0.35. The purpose of this set of graphs is to help us observe
the relationship between the response variable and an explanatory variable and to
compare the different types of smoothers.
374
Table 15.1. Definitions for variables used in Heart Disease Prediction (from
University of California at Irvine Machine Learning Repository)
Variable
Name
Sex
Cp
Fbs
Res
Num
Age
Threstbps
Chol
Thalach
Definition
0: Female
1: Male
Chest pain type
1: typical angina
2: atypical angina
3: non-anginal pain
4: asymptomatic
Fasting blood sugar
0: > 120 mg/dl
1: 120 mg/dl
Resting electrocardiographic results
0: normal
1: having ST-T wave abnormality
2: showing probable or definite left ventricular hypertrophy by Estes criteria
Diagnosis of heart disease
0: < 50% diameter narrowing
1: > 50% diameter narrowing
Age in years
Resting blood pressure (in mm/Hg on admission to the hospital)
Serum cholesterol in mg/dl
Maximum heart rate achieved
The weight function W (z) = (1 |z|3 )3 for z < 1 (zero otherwise) is used in the Loess smoother for weighting the
impact of the points xi a distance |z| from the target point x0 , and this is not a probability density.
45
50
age
0.6
0.5
0.4
0.3
55
0.35
0.45
40
0.25
35
375
0.2
0.50
0.40
0.30
0.20
200
250
300
350
serum cholesterol (mg/dl)
110
130
150
170
maximum heart rate achieved
Fig. 15.1. Running average and running line smoothers for the continuous predictor
variables. The solid line is the running average, and the dashed line is the running
line smoother.
is one candidate, and the resulting estimator is usually called the Gaussian kernel
smoother. Running average smoothers can also be considered as kernel smoothers
that use a box kernel. Most often the kernel is chosen to have bounded support.4
Kernel smoothers use a weight function producing the value of the estimator at
point x0 as a weighted sum of the responses. The weight function has its peak at
the point of estimation x0 and gradually decreases as it moves away to both sides.
The weights are usually standardized so that they sum to unity. A simple weight
function for weighting the contribution
of the point xi to the estimation of y at the
i
, where K() is supported on [1,1] and peaks at
point x0 is, S (xi , x0 , ) = 1 K x0 x
5
zero.
4
The use of kernels with unbounded support can produce some global bias difficulty in the estimators, an issue that
can be restricted to boundary points only if one uses the bounded support kernels (cf., Eubank 1999, exercises,
p. 217). Thus, bounded supported kernels are generally preferred.
Note that any kernel with finite support can be scaled to have support on [1, 1].
40
50
376
60
100
200
300
400
140
160
180
200
500
600
100
120
age
80
100
120
140
160
180
Fig. 15.2. Loess and Lowess smoothers for the continuous predictor variables. The
solid line is the Lowess, and the dashed line is the Loess smoother.
The parameter is called the estimator bandwidth, and it determines the amount
of smoothing. It is allowed to be any non-negative number. If K(u) has support on
[1,1], then K(u/) has support on [, ]. This means the bandwidth determines
how far away observations are allowed to be from x0 and still contribute to the estimate
f (x0 ) of y at x0 . In this case the parameter is the scaling parameter of this density
function. The bandwidth also controls the pointiness of the weight function, affecting
the level of dependency of the estimator on the observations near the target x0 . A
small value of will cause the estimate to depend heavily on the data near x0 , and
hence will result in a rougher or more irregular nonsmooth estimate as one moves
377
M2
1
2
3
5
5
7
1
3
1
5
1
7
The quadratic kernel K(u) = 34 1 u2 I[1,1] (u) is also called the
Epanechnikov kernel (Epanechnikov 1969). It minimizes asymptotic mean squared error of the estimate.
from point to point along the x axis. In contrast, a larger will allow the estimate to
be averaged over a larger region and provide a smoother estimate. The choice of is
a decision to be made ahead of time, often by trial and error, using several values of
and choosing the estimator whose smoothness and performance are considered to
adequately describe the relationship.
The kernel K is usually selected to have the following properties:
K(u)du = 1,
(15.4)
uK(u)du = 0.
(15.5)
V =
1
1
u2 K (u) du = 0,
(15.6)
K(u)2 du < .
(15.7)
Conditions (15.6) and (15.7) are added to (15.4) and (15.5) to improve the asymptotic behavior of the estimate f of f as n and 0, with (15.6) used to control
the asymptotic bias and (15.7) used to make sure the estimator has finite asymptotic
variance (see Eubank 1999). A function K with all these properties combined is called
a second order kernel.
Examples of second-order kernels are shown in Table 15.2.
378
(15.8)
where is a fixed constant and represents twice differentiation. The first term
measures the closeness of fit of the predicted value f (xi ) to the observed value
yi , whereas the second term penalizes the fit for overall nonsmoothness. Of course
one could always fit a curve that exactly went through the data points (yi ,xi ), but
such a curve would be very irregular and rough due to random variations in the
data, and it would capture the idiosyncrasies of the particular sample as opposed
to the sought-after more general relationship between the response variable and the
predictor variable, which is assumed to be smooth and continuous. The relative tradeoff between fit and smoothness is governed by the parameter . It can be shown that
the general optimization problem of minimizing (15.8) for a fixed has an explicit
unique minimizer that is a natural cubic spline (cf., Hastie and Tibshirani 1999, p. 27;
Reinsch 1967, Wasserman 2006, Theorem 5.73; Wood 2006, p. 148).
6
A useful variation is natural cubic splines. These are cubic splines with the additional constraint that the function
must be linear after the knots at the endpoints to cope with the high variance near the boundaries. Because of data
limitations at the endpoints, kernel smoothers suffer from bias at the estimation interval endpoints and thus require
more consideration of how to handle the estimation at the endpoints. The forced linearity at the endpoints addresses
this concern.
379
While knowing that the minimizer of (15.8) is a natural cubic spline is useful,
it does not make it clear how to actually explicitly obtain the function f . We need
a computationally effective manner of doing this, and one such way is based on
Theorem 5.74 of Wasserman (2006).
Consider the function space consisting of all cubic splines with knots at specified points {t1 , . . . , tq } where a x1 t1 < t2 < < tq xn b. Wasserman
(2006; Theorem 5.74) proves that a basis set for this function space is the set of
functions b0 (x) = 1, b1 (x) = x, b2 (x) = x 2 , b3 (x) = x 3 , and bj (x) = (x tj 3 )3 for
j = 4, 6, . . . , q + 3. Since the functions {b0 , b1 , . . . , bq+3 } form a basis, any member
of this space can be written as a linear combination of these basis elements.7 In particular the cubic spline f that minimizes (15.8) is a linear combination of the above
basis elements, i.e.,
f (x) =
q+3
i bi (x).
(15.9)
i=0
For a fixed smoothness parameter , the determination of the cubic spline minimizing (15.8) involves the computation of the parameters in (15.9). Within this
basis-oriented approach, the fact that f () is linear in the basis element coefficients
parameters allows one to formulate the optimization in matrix notation and exploit
linear algebraic methodology to optimize, as shown in Wood (2006) and Wasserman
(2006). The formulation is as follows.
Since f () is linear in the parameters, the penalty term can be written as a
quadratic form in as
b
7 82
f (t) dt = A,
a
where A is a matrix of known coefficients. To see this, write Q (x) as the (row)
matrix of second derivatives of the basis functions {bj (x) : j = 0, . . . , q + 3} so
f (x) = Q (x) and the penalty term is A where
A=
Q (x) Q(x)dx,
(15.10)
>
(i.e., Ai,j is the number bi (x)bj (x)dx, which is computable since bj is at most
linear). Note also that the first two rows and columns of A are 0 because these involve
second derivatives of basis elements b0 and b1 , which are at most linear. A large
then puts most weight on these two bj s.
7
One can construct other bases for the space of cubic splines, and these could be used in (15.9) instead. Wood (2006,
chapter 4) discusses several bases and notes the differences that may arise in computation when using different
bases. Since our goal is to provide an intuitive theoretical framework, we do not delineate these other bases here.
380
The first term in (15.8) the squared error of the fit at the knot points can also be
written in matrix form.8 Therefore the penalized regression spline fitting of (15.8) is
to minimize over
y M2 + A,
where M is the matrix with (i, j ) entry mi,j = bi (xj ) (cf., Hastie and Tibshirani 1999;
Wood 2006). Given , the formal expression of the penalized least squares estimate
of is
1
(15.11)
= M M + A M y.
Since this is a regression, the corresponding hat matrix derived from normal
equations9 can be written (Wood 2006, p. 129) as
1
(15.12)
S = M M M + A M ,
and
y = M = S y.
For actual computation, using orthogonal matrix formulations offers greater stability in computation than the intuitive matrices M and A (Wood 2006, p. 129). Computer
programs such as those written in R (e.g., the program mgcv discussed in Wood 2006,
chapter 5) exploit these representations for computation.
In many applications the knot locations {t1 , . . . , tq } are not prespecified, but rather
are taken to be the data points x1 , . . . , xn themselves (plus the endpoints a and b).
This smoothing spline has a knot for each data point (and endpoints). Although in
doing this the number of internal knots is the same as the number of observations, the
smoothing penalty term in (15.8) will impose a penalty for some spline coefficients to
enforce smoothness. A high enough value of in the objective function (15.8) results
in a much smoother function than n degree of freedom might suggest. This occurs
since some coefficients will be shrunk in magnitude because of , so that the effective
degree of freedom in the fit is much less. If we let , then the fit produces
linear regression with constant slope.10 If we take zero, the change of slope is
less penalized so that smoothness hardly matters, and the fit produces more irregular
curves, approaching an interpolation fit.
8
9
10
Writing the least squares (first) component of (15.8) in matrix form is similar to ordinary least squares, but where
the observed independent variables are the basis elements evaluated at the knots.
The matrix that transforms the observed vector of ys into the predicted values of y s denoted by y is called the hat
matrix (since it transforms the y s to the y ) in ordinary least squares regression. Since y = M = S y, the matrix
S in (15.12) is also called the hat matrix in the additive model context. In ordinary regression, the hat matrix is
1
H = X X X X where X is the design matrix.
The first two rows and columns of A are 0, so a larger value of puts much more weight on the first two basis
elements, which are the linear part of the linear basis set.
381
MSE =
11
12
The hat matrix H in the linear model corresponds to the hat matrix S in (15.12) without the penalty part A.
Hastie and Tibshirani (1999) define two other estimates for the effective degrees of freedom based on analogy with the linear model, namely, n T r(2S S S ) and T r(S S ). For linear models and for smoothers
with S symmetric, these matrices are the same, and this includes linear and polynomial regression smoothers
and regression splines. They are also the same for running line smoothers (cf., Hastie and Tibshirani 1999,
p. 54).
382
The problem here is that f is unknown, so we cannot use this directly. Instead of
obtaining the estimation from MSE, we consider minimizing the predictive squared
error (PSE):
2
1
f (xi ) yi .
n i=1
n
PSE =
CV =
2
1
fi (xi ) f (xi ) 2 fi (xi ) f (xi ) i + i2 .
n i=1
n
%
n
2
1
E (CV ) = E
+ 2.
fi (xi ) f (xi )
n i=1
Intuitively, because for a large sample n the influence of any single point is not
large, we should have f (xi ) fi (xi ). Accordingly, it should be the case that
383
i=1 f (xi ) yi
n
(15.14)
GCV () =
1
2 .
T
r[I
S
]
n
GCV is computationally more tractable than ordinary CV, and it also has the
advantage of invariance (Wahba 1990). The estimator of that minimizes the GCV
is denoted by GCV , and in terms of the estimators loss performance, the choice of
GCV is asymptotically the best possible value of the smoothing parameter (Eubank
1999), p. 240; Li 1986). Computer packages allow the computation of GCV .
15.2.6 Additive Models: Multivariate Case and the Backfitting Algorithm
The previous discussion considered the situation of fitting the additive model in the
case of a single predictor variable x and a response variable y (essentially univariate
nonparametric regression). The additive model in (15.2), however, has k possible
predictors x1 , . . . , xk for the response variable y. Here we address the steps necessary
to incorporate k predictor variables (the backfitting algorithm).
13
Stone (1977) shows that cross-validation is asymptotically equivalent to the Akaike information criterion (AIC).
384
To motivate the backfitting algorithm, consider first the familiar linear model:
E (y|X = x) = +
k
j xj .
j =1
= j xj + E +
m Xm |Xj = xj .
m=j
Therefore, we find
j xj = E y|Xj = xj E +
m Xm |Xj = xj
m=j
= E y
m Xm |Xj = xj .
m=j
The term inside the expectation is the residual error in predicting y using all the
information about y that is contained in all the other predictor variables xm except
for xj . Therefore the idea is this. First run the partial regression predicting y using
all the other predictors except xj , and then calculate the residual of this regression
call it e. The resulting equation is E(e) = j xj so this suggests running a simple
regression with xj and the partial residual e to obtain an estimate of j . The problem,
however, is that initially we do not know any of the other m s to get e. Therefore,
instead we take an initial guess of the m s, m = j , and use these to get an estimate
j . We do this process for each of the m s in turn, which at the end provides an
updated estimate of each of the m s. The same process can then be repeated again,
but now using these updated estimates of the m coefficients to create a second round
of partial residual regressions, which are in turn regressed with the omitted variable to
get another updated estimate of the m s; this process is continued until convergence
takes place (i.e., the m s do not change from iteration to iteration). In this way we
end up having estimated all the m s in the original linear regression.
An example involving only two explanatory variables should clarify this process.
Suppose that we have a model
yi = + xi1 1 + xi2 2 + ei .
385
xi1
Let yi = yi y,
= xi1 x1 , and xi2
= xi2 x12 . ComWe observe that = y.
pute the partial residuals for x1 and x2 at iteration k as
(k)
= yi 2(k1) xi2
= ei + 1(k1) xi1
,
ei[1]
(k)
ei[2]
= yi 1(k1) xi1
= ei + 2(k1) xi2
.
(k)
We run a regression involving e[1]
and x1 to obtain 1(k) . Similarly, we can obtain 2(k)
with a second partial residual regression. Thus, we first start with an initial estimate
1(0) and 2(0) (e.g., we could just assume them to be zero), and then we run the
procedure to get updated estimates 1(1) and 2(1) . The same process can be done again,
this time starting from estimates 1(1) and 2(1) . This iteration continues until the 1(k)
and 2(k) estimates become stable (do not change from iteration to iteration).
To apply this same intuitive process to the fitting of the additive model, we observe
that the important assumption that was used in this process was the additivity, not
the fact that the predictor function involving xj was the monomial j xj . We can use
functions fj (xj ) in place of the terms j xj and can perform a similar iterative fit. We
show this now.
Suppose that we have n observations and a simple additive model (15.2). Each of the
individual functions fj in (15.2) can be initially estimated using a smoother (kernel,
cubic spline, etc.) as described earlier. Suppose Sj () is the smoother chosen for the j th
predictor variable in (15.2). It will provide a function fj for the general model (15.2).
Since the smoothers as discussed previously have a constant term in their formulation
and so does the model (15.2), to have identifiability (unique estimates), we center
variables by adding the constraint E[ kj =1 fj (xj )] = 0 or equivalently E(y) = when
we perform our estimations. We improve the resultant fit when the k smoothers are put
together, by iteratively considering the functions to be fit simultaneously in a manner
similar to what we did with linear regression.
Assuming that the additive model (15.2) is correct and y = + kj =1 fj (xj ) + e,
it follows that for any j ,
fm (xm ) |xj ].
(15.15)
fj xj = E[y
m=j
The term inside the expectation in (15.15) is the residual left in y (holding xj fixed)
after accounting for the other predictors fm (xm ), m = j . Thus, as before, the intuition
is that if for each m = j we can find good estimates fm fm , then we can use these
to get a good estimate fj of fj .
386
This suggests the following iterative algorithm for fitting the model (15.2) similar
to that described previously for linear regression. First, make initial estimates of the
fm , m = 1, . . . , k, m = j (say, linear regression estimates of xm on y, or even set
Use these to estimate fj . This new
them equal to zero) and set the initial = y.
estimate of fj can then be used in a similar manner to (15.15) to get new estimates
of the other fm . In each iterative step, we smooth out the response variable y after
removing the effect from other predictors:
fj = Sj y
fl (xl ) |xj .
l=j
n
fj xij ,
i=1
Centering:
1 (p)
(p)
(p)
fj = fj
f
(xij )
n i=1 j
n
1 (p)
f
(xij )
n i=1 j
C
C
C (p)
(p1) C
(3) Repeat step 2 until convergence: Cfj fj
C < for a specified accuracy
level .
n
(p) = (p1) +
This process is called the backfitting algorithm, and by performing the backfitting
algorithm we have accounted for the influence of the other predictors fm on fj ,
14
387
yi
i=1
k
42
fm (xim )
m=1
k
0
m
12
fm (t) dt,
(15.16)
m=1
where the fm s are twice continuous differentiable. Each m is used for penalizing
each smooth function fm . Again, the solution to (15.16) occurs when the smoother
fm for the predictor xm is a cubic spline (Schimek 2000, p. 283). Equation (15.16)
can be rewritten as
$
y
k
% $
fm
m=1
k
%
fm +
m=1
k
m fm Am fm ,
(15.17)
m=1
fj = I + j Aj
1
m=j
fm , j = 1, . . . , k.
(15.18)
388
1
Utilizing the individual smoothers (hat matrices), Sj = I + j Aj , j =
1, . . . , k, and rearranging terms, we may rewrite the entire set of equations (15.16) in
a giant (nk nk) matrix form as
f1
S1 y
I S1 S1 S1
S2 I S2 S2 f2 S2 y
(15.19)
..
..
.. . .
.. .. = .. .
.
. .
.
.
.
.
Sk Sk Sk
fk
Sk y
(15.20)
Solving (15.19) or (15.20) yields the desired smoothers for (15.2). According to
Hastie and Tibshirani (1999, p. 111), the second summation (the penalty term in
(15.17)) can be interpreted as a down-weighting of each component of fj . The downweighting is determined by the corresponding eigenvalue of that component and j .
In practice, the mgcv program written by Simon Wood and available in the computer
package R will supply the resulting estimated functions fm .
k
j =1
fj x j ,
(15.21)
389
where g() is a nonlinear differentiable and strictly monotonic link function. Since the
link function g() is invertible, the equation (15.21) can also be rewritten as
k
E [y] = g 1 +
(15.22)
fj x j .
j =1
Equation (15.22) looks very similar to (5.6) in the GLM framework, except the
systematic function inside the brackets in GLM is a linear model and in GAM it is an
additive model. The introduction of the additive nonparametric regression inside the
brackets in (15.22) complicates the estimation, but utilizing an estimation approach
similar to that used in GLM (Fisher scoring) and combining it with the estimation
method previously described for the additive model (backfitting), we are able to obtain
estimates in the GAM. The method is called local scoring and can be performed using
mgcv in R. Details and proofs can be found in Wood (2006) or Wasserman (2006).
(15.23)
where xj are the predictors and the j i are coefficients of the smoother fj that are to
be estimated. Once the basis is selected (and j is determined), as in Section 15.2.4,
for each j we have a matrix Mj so that fj = Mj j . For identifiability we center the
fj by making the sum (or mean) of the elements of fj be zero (and we incorporate
the constant term into our estimate of in the general formula). Thus we can create
390
1
j Aj ,
2 j
(15.24)
where Aj is the penalty matrix for predictor j discussed in Section 15.2.4. The penalty
is needed so that the fit is smooth (otherwise the fit could be a very rough interpolative
fit of a saturated model that maximizes the unpenalized likelihood). The vector of j s
controls the trade-off between the roughness of smoother functions and the goodness
of fit of the model. The optimal vector is solved for by maximizing lp () given a
set of j s. In practice, all this is done in the computer program in the packages that
fit GAMs.
Example 15.2 (GAM Fit to Coronary Heart Disease with Logit Link Function).
Returning to the example of coronary heart disease prediction (Example 15.1), we can
apply the GAM methodology to determine the influence of the continuous predictor
variables on the likelihood of an individual having coronary heart disease. In this
analysis we are interested in estimating the probability of an individual having heart
disease, rather than the zero-one response variable gathered in the data. Therefore,
as in Chapters 3 and 5, we introduce a link function (the logit link function g in
a GAM), which results in an ability to estimate this probability using the logistic
distribution. Refer to the discussion in Chapters 3 and 5 for details of this logit link
function.
Figure 15.3 displays the results of fitting a cubic spline function to the data and
plots the results against the log odd-ratio that would result from using a logit link
function g in a GAM. The response variable of interest in this case is the probability
of an individual having coronary heart disease given the set of predictor variables.
This probability is best displayed as a log odds-ratio, which
the left-hand side of a
is
y
GAM with the function g as the logit function g (y) = ln 1y . As can be seen, the
cubic spline fits exhibit the nonlinear nature of the fit to the log odds, and hence the
influence of each predictor variable on log odds is not a simple function.
391
1.2
log Odds
0.8 0.4
log Odds
1.0
0.5
0.0
0.0
35
40
45
age
50
55
60
150
100
120
140
160
resting blood pressure (mm/Hg)
2.5
log Odds
1.5 0.5
log Odds
1.5 1.0 0.5 0.0
0.5
0.5
30
80
Fig. 15.3. Fit of the continuous predictors variables to the log odds-ratio of coronary
heart disease using a GAM with the logit link function.
392
zero corresponding to the constant term and linear term in the basis representation.
A larger value of means a smoother function should be applied, and if is infinity,
then all the weight must go to the first two basis elements, implying the model is a
linear model and the fit is a GLM. Therefore, in testing between a GAM and a GLM
the hypothesis to be tested can be phrased as
H0 : = ,
Ha : < .
Wang (2011) presents a statistical test of this hypothesis using a ratio test statistic
based on using the smoothing parameter GCV that minimizes the GCV measure
previously discussed in 15.2.5:
tGCV =
GCV (GCV )
.
GCV ( = )
393
smoothness penalty j in the hat matrix, the trace can shrink and the effective degrees
of freedom can be less. This was the shrinkage of effective degrees of freedom alluded
to at the end of Section 15.2.4.
For a GAM, it still makes sense to consider deviance as the means to measure the
difference when comparing the two models, but now we need to define what we
mean by a saturated model, and we need to use the penalized log-likelihood function
lp instead of the log-likelihood function l. However, the deviance as defined earlier
with the penalized log-likelihood is not generally provable to be asymptotically 2 distributed, and its distribution theory is not yet well developed. In practice, however,
there is some empirical justification (Hastie and Tibshirani 1999) for informally using
the deviance test by referencing the 2 distribution with degrees of freedom equal to
the difference of the two models effective degrees of freedom.
394
Table 15.3. GAM Model 1 with Four Continuous Predictor Variables, g = Logit
Link Function, Predictor Variables Smoothed Individually
Estimate
Intercept
age
threstbps
chol
thalach
AIC
0.5975
edf
1
3.396
1
1
312.18
Std. Error
0.1421
Chi sq.
0.001
4.827
7.671
20.568
z Value
4.206
p-value
0.97285
0.34039
0.00561
0
p-Value
0
15.4 Conclusion
395
Table 15.4. GAM Model 2 with Two Significant Continuous Predictor Variables,
g = Logit Link Function, Two Predictor Variables Smoothed Together
Estimate
Intercept
chol, thalach
AIC
0.5903
edf
2
310.18
Std. Error
z Value
p-Value
0.1409
Chi.sq
33.59
4.189
p-value
0
results of GAM analysis in R. We used the R program mgcv designed for applying
GAMs that is available for download at http://cran.r-project.org/bin/
windows/base/.
The first model (GAM 1) uses the four continuous predictor variables discussed
earlier, and the logit link function (so the log odds-ratio of the likelihood of exhibiting
coronary heart disease is the dependent variable in the GAM model). The results
obtained using the R program package mgcv are presented in Table 15.3. As can be
seen, the variables age and threstbps are not significant.
Since the two variables age and threstbps were not significant, in the interests
of parsimony, another model was created with just the two significant continuous
predictor variables included. In Table 15.4 the two variables were smoothed. As can
be seen, by eliminating the nonsignificant variables we obtain a better model (lower
AIC) than in Table 15.3.
Since we also have other categorical variables available in the dataset, we can
include them in the model. The categorical variables do not need smoothing, so they
are just included in the GAM formulation. Table 15.5 show the GAM formulation
that has the two significant continuous predictor variables chol and thalach each
smoothed individually, as well as the categorical predictor variables.
Thus, the final GAM formulation involving the two significant continuous predictor
variables as well as the additional categorical variables produces the best fit (lowest
AIC).
15.4 Conclusion
Generalized additive models (GAMs) combine the flexibility of modeling the relationship between the response variable and the predictor variables exhibited in GLM
with the flexibility of allowing nonlinear relationships between individual predictors
and the response variable exhibited in nonparametric regression models. The additive
model, a compromise between the full flexibility of multidimensional nonparametric regression and fully parametric linear models, provides the basis for the GAM
396
Table 15.5. GAM Model 3 with Two Significant Continuous Predictor Variables
Smoothed Individually and Categorical Predictor Variables Included,
Link Function g = Logit Link
Estimate
Intercept
Sex
Cp
Fbs
Res
chol
thalach
AIC
4.9986
1.0624
1.1221
1.5154
0.2715
edf
1
1
257.69
Std. Error
z Value
p-Value
0.7391
0.407
0.1969
0.6398
0.3649
Chi.sq
4.23
6.802
6.763
2.61
5.699
2.369
0.744
p-value
0.0397
0.0091
0
0.00904
0
0.01785
0.45679
that provides enough structure to allow inference as well as improved model fit for
prediction. We have shown the intuition behind the models in this chapter, but the
computation of the GAM is substantial and requires care. Fortunately, such computational considerations have been incorporated into statistical packages such as SAS
and R, and the computing power now available has made such computations and
model fits available to researchers without trouble. Books such as Clark (2013) and
Wood (2006) can provide instruction for using the freeware in R and can provide
access to improved modeling of complex relationships between response variables
and predictors in actuarial science, insurance, and other domains.
References
Buja, A., T. Hastie, and R. Tibshirani (1989). Linear smoothers and additive models (with
discussion). Annals of Statistics 17(2), 453510.
Clark, M. (2013). Generalized Additive Models: Getting Started with Additive Models in R.
Center for Social Research, University of Notre Dame, Available at http://www3.
nd.edu/mclark19/learn/GAMS.pdf.
Cleveland, W. S. (1979). Robust locally weighted regression and smoothing scatterplots.
Journal of the American Statistical Association 74(368), 829836.
Cleveland, W. S. (1981). Lowess: A program for smoothing scatterplots by robust locally
weighted regression. The American Statistician 35(1), 54.
Craven, P. and G. Wahba (1979). Smoothing noisy data with spline functions: Estimating the
correct degree of smoothing by the method of generalized cross-validation. Numerische
Mathematik 31, 377403.
Epanechnikov, V. A. (1969). Nonparametric estimation of a multidimensional probability
density. Theory of Probability and its Application 14(1), 156161.
Eubank, R. L. (1999). Nonparametric Regression and Spline Smoothing (2nd ed.). Marcel
Dekker, New York.
References
397
Golub, G. H., M. Heath, and G. Wahba (1979). Generalized cross-validation as a method for
choosing a good ridge parameter. Technometrics 21(2), 215223.
Hastie, T. J. and R. J. Tibshirani (1999). Generalized Additive Models. Chapman & Hall,
Boca Raton, FL.
Johnson, M. L., V. L. Bengtson, P. G. Coleman, and T. B. Kirkwood (2005). The Cambridge
Handbook of Age and Ageing. Cambridge University Press, Cambridge.
Li, K.-C. (1986). Asymptotic optimality of CL and generalized cross-validation in ridge
regression with application to spline smoothing. Annals of Statistics 14(3), 11011112.
Liu, H. (2008). Generalized Additive Model. Department of Mathematics and Statistics,
University of Minnesota Duluth, Duluth, MN 55812.
Reinsch, C. H. (1967). Smoothing by spline functions. Numerische Mathematik 10(3),
177183.
Schimek, M. G. (2000). Smoothing and Regression: Approaches, Computation, and
Application. John Wiley & Sons, New York.
Stone, M. (1977). Asymptotics for and against cross-validation. Biometrika 64(1), 2935.
Wahba, G. (1990). Spline Models for Observational Data. Number 59. Siam.
Wang, Y. (2011). Smoothing Splines: Methods and Applications. Taylor & Francis Group,
New York.
Wasserman, L. (2006). All of Nonparametric Statistics, Volume 4. Springer, New York.
Wood, S. (2006). Generalized Additive Models: An Introduction with R. Chapman & Hall,
Boca Raton, FL.
16
Nonlinear Mixed Models
Katrien Antonio and Yanwei Zhang
Preview of the Chapter. We start with a discussion of model families for multilevel
data outside the Gaussian framework. We continue with generalized linear mixed
models (GLMMs), which enable generalized linear modeling with multilevel data.
The chapter includes highlights of estimation techniques for GLMMs in the frequentist
as well as Bayesian context. We continue with a discussion of nonlinear mixed models
(NLMMs). The chapter concludes with an extensive case study using a selection of R
packages for GLMMs.
16.1 Introduction
Chapter 8 (Section 3) motivates predictive modeling in actuarial science (and in many
other statistical disciplines) when data structures go beyond the cross-sectional design.
Mixed (or multilevel) models are statistical models suitable for the analysis of data
structured in nested (i.e., hierarchical) or non-nested (i.e., cross-classified, next to each
other, instead of hierachically nested) clusters or levels. Whereas the focus in Chapter
8 is on linear mixed models, we now extend the idea of mixed modeling to outcomes
with a distribution from the exponential family (as in Chapter 5 on generalized linear
models (GLMs)) and to mixed models that generalize the concept of linear predictors.
The first extension leads to the family of generalized linear mixed models (GLMMs),
and the latter creates nonlinear mixed models (NLMMs). The use of mixed models
for predictive modeling with multilevel data is motivated extensively in Chapter 8.
These motivations also apply here. We focus in this chapter on the formulation and
interpretation of mixed models for non-normal outcomes and nonlinear modeling
problems. Estimation, inference, and prediction with a range of numerical techniques
are discussed. Readers who are not interested in the technicalities of estimation
techniques can skip Sections 16.3.3.1 to 16.3.3.3.
398
399
400
(.) and c(.) are known functions; is the natural and the scale parameter. Using
vector notation the following relations hold:
= E[ y] = ( ) and Var[ y] = () = V (),
(16.2)
where derivatives are with respect to and V (.) is the so-called variance function.
The latter function captures the relationship between the mean and variance of y.
GLMs provide a way around transforming data by specifying a linear predictor for a
transformation of the mean regression parameters
g() = = X,
(16.3)
ui fU (ui ),
(16.4)
(16.6)
with ui the vector of random effects for cluster i, and xij and zij the p and q
dimensional vectors of known covariates corresponding with the fixed and random
401
(16.7)
ni
m + "
ij
"
1
2
ij e
1 e 2 2 ui dui .
(16.8)
L(, | y) =
2
y
!
ij
2
i=1
j =1
Straightforward calculations using mean and variance of a lognormal distribution
show
E(yij ) = E(E(yij |ui )) = E(exp (xij + ui ))
= exp (xij ) exp ( 2 /2)
(16.9)
and
Var(yij ) = Var(E(yij |ui )) + E(Var(yij |ui ))
= E(yij )(exp (xij )[exp (3 2 /2) exp ( 2 /2)] + 1),
(16.10)
and
Cov(yij , yik ) = Cov(E(yij |ui ), E(yik |ui )) + E(Cov(yij , yik |ui ))
= exp (xij ) exp (xik
)(exp (2 2 ) exp ( 2 )).
(j = k)
(16.11)
402
The expression in round parentheses in (16.10) is always greater than 1. Thus, although
yij |ui follows a regular Poisson distribution, the marginal distribution of yij is overdispersed. According to (16.11), due to the random intercept, observations on the same
subject are no longer independent, as is desirable for clustered data. Actuarial literature
on ratemaking (see, e.g., Denuit et al. 2007, and Antonio and Valdez 2012) often uses
2
a slightly modified version of the normality assumption, namely ui N( 2 , 2 ).
This leads to
E[yij ] = E[E[yij |ui ]] = exp (xi
2 2
+ )
2
2
= exp (xi ),
E[yij |ui ] = exp (xi + ui ).
(16.12)
In actuarial parlance, the so-called a priori premium (E[yij ]), specified as exp (xi ),
uses only a priori measurable risk factors (such as gender, age, and car capacity).
It is the marginal mean of yij and is therefore correct on average. The a posteriori
correction factor, exp (ui ), adjusts the a priori tariff based on the observed claim
history of the insured. We estimate this factor by predicting ui .
Example 16.2 (An Illustration of Shrinking). We consider a claim frequency model
using the auto claim data from Yip and Yau (2005), where we specify a log-linear
Poisson model with Jobclass as a random effect. In particular, we are interested
in how the estimate for each job class level differs between the mixed model and the
GLM where Jobclass enters as a factor fixed effect. This is the difference between
a partial pooling approach (with mixed models) and the no pooling approach (with
cluster-specific intercepts); see our discussion in Chapter 8. Figure 16.1 shows such
a comparison on the estimation of job class levels. The horizontal dotted line corresponds to the estimated intercept from the mixed model and represents the average
effect for all job categories because all the random effects have zero means. That
is, it is roughly the estimate when all job categories are pooled together. In contrast,
the estimates from the generalized linear model (the points) can be viewed as the
individual estimate for each job class level, ignoring the other levels indeed, fitting
a GLM with only the job class as a predictor is equivalent to fitting eight) separate
GLMs on each subset of data with a unique job class because of the orthogonal design
matrix corresponding to the job class. We see that the mixed model (the triangle)
shrinks the separate estimates from the GLM toward the pooled group-level estimate
across all the job classes. The shrinkage is most significant for Lawyer, Professional,
and Student. Therefore, the generalized linear mixed model captures the core insight
of the credibility models, where the estimates from the mixed models can be viewed
as the weighted average between the pooled group-level estimate and the separate
403
Student
Professional
Manager
Lawyer
HomeMaker
Doctor
Clerical
Blue Collar
Fig. 16.1. The job class estimates from the generalized linear model () and the
Poisson mixed models () in the auto insurance frequency model. The horizontal line
is the average estimate for all job classes, and the vertical lines show the uncertainty
intervals based on one standard error.
individual estimates. As a result, the mixed model produces less extreme estimates
while still accounting for the heterogeneity across the various levels.
16.3.3 Estimation
Using the model specifications in (16.4) it is straightforward to write down the likelihood of the corresponding GLMM:
(16.13)
L(, D| y) = fY |u ( y|u)fU (u)du,
where the integral goes over the random effects vector u (with covariance matrix D).
The presence of the integral in (16.13) hinders maximum likelihood estimation and
prohibits explicit expressions for estimators and predictors, such as those derived for
LMMs. Only so-called conjugate distributional specifications lead to a closed-form
solution in (16.13), a normal distribution for the response, combined with normally
distributed random effects (as with LMMs), being one example. More general model
assumptions require approximate techniques to estimate , D and predict the random effect for cluster i, ui . As in Molenberghs and Verbeke (2005) we distinguish
three approaches to tackle this problem: approximating the integrand, approximating the data, and approximating the integral (through numerical integration). Having
404
Pinheiro and Bates (2000), McCulloch and Searle (2001), (chapters 8 and 10) and
Tuerlinckx et al. (2006) as main references, we discuss next some highlights of
these methods. This discussion will help readers understand the differences among
different R packages available for data analysis with GLMMs (as demonstrated in
Section 16.6). Section 16.3.4 presents pros and cons of the techniques mentioned
in 16.3.3.1, 16.3.3.2, and 16.3.3.3, as well as references to other techniques (not
discussed here). We postpone a discussion of a Bayesian approach to Section 16.5.
16.3.3.1 Approximating the Likelihood: the Laplace Method
The Laplace method (see Tierny and Kadane 1986) approximates integrals of the form
(16.14)
eh(u) du.
for some function h of a q-dimensional vector u. The method relies on a second-order
+ (u u)
u),
h(u) h(u)
2
(16.15)
h(u)
|u=u = 0,
u
(16.16)
with
2
2
h(u) 2
= u
h(u)
u)
e d u eh(u)+
du.
(16.17)
Approximating the density of u with a multivariate Gaussian distribution
(h (u))
1 ) leads to
N (u,
21/2 h(u)
2
2
e .
(16.18)
eh(u) d u (2 )q/2 2h (u)
This technique is readily available to approximate the likelihood in a GLMM (see
Breslow and Clayton 1993, and McCulloch and Searle 2001, among other references):
= log fY |U ( y|u)fU (u)du
= log elog fY |U ( y|u)+log fU (u) du
(16.19)
= log eh(u) du,
405
q
2
log 2
(16.20)
where W and are diagonal matrices with elements [V (i )(g (i ))2 ]1 and g (i ),
respectively.1 Hereby g(i ) and V (i ) are the mean and variance of yi , conditional
on ui , as introduced in (16.2).
The matrix of second-order derivatives is (see (16.18))
2 h(u)
1
1 W
= Z W + Z
( y ) D1 .
u u
u
(16.21)
The random vector corresponding with the second term in this expression has expectation zero, with respect to fY |U ( y|u), and will be ignored. Therefore,
1
2 h(u)
=
Z W 1 Z + D1
u u
1
Z W Z D + I D1 .
=
(16.22)
yi
u
u
i u
i
1
1
1
(yi i )
=
z .
V (i ) g (i ) i
i
(16.24)
406
Therefore, the following set of equations has to be solved simultaneously with respect
to and u (using a numerical optimization method):
1
X W ( y ) = 0
1
Z W ( y ) = D1 u.
(16.25)
This set of equations also arises by jointly maximizing (with respect to and u)
1
log fY |U ( y|u) u D1 u,
2
(16.26)
(16.27)
(16.28)
(16.29)
with a diagonal matrix with entries g (i ). Calculating the variance of t is complicated because of the dependence of on (and therefore on the random vector u).
407
(16.30)
The working variate t approximately follows a linear mixed model (as in Chapter 8),
with design matrices X (fixed effects), Z (random effects), D the covariance matrix of
the random effects, and the covariance matrix of the error terms. In this LMM it is
straightforward to estimate , u, and the unknown variance components. Therefore,
the pseudo-likelihood algorithm goes as follows. Starting from initial estimates of
, u, and the variance components, the working variates in (16.29) are evaluated.
Consequently, using LMM methodology, updated estimates follow from (16.29) and
(16.30). These steps are repeated until there is convergence of the estimates.
16.3.3.3 Approximating the Integral: Numerical Integration Techniques
Approximating the integral in (16.13) with a so-called (adaptive) quadrature rule for
numerical integration is based on Liu and Pierce (1994). For ease of explanation we
consider next the case of a one-dimensional integral. The case with multidimensional
integrals is documented in Tuerlinckx et al. (2006).
Nonadaptive GaussHermite quadrature. Non-adaptive Gauss-Hermite quadrature approximates an integral of the form
+
h(z) exp (z2 )dz,
(16.31)
Q
wl h(zl ).
(16.32)
l=1
Q is the order of the approximation, the zl are the zeros of the Qth order Hermite
polynomial, and the wl are corresponding weights. The nodes (or quadrature points)
zl and the weights wl are tabulated in Abramowitz and Stegun (1972, p. 924). The
quadrature points used in (16.32) do not depend on h. Therefore, it is possible that
only a very few nodes lie in the region where most of the mass of h is, which would
lead to poor approximations.
Adaptive GaussHermite quadrature. With an adaptive Gauss-Hermite quadrature rule, the nodes are rescaled and shifted such that the integrand is sampled in a
408
suitable range.
> + Assume h(z)(z; 0, 1) is unimodal, and consider the numerical integration of h(z)(z; 0, 1)dz. Let and be
2 1
2
2
2
(16.33)
= mode [h(z)(z; 0, 1)] and = 2 ln (h(z)(z; 0, 1))2
z=
z
Acting as if h(z)(z; 0, 1) were a Gaussian density, and would be the mean
and variance of this density. The quadrature points in the adaptive procedure, zl" , are
centered at with spread determined by ,
namely
zl" = + 2 zl
(16.34)
> +
with (l = 1, . . . , Q). Now rewrite h(z)(z; 0, 1)dz as
+
h(z)(z; 0, 1)
(z; , )dz,
(16.35)
(z; , )
where (z; , ) is the Gaussian density function with mean and variance 2 . Using
simple manipulations it is easy to see that for a suitably regular function v,
$
%
+
+
1 z 2
2 1/2
dz
v(z)(z; , )dz =
v(z)(2 )
exp
2
+
v( + 2z)
exp z2 dz
=
Q
v( + 2zl )
wl .
(16.36)
l=1
instead of v(z) and replacing and with their estimates from
Using h(z)(z;0,1)
(z;,)
(16.33), the following quadrature formula results:
+
Q
h(z)(z; 0, 1)dz 2
wl exp (zl2 )(zl" ; 0, 1)h(zl" )
l=1
Q
wl" h(zl" ),
(16.37)
l=1
with adaptive weights wl" := 2 wl exp (zl2 )(zl" ; 0, 1). Expression (16.37) is an
adaptive Gauss-Hermite quadrature formula.
Link with Laplace approximation. We illustrate the connection between the
Laplace approximation (from Section 16.3.3.1) and the adaptive GaussHermite
quadrature with a single node. Indeed, when Q = 1 (i.e., the case of a single node),
z1 = 0 (from the Hermite polynomial) and w1 = 1. The corresponding adaptive
409
node and weight are z1" = and w1" = 2 (z1" ; 0, 1), respectively. The adaptive
GH quadrature formula then becomes
1/2
2
2
2 2
2 21/2
2
2 2
exp {log ((;
0, 1)h())},
(16.38)
2 z2 log (h(z)(z; 0, 1)) z= 2
where = z1" maximizes h(z)(z; 0, 1). This corresponds with the Laplace formula
from (16.18).
Adaptive GaussHermite quadrature for GLMMs. We describe the case of a
GLMM with a single, normally distributed random effect ui N(0, 2 ) for each
cluster i. The use of an adaptive GaussHermite quadrature with GLMMs starts from
determining the posterior mode of ui . Since this posterior distribution depends on
unknown fixed effects and variance parameters, we replace the unknown , , and
with their current estimates: (c) , (c) , and (c) . Using these current estimates u i
maximizes
f ( yi |ui )f (ui | (c) ),
(16.39)
(16.40)
ni
"
j =1
Q
l=1
wl"
ni
"
j =1
fY |U (yij |zl" ) ,
(16.42)
410
with adaptive weights wl" = 2i wl exp (zl2 )(zl" ; 0, 1) and zl" = i + 2i zl . In this
expression the linear predictor corresponding with fY |U (yij |i ) and fY |U (yij |zl" ),
respectively, is xij + i and xij + zl" . Multiplying (16.42) over all clusters i
leads to the total likelihood. Maximizing the latter over the fixed effects regression
parameters, the dispersion parameter, and the variance components leads to updated
parameter estimates (c+1) , (c+1) , and (c+1) . We predict the cluster-specific random
effects with the posterior modes from (16.39).
411
corresponding null hypothesis constrains the variance of the random effect to the
boundary of its parameter space. With respect to inference with (G)LMMs a Bayesian
analysis has some additional features (see Section 16.5 for discussion).
16.4 Nonlinear Mixed Models
LMMs and GLMMs model the mean (in LMMs) or a transformation of the (conditional) mean (in GLMMs) as linear in the fixed effects parameters and the random
effects u. Nonlinear mixed models (NLMM) extend the concept of linear predictors.
In an NLMM the conditional distribution of yij (being the j th response on cluster i),
given ui , belongs to the exponential family with mean structure
E[yij |ui ] = h(xij , , zij , ui ),
(16.43)
m
"
i=1
fi ( yi |, , u1 , . . . , um )
m
"
i=1
412
each parameter (vector), solves this issue and provides a way to circumvent otherwise
intractable calculations. The likelihood approximations discussed in Section 16.3.3 are
replaced in a Bayesian analysis with general MCMC methodology for sampling from
posterior distributions. This allows specification of more complex hierarchical model
structures, such as the spatial structures in Chapter 11 or the three-level count data
models in Antonio, Frees, and Valdez (2010). Moreover, the Bayesian methodology is
not limited to Gaussian random effects. For predictive modeling in actuarial science,
Bayesian statistics are particularly useful for simulation from the posterior (predictive)
distribution of quantities of interest, such as a policys random effect or the number
of claims in a future time period.
ij = exp (0 + 1 Yearij ).
(16.45)
413
1Q
Median
3Q
Max
-26.8194
-1.0449
0.2456
2.3197
18.1740
Coefficients:
Estimate Std. Error
z value Pr(>|z|)
(Intercept) -3.702172
0.008648 -428.105
<2e-16 ***
yearcentr
0.005098
0.0464 *
-0.010155
-1.992
--Signif. codes:
on 766
degrees of freedom
on 765
degrees of freedom
AIC: 14904
Number of Fisher Scoring iterations: 5
The estimate 0 for the intercept is 3.702 (with s.e. 0.00865) and 1 = 0.0102
(with s.e. 0.00510).
No pooling. We continue the analysis with a fixed effects Poisson model that specifies an occupation class specific intercept, say 0,i , for each of the 113 occupation
classes in the dataset, as well as a global fixed Year effect. The intercepts 0,i are
unknown, but fixed. We fit the model in R with the glm function, identifying the
occupation class as a factor variable:
Countij POI(Payrollij ij )
(16.46)
414
3Q
0.7186
Max
7.1909
Coefficients:
Estimate Std. Error
yearcentr
9.918e-03 5.157e-03
factor(riskclass)1
-2.578e+00 2.425e-01
factor(riskclass)2
-3.655e+00 4.082e-01
factor(riskclass)3
-3.683e+00 1.374e-01
factor(riskclass)4
-1.309e+01 2.103e+03
factor(riskclass)5
-2.737e+00 9.325e-02
...
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05
z value Pr(>|z|)
1.923 0.054448 .
-10.630 < 2e-16 ***
-8.952 < 2e-16 ***
-26.810 < 2e-16 ***
-0.006 0.995035
-29.347 < 2e-16 ***
. 0.1
on 767
on 636
degrees of freedom
degrees of freedom
Comparing the deviance of the complete and the no pooling model results in a
drop-in-deviance of 12,2701,192 = 11,078, with a difference in degrees of freedom of 129. In R the anova function is available for this comparison. With a pvalue of < 2.2 1016 the no pooling model outperforms the complete pooling
model.
415
> anova(fitglm.CP,fitglm.NP,test="Chisq")
Analysis of Deviance Table
Model 1: count ~ yearcentr
Model 2: count ~ 0 + yearcentr + factor(riskclass)
Resid. Df Resid. Dev Df Deviance Pr(>Chi)
1
765
12270.4
2
636
1192.9 129
11078 < 2.2e-16 ***
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1
1
GLMMs: Random intercepts, Laplace approximation with lmer. We investigate a Poisson regression model with random occupation class specific intercepts as
a meaningful alternative to the no pooling model. The model formulation is
Countij |ui,0 POI(Payrollij (ij |ui,0 ))
(16.47)
We first fit this random intercepts model with the lmer (or: glmer) function from the
R library lme4. By default lmer uses Laplace approximation (see Section 16.3.3.1)
to approximate the Poisson likelihood. Adaptive GaussHermite quadrature is also
available within the lme4 package (see infra).2
> hlm1 <- glmer(count~(1|riskclass)+yearcentr+offset(log(payroll)),
+
family=poisson(link="log"), data=wcFit)
> print(hlm1)
Generalized linear mixed model fit by the Laplace approximation
Formula: count ~ (1 | riskclass) + yearcentr + offset(log(payroll))
Data: wcFit
AIC BIC logLik deviance
1771 1785 -882.6
1765
Random effects:
Groups
Name
Variance Std.Dev.
riskclass (Intercept) 0.80475 0.89708
Number of obs: 767, groups: riskclass, 130
The glmmmML package is another R package for Laplace approximation and GaussHermite quadrature for Binomial
and Poisson random effects models; see the books website for sample code.
416
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -3.562125
0.083489 -42.67
<2e-16 ***
yearcentr
0.009730
0.005156
1.89
0.0592 .
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1
1
Correlation of Fixed Effects:
(Intr)
yearcentr -0.010
From the R output just given, we conclude that 0 = 3.5621 (with s.e. 0.0835),
1 = 0.00973 (with s.e. 0.00516), and u2 = 0.805 is the estimate of the variance of
random intercepts. Extracting the estimates of fixed and random effects, as well as
prediction intervals for the random intercepts, goes as follows.
417
We are now ready to compare model (16.46) (no pooling) to (16.47) (random intercepts). The left panel in Figure 16.2 shows the intercepts and corresponding error bars
from the no pooling model, together with one standard error, against the total size of
the occupation class (i.e., j Payrollij ). Figure 16.2 (right) shows the point predictions of the random intercepts. To create this plot we refit the random effects model and
3
Estimate (+/ s.e.)
3
0
3
6
9
4
4
log(Size)
12
0
3
6
9
4
4
log(Size)
Fig. 16.2. Point estimates for occupation class specific intercepts, plus/minus one
standard error. Results are from no pooling approach (left) and Poisson random
intercepts model (right). The dashed line is y = 3.702 (i.e., the overall intercept
from the complete pooling model).
12
418
do not include an intercept.3 The light grey dashed line in the Figure is y = 3.702,
the overall intercept from the complete pooling model.
As discussed in the introduction to Chapter 8 the no pooling model results in unreasonable estimates for certain occupation classes. The output printed next compares the
size, random intercept estimate, fixed intercept estimate, and corresponding standard
errors for a selection of occupation classes.
## occupation class 122 (our numbering)
# random intercept model
num size
lower estimate
upper logsize
122 33.32 -6.228015 -4.635567 -3.04312 3.506158
# no pooling model
num size
lower estimate
upper logsize
122 33.32 -854.0756 -17.96726 818.1411 3.506158
# data for this class (i.e. zero claims on 33.32 payroll total)
riskclass year count
payroll
122
1
0
3.28
122
2
0
5.69
122
3
0
4.51
122
4
0
4.80
122
5
0
9.07
122
6
0
5.97
## occupation class (our numbering)
# random intercepts model
num size
lower estimate
upper logsize
61 23.16 -3.840979 -2.955994 -2.071008 3.142427
# no pooling model
num size
lower estimate
upper logsize
61 23.16 -4.144029 -3.144028 -2.144028 3.142427
# data for this class (i.e. 1 claim on 23.26 payroll total)
riskclass year count payroll
61
1
0
3.12
61
2
0
3.68
61
3
0
3.76
61
4
0
3.83
61
5
1
4.99
61
6
0
3.78
## occupation class (our numbering)
# random intercepts model
num size
lower
estimate
upper
logsize
52
0.08
-3.5900335 -0.6187492
2.3525350
-2.525729
3
As explained in Chapter 8 on LMMs, lme4 evaluates the variance of the distribution of u| y, conditional on the
maximum likelihood estimates for unknown parameters.
419
# no pooling model
num size
lower
estimate
upper
logsize
52
0.08 -2117.1442456 -13.7817186
2089.5808085
-2.525729
# data for this class (i.e. 0 claims on 0.08 payroll total)
riskclass year count payroll
52
4
0
0.08
The parameter estimates and standard errors obtained with GH quadrature are very
close to those obtained with Laplace approximation (see the results of hlm1). Changing the number of quadrature points has a very minor impact on the results.
420
GLMMs: Random intercepts, approximating the data with glmmmPQL. To illustrate the approach from Section 16.3.3.2 (i.e., repetitive fits of a linear mixed model
to pseudo-data), the function glmmmPQL from library(MASS) is available in R.
For the Poisson random intercepts model from (16.47), convergence is reached in 10
iterations. Parameter estimates and corresponding standard errors are printed next.
They are different from but close to the results obtained with lme4. Based on the
discussion in Section 16.3.4, however, we prefer the use of adaptive GaussHermite
quadrature whenever possible for the model under consideration.
> library(MASS)
> PQL1 <- glmmPQL(count ~ yearcentr + offset(log(payroll)),
random = ~ 1 | riskclass,family = poisson, data = wcFit)
iteration 1
iteration 2
iteration 3
iteration 4
iteration 5
iteration 6
iteration 7
iteration 8
iteration 9
iteration 10
> summary(PQL1)
Linear mixed-effects model fit by maximum likelihood
Data: wcFit
AIC BIC logLik
NA NA
NA
Random effects:
Formula: ~1 | riskclass
(Intercept) Residual
StdDev:
0.9290198 1.50974
Variance function:
Structure: fixed weights
Formula: ~invwt
Fixed effects: count ~ yearcentr + offset(log(payroll))
Value Std.Error DF
t-value p-value
(Intercept) -3.489754 0.08911097 636 -39.16189 0.0000
yearcentr
0.009496 0.00656277 636
1.44688 0.1484
421
Correlation:
(Intr)
yearcentr -0.012
Standardized Within-Group Residuals:
Min
Q1
Med
Q3
Max
-2.5749914 -0.5294022 -0.1518360 0.4497736 12.6121268
Number of Observations: 767
Number of Groups: 130
where we use normal N(0, 104 ) priors for 0 and 1 , and a (0.01, 0.01) prior for
(u2 )1 . The posterior densities of these parameters are illustrated in Figure 16.3.
With respect to predictive modeling, a Bayesian approach is most useful, because
it provides the full predictive distribution of variables of interest (here: Counti7 ). We
illustrate this in Figure 16.4 for a selection of risk classes. Histograms are based on
422
60
3
2
1
0
40
20
0
3.8 3.6 3.4 3.2
0
0
0.4 0.6 0.8 1.0 1.2 1.4
2u
Fig. 16.3. Posterior simulations for parameters used in (16.47) (from left to right:
0 , 1 , and u2 ), workers compensation insurance (frequencies). Results are based
on two chains, 50,000 simulations each, a thinning factor of 5, and burn-in of 2,000
simulations.
50,000 simulations from the relevant predictive distribution (using model (16.47)).
For each risk class the observed number of claims is indicated, as well as the point
prediction obtained with a frequentist approach, using Laplace approximation from
(g)lmer.
0.06
0.10
0.04
0.05
0.02
0.00
0.00
0 5 10 15 20 25
Occupation class = 64
0.06
0
20 30 40 50 60 70
Occupation class = 18
0 1 2 3 4 5
Occupation class = 68
0.06
0.04
0.04
0.02
0.02
0.00
0.00
10 20 30 40 50 60
10 20 30 40 50 60 70
Occupation class = 87
Occupation class = 110
Fig. 16.4. Posterior predictive simulations for the number of claims in year 7 for
a selection of risk classes. Simulations are based on Bayesian analysis of (16.47),
using two chains, 50,000 simulations each, a thinning factor of 5, and burn-in of
2,000 simulations. The dashed light grey line is the point prediction as obtained with
(g)lmer; the dashed light grey line is the observed number of claims.
References
423
References
Abramowitz, M. and I. Stegun (1972). Handbook of Mathematical Functions: With
Formulas, Graphs and Mathematical Tables. Dover, New York.
Antonio, K., E. Frees, and E. Valdez (2010). A multilevel analysis of intercompany claim
counts. ASTIN Bulletin: 40(1), 151177.
Antonio, K. and E. Valdez (2012). Statistical aspects of a priori and a posteriori risk
classification in insurance. Advances in Statistical Analysis 96(2), 187224.
Breslow, N. and D. Clayton (1993). Approximate inference in generalized linear mixed
models. Journal of the American Statistical Association 88(421), 925.
Breslow, N. and X. Lin (1995). Bias correction in generalized linear mixed models with a
single component of dispersion. Biometrika 82, 8191.
Denuit, M., X. Marechal, S. Pitrebois, and J.-F. Walhin (2007). Actuarial Modelling of Claim
Counts: Risk Classification, Credibility and Bonus-Malus Scales. Wiley, New York.
Haberman, S. and A. Renshaw (1996). Generalized linear models and actuarial science. The
Statistician 45(4), 407436.
Kim, Y., Y.-K. Choi, and S. Emery (2013). Logistic regression with multiple random effects:
A simulation study of estimation methods and statistical packages. The American
Statistician 67(3), 171182.
Liang, K. and S. Zeger (1986). Longitudinal data analysis using generalized linear models.
Biometrika 73(1), 1322.
Lin, X. and N. Breslow (1996). Analysis of correlated binomial data in logistic-normal
models. Journal of Statistical Computation and Simulation 55, 133146.
Liu, Q. and D. Pierce (1994). A note on Gauss-Hermite quadrature. Biometrika 81(3),
624629.
McCulloch, C. and S. Searle (2001). Generalized, Linear and Mixed Models. Wiley Series in
Probability and Statistics, Wiley, New York.
Molenberghs, G. and G. Verbeke (2005). Models for Discrete Longitudinal Data. Springer
Series in Statistics, Springer, New York.
Nelder, J. and R. Wedderburn (1972). Generalized linear models. Journal of the Royal
Statistical Society, Series A 135, 370384.
Pinheiro, J. and D. Bates (2000). Mixed Effects Models in S and S-Plus. Springer, New York.
Purcaru, O., M. Guillen, and M. Denuit (2004). Linear credibility models based on time
series for claim counts. Belgian Actuarial Bulletin 4(1), 6274.
Tierny, L. and J. Kadane (1986). Accurate approximations for posterior moments and
marginal densities. Journal of the American Statistical Association 81, 8286.
Tuerlinckx, F., F. Rijmen, G. Verbeke, and P. D. Boeck (2006). Statistical inference in
generalized linear mixed models: A review. British Journal of Mathematical and
Statistical Psychology 59, 225255.
Wolfinger, R. and M. OConnell (1993). Generalized linear mixed models: A
pseudo-likelihood approach. Journal of Statistical Computation and Simulation 48,
233243.
Yip, K. C. and K. K. Yau (2005). On modeling claim frequency data in general insurance
with extra zeros. Insurance: Mathematics and Economics 36(2), 153163.
Zhao, Y., J. Staudenmayer, B. Coull, and M. Wand (2006). General design Bayesian
generalized linear mixed models. Statistical Science 21, 3551.
Part IV
Longitudinal Modeling
17
Time Series Analysis
Piet de Jong
Chapter Preview. This chapter deals with the analysis of measurements over time,
called time series analysis. Examples of time series include inflation and unemployment indices, stock prices, currency cross rates, monthly sales, the quarterly number of
claims made to an insurance company, outstanding liabilities of a company over time,
internet traffic, temperature and rainfall, and the number of mortgage defaults. Time
series analysis aims to explain and model the relationship between values of the time
series at different points of time. Models include ARIMA, structural, and stochastic
volatility models and their extensions. The first two classes of models explain the level
and expected future level of a time series. The last class seeks to model the change
over time in variability or volatility of a time series. Time series analysis is critical to
prediction and forecasting. This chapter explains and summarizes modern time series
modeling as used in insurance, actuarial studies, and related areas such as finance.
Modeling is illustrated with examples, analyzed with the R statistical package.
The distinction between counts and continuous variables disappears if the counts are large. For example, the GNP
of a country is a count of, say, dollars. However, for practical purposes GNP is regarded as a continuous variable.
427
428
sales
units
GL/year
250
100
50
0
200
150
0
186018801900192019401960
15
10
$/share
100
150
temperature
earnings
50
-1
0
1960 1965 1970 1975 1980
1850
1900
1950
2000
Fig. 17.1. Four time series described at the end of Section 17.1.1 and used to illustrate
time series analysis in this chapter.
points of time. These relationships are of interest for one or more of the following
reasons:
Forecasting. Forecasting aims to predict future outcomes or distributions. It requires
knowledge and understanding of the relationship between what is happening or has
happened and what will happen in the future. Forecasting requires a special case of
predictive modeling as generally discussed in this book: the events to be predicted
are in the future, and data used for prediction are the similar outcomes in the past
or present. This is well illustrated with the temperature series in Figure 17.1 that
measures global average temperature. Of interest is the likely progression of future
temperature based on historical temperature. Forecasts often provide a basis for action.
For example, the forecast that insurance claims will rise next quarter might necessitate
an increase in insurance reserves.
Understanding. We may simply want to know the whys and wherefores in the
movements of a time series.
Description. Can the information in a time series be compressed into a few meaningful
summary statistics. Any summary must hold across time relationships.
Control. Controlling future outcomes requires knowledge of the relationship between
variables happening now and the outcomes of these same or other variables in the
future.
429
Signal Extraction. Many time series are noisy indicators of underlying signals.
For example, the temperature time series has movements that appear to mask an
underlying trend. Signals often have some kind of continuity in time, and estimating
the relationship between measurements at different points of time makes for a better
recovery of a signal.
Inference. Inferring the effect of one variable on another requires partialing out the
effects of other, possibly across time, effects. For example, to detect outliers in the
Nile river flow time series, one has to understand the dynamics of Nile river flow.
Time series analysis with R. The analyses described in this chapter require detailed
calculations facilitated by a computer package such as R (R Development Core Team
2010). The R package provides a large number of time series datasets, especially if
the tseries package is loaded. Therefore, ensure the commands are accessible
using the library(tseries). Other R packages for time series analysis include
forecast, ltsa, FitAR, and FitARMA. Many time series are available in the R
datasets package. Issue the command data() to list the available time series,
including the first three plotted in Figure 17.1 and listed next.
r
r
r
r
where url is the http address in Jones (2010). The series is continually updated and
extended, requiring the updating of c(2013,8).
430
diff(sales)
5
4
3
2
1
0
-1
-2
-3
2
1
0
-1
1960
1965
1970
1975
1980
50
100
150
All logarithms discussed and used in this chapter are natural logarithms.
431
The first difference of the natural logarithm of a time series has a special significance:
yt
yt yt1
yt yt1
ln(yt ) ln(yt1 ) = ln
,
(17.1)
= ln 1 +
yt1
yt1
yt1
where the approximation is accurate if the difference yt yt1 is small compared to yt1 .
Hence analyzing differences in logs is tantamount to analyzing approximate percentage
changes.
s = 0, 1, 2, . . . .
Note that s = s and 0 0 is the constant variance of the time series. The sequence s as
a function of s is called the auto or lag covariance function. Lag covariances s satisfy other
properties such as |s | < 0 . For convenience it is often assumed that yt denotes the value
of the time series after subtracting its mean, implying that the mean of the adjusted series is
= 0 and s = E(yt yts ).
The covariance s divided by the product of the standard deviations of yt and yts yields a
correlation, called the lag correlation. By stationarity, the two standard deviations both equal
0 , and hence the lag correlation lag s is s /0 . Lag correlation as a function of the s is
called the auto or lag correlation function (ACF). The adjective auto emphasizes that it is
the correlation of yt with itself at another point in time.
Autocovariance s is estimated as
s
n
1
ts y),
(yt y)(y
n t=s+1
s = 0, 1, . . . n 1.
(17.2)
Here y is the sample mean of the time series. The divisor in s is n, even though the sum
contains only n s n terms.
432
Computing the ACF with R. Use acf(y) where y is the time series. Optional
parameters include the maximum lag and whether the auto covariance or
autocorrelation function is computed. Estimated ACFs corresponding to the time
series in Figures 17.1 and 17.2 are displayed in Figure 17.3.
Interpreting the ACF requires care because s is the unconditional covariance between yt
and yts . Other variables, including the intermediate variables yt1 , . . . , yts+1 , are not held
constant. A large value for s may be a consequence of relationships of yt and yts to these
other variables, as is further discussed later.
The ACFs in Figure 17.3 are those for each series in Figure 17.1 as well as their first
differences yt yt1 . Also plotted are the horizontal 2/ n confidence bounds discussed
later. The nile series displays mild autocorrelation, whereas the sales series has more persistent autocorrelation even in the first differences. The logarithm of earnings indicates strong
autocorrelation and periodicity in the autocorrelation of the first differences, with the ACF
peaking at lags 4, 8, . . . corresponding to an annual cycle. The temperature series also has
strong autocorrelation, which largely disappears once differences are taken.
within about 2/ n of zero for 95% of samples. Values of the ACF outside these bounds
suggest the time series is not noise. The 95% confidence limits are plotted in each of the panels
in Figure 17.3. Note that all the autocorrelations for small lags corresponding to each of the
series are significant. However, the significance of many of the autocorrelations disappears
once differences are taken.
A variant of noise is the random walk, which is also called integrated noise. Time series yt
is a random walk if the first difference series yt yt1 is noise or yt = yt1 + t . A random
walk is nonstationary because, using repeated substitution, yt = y0 + 1 + + t . Hence
the variance of yt is var(y0 ) + t 2 , diverges as it becomes large, provided 2 = 0. A random
433
nile
sales
0.5
0.5
-0.5
-0.5
0
10
15
20
log(earnings)
0.5
0.5
-0.5
-0.5
5
10
15
20
15
20
temperature
10
15
20
10
Fig. 17.3. Estimated ACFs of the series in Figure 17.1 as well as their first differences. For earnings, the estimates are for the log of the series. Lines in each panel
approaching zero more
rapidly correspond to the first differences. Horizontal lines in
each graph are the 2/ n approximate 95% limits for testing zero autocorrelation.
In R use plot(acf(y)) or plot(acf(diff(y))) where y is the series.
walk with drift is where yt = yt1 + + t . To test for a random walk, take the first
differences yt yt1 and see if the resulting time series is noise. From Figure 17.3 the nile
series does not appear to be a random walk because the ACF of the first differences at lag
1 lies outside the 95% limits. Similarly, none of the other three series appear to be random
walks.
Overdifferencing. If yt is noise t then the difference yt yt1 = t t1 has an ACF
that is zero everywhere, except at lag s = 1 where it is 1/2. Thus differencing can induce
autocorrelation where previously there was none, a situation often called overdifferencing.
Thus negative autocorrelation at lag 1 in the first difference of a series may be a consequence of
differencing, rather than of the model. For example, consider the ACFs of the first differences
in Figure 17.3. The nile, (log) earnings, and temperature series have negative autocorrelation
in the first differences that is partially explainable by differencing.
Geometric random walk (GRW). Suppose the percentage changes in the right hand side of
(17.1) have mean and variance 2 . Taking the exponential of (17.1) and rearranging yields
yt = yt1 e+t with t (0, 2 ). In this equation exact equality has replaced approximate
equality, appropriate if and are small, which is in turn enforced if the frequency of
observation is large. Thus the log of a GRW is a random walk with drift .
434
(17.3)
(17.4)
which is an MA(). The moving average (17.4) only makes sense if |a| < 1; otherwise the
coefficients in (17.4) blow up. The condition |a| < 1 ensures the variance of yt is finite and
is called the stationarity condition for an AR(1). Stationarity conditions are further discussed
later. Note the ACF of an AR(1) is never zero because yt and yts have, according to (17.4)
and for all s, common noise terms and hence are correlated.
MA(1). Here yt = t + bt1 is again assumed to hold for all t, and hence t1 = yt1 bt2
with similar expressions for t2 , t3 , and so on. Successively substituting these expressions
for past errors leads to
yt = t + byt1 + b2 yt2 + ,
which is an AR(). This representation only makes sense if |b| < 1, which is called the
invertibility condition for an MA(1). Again t represents the new part of the series at time t.
Note an MA(1) is always stationary, but need not be invertible. The ACF of an MA(1) is zero
for lags s > 1 because for such lags yt and yts are based on distinct noise terms.
MA(q). Here yt = t + + bq tq has a lag covariance function s that is zero for s > q
because yt and yts for s > q involve different noise terms that are uncorrelated. Further if
s < q, then s = 2 j bj bj s . This shows that every MA(q) is stationary provided q < .
An AR(p) is equivalent to an MA(), and hence the ACF does not cut out. More generally
the ACF of ARMA(p, q) does not cut out if p > 0.
3
Models other than ARMA models can also produce stationary time series.
435
(17.5)
Fitting ARMA models with R. Back casting and the numerical search for the
optimum values of the AR and MA coefficients are implemented in R with the
command arima(y,c(p,d,q)). For example, arima(nile,c(1,0,1)) fits an
ARMA(1,1) model to the nile data mentioned earlier. In the arima(y,c(p,d,q))
command y specifies the time series and p and q specify p and q, respectively. The d
value is the degree of differencing to be performed on the series prior to fitting. This
is further discussed later.
Note the right-hand side equals t if the AR and MA coefficients are the true coefficients. The notation et is used
here to indicate that the AR and MA coefficients are estimates.
436
(17.8)
b(L) 1 + b1 L + + bq Lq .
Equation (17.8) is shorthand for (17.3), with all yt terms taken to the left-hand side and L, the
lag operator Lk yt = ytk , serving to lag the time index on the yt or t terms. The ARMA(1,1)
has a(L) = 1 aL and b(L) = 1 + bL, whereas the AR(1) has a(L) = 1 aL and b(L) = 1,
and for the MA(1) a(L) = 1 and b(L) = 1 + bL.
In terms of the lag operator notation, the AR and MA representations of an ARMA(p, q)
are, symbolically,
yt =
b(L)
t ,
a(L)
a(L)
yt = t ,
b(L)
(17.9)
437
where the 1/zi are the roots (real or complex) of a(L). Both of these equations follow from
algebra. The expansion on the right is cogent if |zi | < 1 or, in terms of the root, |1/zi | > 1.
Further, 1/a(L) makes sense as an expansion in powers of L if all the factors can be expanded;
that is, if all the roots of a(L) lie outside the unit circle. This guarantees that the ARMA can
be written as a (convergent) MA. Since MAs with convergent coefficients are stationary, all
roots of a(L) outside the unit circle are called the stationarity condition for an ARMA(p, q).
The same argument applied to b(L) yields the invertibility condition for an ARMA(p, q):
an ARMA(p, q) can be written as an AR() if all the roots of b(L) lie outside the unit circle.
Invertibility has no bearing on the issue of stationarity.
The random walk is an ARMA(1,0) model with a(L) = 1 L that has a single root 1,
on the unit circle. Hence a random walk is not stationary. The model yt = t + t1 is an
ARMA(0,1) with b(L) = 1 + L. The model is stationary but not invertible because the root
of 1 + L is 1. Hence the latter model cannot be written as a pure AR. Nevertheless it is
obviously stationary. To see the issue, note t = yt t1 . Lagging once and substituting into
the equation for yt yields yt = t + yt1 t2 . Repeating this argument by substituting out
t2 and so on shows
yt = t + yt1 yt2 + yt3 ,
where the coefficients attaching to y values in the remote past are 1. Obviously this is not
a satisfactory representation: the model is noninvertible. Practically speaking, a noninvertible model implies that t cannot be written as the difference between yt and some linear
combination of previous yt values.
438
ACF
PACF
MA(q)
AR(p)
ARMA(p, q)
dies out
cuts out: zero for s > p
dies out
two variables after removing any effects of the intermediate variables yt1 , . . . , yts+1 . If
s = 1 there are no intermediate variables, and hence 1 = 1 /0 . At lag s = 2 the partial
autocorrelation removes the effect of the single intermediate variable yt1 . If the time series
is an AR(1) yt = ayt1 + t , then removing the effect of yt1 from yt leaves yt ayt1 = t ,
and the correlation between the latter and yt2 cyt1 is 0 because t is noise, unrelated to
the past. Hence the PACF for an AR(1) is zero at s = 2. Similarly at lag s > 2 an AR(1) has
PACF equal to zero because removing the intermediate variables is equivalent to removing
just yt1 , which leaves just t , as s = 2. Hence the PACF for an AR(1) cuts out: all values are
zero for s = 2, 3, . . ..
The PACF for an AR(p) cuts out for s > p. This holds because removing from yt the effects
of the intermediate variables yt1 , . . . , yts+1 for s > p leaves t , which is uncorrelated to
any previous values yt1 , yt2 , . . .. Since invertible MA(q) and ARMA(p, q) time series are
equivalent to AR(), it follows that the PACF for models with MA components never cuts
out (i.e., dies out). These results are displayed in the third column of Table 17.1.
The results in Table 17.1 are used to identify appropriate values for p and q given an ACF
and PACF function. Thus cutting-out behavior in either the ACF or PACF indicates a pure
AR or MA model, with p and q determined by the cutting-out point. If neither cuts out, an
ARMA(p, q) is indicated with neither p nor q equal to zero. In all cases p and q are assumed
small if not zero.
In practice, estimates of the ACF and PACF are available. Estimated PACFs of the series in
Figure 17.1 are displayed in Figure 17.5. The PACF lag s is estimated, in effect, by correlating
the residuals from the regressions of yt and yts on the intermediate values yt1 , . . . yts+1 .
These regressions and subsequent correlations are recursively computed in R with the pacf(y)
command.
For example, with the nile time series, the PACF for the undifferenced series appears to
cut out quite rapidly, whereas that of the differenced series has a negative early spike. Similar
to the ACF, the approximate 95% confidence limits for testing whether the PACF is zero
are 2/ n; these horizontal lines are indicated in each of the panels of Figure 17.5. The
PACFs for sales and log earnings cut out very early on, whereas for temperature there is no
early cutting out. The PACFs of the differences of log earnings and temperature have negative
spikes, suggesting possible overdifferencing.
The 2/ n limits on the ACF and PACF are applicable for large n and assume that all
population correlations for lags s > 0 are zero: that is the series is noise. If the time series is
not noise then the confidence interval is adjusted to c/n where, for the case of an AR(p),
sales
0.5
0.5
-0.5
-0.5
0
10
439
15
20
log(earnings)
10
15
20
15
20
temperature
0.5
0.5
-0.5
-0.5
0
10
15
20
10
Fig. 17.5. Estimated PACFs of the four series in Figure 17.1 for the original series
and first differences. For each series the lines approaching zero more rapidly cor
respond to the first differences. Horizontal lines in each graph are the 2/ n
approximate 95% limits for testing zero correlation. In R use plot(pacf(y))
or plot(pacf(diff(y))) where y is the series.
k1
rs2 .
(17.10)
s=1
(17.11)
called the LjungBox test statistic. If (17.11) is large, compared to the chisquare with k
degrees of freedom, then the hypothesis that there is zero autocorrelation at lags s = 1, . . . , k
is rejected. Use Box.test(y,k,"Ljung-Box") in R to compute (17.11) for time series y.
440
t (0, 2 ),
where a(L)(1 L)d has d roots on the unit circle and all remaining roots that is, the roots
of a(L) outside the unit circle.
ARIMA models are also called BoxJenkins models. With these models, nonstationarity, if
present, is assumed to be induced by roots on the unit circle, called unit roots. Non-unit roots
inside the unit circle are always excluded on practical grounds. The argument is that roots
inside the unit circle induce explosive behavior and few time series display such behavior.
The appropriate degree of differencing d is judged from the plot of the series and the
behavior of the estimated ACF and PACF. Evidence of unit roots is provided when both the
ACF and PACF converge very slowly to zero. This suggests d 1. If the first difference has
an ACF and PACF that also converge very slowly, then perhaps d 2, and so on. Practically
speaking d is rarely greater than 2. Once the appropriate degree of differencing is identified,
the stationary model for the differences is estimated as in the previous section.
The Box-Jenkins methodology proceeds according to the following steps:
(1) Check for and remove nonstationarity. Methods of detecting nonstationarity include
plotting the data against time and examining the estimated ACF and PACF. Loosely
speaking, nonstationarity can also refer to outliers and so on. Nonstationary in the mean
is generally removed by differencing. Nonstationarity in variance may be removed by
taking, for example, logarithms. Other ways of removing trends include regression
against time or the use of dummy variable regression to take out outliers.
(2) Identify the AR and MA orders p and q. Typically there are two or three orders at
the most. Overfitting, and eliminating nonsignificant AR or MA coefficients, is also
used to determine the appropriate p and q.
(3) Estimate the AR and MA coefficients. This step uses the techniques discussed in
Section 17.3.2.
(4) Check the fitted model. Diagnostic checking focuses on the fitted residual yt yt .
With a proper model that is, a model that correctly captures across time dependence
the residuals behave similarly to noise. The ACF and PACF of the residuals are thus
examined directly or with (17.11) to confirm noise. Also the sizes of the standard
errors in relation to the estimates are examined to see if any estimated coefficients can
be set to zero.
441
coef
se
,
ARIMA(1, 1, 0)
ARIMA(1, 1, 1)
0.31
0.08
1.37
0.84
0.08
1.32
258
0.61
0.12
253
ARIMA(2, 1, 0)
0.25
0.08
1.34
0.20
0.08
255
ARIMA(2, 1, 1)
0.77
0.20
1.32
0.04
0.12
253
0.56
0.19
To illustrate, consider the sales time series plotted in the top right panel of Figure 17.1 and
in the right panel of Figure 17.2 The ACF and PACF of the series and its first differences
are plotted in Figure 17.3, suggesting the series is nonstationarity, but the first difference
may be regarded as stationary. Results from fitting ARIMA(p, 1, q) models are displayed in
Table 17.2. For each fit the AR and MA coefficients are presented (in order), with standard
errors reported below each coefficient estimate. The final row reports the estimated and the
loglikelihood value for each fitted model.
The highest loglikelihood is achieved with the (1,1,1) and (2,1,1) models. With the
latter, the second AR coefficient is not significant, suggesting that the more parsimonious
ARIMA(1,1,1) is adequate. Here overfitting is used to show that a higher order model does
not provide statistically significant additional explanation.
The value of the statistic (17.11) computed from the ACF of the residuals of the ARIMA
(1,1,1) model with k = 10 is 8.20. Using a chisquared distribution with k 2 parameters (2 degrees of freedom are subtracted because two ARMA parameters are estimated)
the value 8.20 indicates no statistical significant autocorrelation in the residuals. Hence the
ARIMA(1,1,1) model appears adequate in explaining the across time dependence.
The estimated mean of the first differences yt yt1 is 0.04. The preferred model is thus
(yt yt1 0.04) = 0.84(yt1 yt2 0.04) + t 0.61t1 .
This equation is rearranged to yield an ARMA(2,1) equation appropriate for forecasting:
yt = 0.006 + 1.84yt1 0.84yt2 + t 0.61t1 ,
(0, 1.322 ).
This model is nonstationary: 1 1.84L + 0.84L2 = (1 L)(1 0.84L) has a unit root.
442
Common factors. Model multiplicity also arises in the context of common factors. Suppose
a(L)yt = b(L)t . Multiplying both sides by, say, 1 + cL leads to a different, though equally
valid representation. Obviously the resulting model is not parsimonious. In the model (1 +
cL)a(L)yt = (1 + cL)b(L)t , the term 1 + cL is a common factor that must be canceled out.
One step ahead prediction. Given the ARMA(p, q) model (17.3), the one step ahead forecast
of yn+1 given y1 , . . . , yn is
yn+1 = a1 yn + + ap yn+1p + b1 n + + bq n+1q .
(17.12)
This differs from (17.3) with t = n only in that n+1 is missing from the right-hand side. This
omission is justified given that n+1 is the innovation or new element entering at t = n + 1
and nothing is known about it other than it has mean zero and variance 2 . The values
n , . . . , n+1q in the right-hand side are estimated with yt yt where yt is the one step ahead
predictor of yt given its past. In this context, the variance of yn+1 yn+1 is called the mean
square error of one step ahead prediction. Once yn+1 is available, the next n+1 is calculated.
Thus one step ahead prediction is recursive: as each new observation becomes available, the
error of prediction is calculated, which in turn is used in the one step prediction of subsequent
observations.
MMSE prediction. One step ahead prediction discussed earlier is an example of minimum
mean square error (MMSE) prediction, in that the predictor minimizes the average squared
error: E{(yt yt )2 }. If the noise terms t are normal, then yt and the associated mean square
error are the conditional mean and variance yt given the past data. MMSE prediction is,
however, subject to criticism. First the MSE criterion penalizes over- and underprediction
equally. Further, squaring the error implies that the predictor is focused on avoiding large
errors. Despite these criticisms MMSE prediction is a favorite because it is relatively simple.
More than one-step-ahead prediction. The prediction of yn+m given the observations up
to time t = n is also computed recursively. Here m > 1 is called the lead time of the
prediction and n is fixed. Consider (17.3) with t = n + m. In the right of this equation any
yt not observed is replaced by its prediction given y1 , . . . , yn ; any future t that is, for
t = n + 1, n + 2, . . . is replaced by zero and the present and past errors t yt yt
for t = n, n 1, . . .: the one step ahead prediction errors. The mean square error of these m
step ahead predictions is found by substituting out successively yn+m1 , . . . , yn+1 , yielding
yn+m as a linear combination of n+m , . . . , n+1 plus a remainder term in n , n1 , . . . and
yn , yn1 , . . .. The future error terms are unknown, whereas the remainder term is known. The
variance of the prediction is thus 2 times the sum of squares of the coefficients associated
with the future noise terms. As with one step ahead prediction, uncertainty associated with the
AR and MA parameters is ignored.
Non-normal noise. ARIMA modeling assumes the approximate normality and independence
of the noise terms t . A histogram plot of the computed noise terms t as in (17.6), computed
with the estimated ARMA coefficients, will display departures from normality, if they are
present. Many financial time series are leptokurtic in that, despite symmetry, the tails of the
443
distribution are fat compared to the normal. Leptokurtic computed residuals et suggest that
an ARIMA model does not adequately capture the large shocks and that stochastic volatility
modeling, for example, is needed, as discussed in Section 17.4.2.
t+1 = t + t + at ,
t+1 = t + bt .
(17.13)
Only yt is observed and t , t , and t are mutually uncorrelated noise or disturbance terms
with common variance 2 . Further t is the level of the series, whereas t is the slope,
both varying with t. If (17.13) applies, then (1 L)2 yt is a linear combination of the noise
terms t , t , and t and hence is stationary. The time series thus has two unit roots.
444
t (0, 2 ).
(17.14)
This states that the sum of four consecutive seasonal components t adds to ct where t
is approximately zero, independent of the other disturbance terms. If c = 0 the components
add exactly to zero, corresponding to seasonal components modeled as in dummy variable
regression. Initial 0 , 1 , and 2 are estimated, because they determine subsequent seasonal
effects.
Structural models frame time series in terms of concrete and interpretable constructs: level
t , slope or trend t , seasonal t , and the associated parameters a, b, and c. This interpretability
contrasts with ARMA models where the interpretation attached to AR or MA coefficients is
more obscure. With structural models the emphasis is as much on estimating components such
as t and t as on estimating a and b. For example, with the temperature time series a pertinent
question is, What is the latest trend t ?
Fitting structural models with R. The command StructTS(y)) fits a structural
model to time series y including, where appropriate, a seasonal component. To
eliminate components, use keywords such as "level," which eliminates slope and
seasonal components, or "trend", which eliminates just the seasonal.
Figure 17.6 displays the output from the R when fitting the structural model to the log
earnings time series displayed in Figure 17.2. The output is in terms of variances of t , at ,
and bt . Thus,
!
!
1.033
3.471
a =
= 0.86,
b =
= 0.157.
= 0.001404 = 0.037,
1.404
140.4
445
Random walk with fixed drift. If a = 0 and b = 0 then t = 0 is constant and yt is a random
walk with fixed drift 0 . Zero drift results if = 0.
Starting conditions. The straight line and random walk with drift special cases indicate the
relevance of 0 and 0 . These starting conditions are not necessarily nuisance parameters,
but are often practically important and warrant estimation.
Testing restrictions. The significance of a component in (17.13) is judged by computing the drop in the loglikelihood when it is eliminated. For example, suppose r=StructTS(log(earnings)) is the R command. Then r$loglik displays
the value of the log-likelihood of the fitted model in this case 289.559. With
r=StructTS(log(earnings),"trend") the seasonal terms t are eliminated; in this
case r$loglik is 258.446, implying a drop of 289.559 258.446 = 31.113. The significance of the drop is evaluated with the likelihood ratio test:
Unrestricted Likelihood
= 2 (289.559 258.446) = 62.226.
2 ln
Restricted Likelihood
This test statistic is compared to the chisquared distributions with degrees of freedom equal
to the number of parameters set to zero: in this case c = 0 = 1 = 2 = 0, a total of four
parameters. Clearly the drop is very significant, and there appears to be a seasonal effect.
Stationary components. ARMA components can be included in the model to model stationary
dependence. For example t in (17.13) may be replaced by an AR(1) if there is short-term
transitory dependence in yt on top of the dependence induced by the nonstationary components.
Kalman filter. This is an algorithm for the recursive computing of one step ahead predictions
for models of the form (17.13) and its generalizations, often called statespace models. The
predicted value given the past, yt , and the realized value yt are combined with ancillary
information accumulated with the filter to derive the next one-step-ahead prediction yt+1 . The
weights in the weighted average are automatically adjusted as more observations come to
hand. The algorithm at the same time computes variance of the prediction error, as well as
estimates of the state (in this case t and t ), and their prediction variances and covariances.
Calculations assume values of a, b, c, and 2 .
Maximum likelihood estimation of structural models. The parameters of (17.13) and its
seasonal extension (17.14) are estimated iteratively. For example, with (17.13), given a guess
of the unknown parameters a, b, and 2 , the one step ahead MMSE predictor yt is calculated as
well as the MSE: nt=1 (yt yt )2 /n. Maximizing the likelihood is equivalent to minimizing the
MSE plus second-order terms. Hence maximum likelihood estimation proceeds by repeatedly
guessing the unknown parameters, evaluating the MSE and the second-order adjustment, and
using a numerical search to find the maximizing parameters. The search is facilitated by using
numerical optimization.
Smoothing filter. This recursive algorithm computes estimates of t , t , and the noise terms,
for t = 1, . . . n based on all available data y1 , . . . , yn , not just the past. For example, given a
time series y1 , . . . , yn the slope t near the middle of the series may be required as well as its
446
5
1.5
2.5
0
-2.5
0.5
-5
0
1995
1996
1997
1998
1999
1995
1996
1997
1998
1999
t+1 = t + + 0.387t .
Hence , the initial value, is critical because it indicates the constant slope of the process. The
initial value is estimated as = 0.0007, corresponding to an increase of 100 12 0.0007 =
0.84 C per century.
2
t+1
= a0 + a1 t2 + b1 (yt )2 .
(17.15)
447
448
series are measured simultaneously. In this case yt is a vector of measurements with different
components corresponding to, for example, different stock prices, a variety of economic
indicators, or mortality at different ages. The concept of serial correlation is now extended to
correlation between different components of yt at different points of time.
The Lee-Carter (LC) model is a special, idiosyncratic multiple time series model much used
in the analysis of mortality. The time series measurements are mortality deaths at different
ages i = 0, 1, . . . , p, say, where p is the maximum age and it is assumed that
yt = a + bt + t .
Here a and b are vectors with as many components as there are ages, whereas t is an
unobserved scalar time series modeling the trend in overall mortality, and t is a vector of
mutually uncorrelated noise terms. The vector of mortalities yt is usually taken as the logarithm
of the death rate; that is, the number of deaths divided by exposure. The aim of the analysis
is to estimate a and b as well as make inferences about the overall trend t and, if required,
forecast t to obtain an understanding of likely future death rates.
References
Box, G. E. P., G. M. Jenkins, and G. C. Reinsel (1994). Time Series Analysis: Forecasting
and Control (3rd ed.). Prentice-Hall, Englewood Cliffs, NJ.
Chatfield, C. (1989). The Analysis of Time Series: An Introduction (2nd ed.). Chapman and
Hall, London.
Harvey, A. C. (1989). Forecasting, Structural Time Series Models and the Kalman Filter.
Cambridge University Press, Cambridge.
Jones, P. (2010, April). Climate Research Unit at the University of East Anglia.
http://www.cru.uea.ac.uk/cru/data/temperature/
CRUTEM3-gl.dat.
Kendall, M. G. and J. K. Ord (1990). Time Series (3rd ed.). E. Arnold, Seven Oaks, UK.
R Development Core Team (2010). R: A Language and Environment for Statistical
Computing. Vienna, Austria: R Foundation for Statistical Computing.
Shumway, R. H. and D. S. Stoffer (2011). Time Series Analysis and Its Applications: With R
Examples. Springer, New York.
18
Claims Triangles/Loss Reserves
Greg Taylor
Chapter Preview. This chapter considers the application of predictive models to insurance claims triangles and the associated prediction problem of loss reserving (Section 18.1). This is approached initially by reference to the chain ladder, a widely used
heuristic reserving algorithm. Rigorous predictive models, in the form of Generalized
linear models (GLMs) that reproduce this algorithm, are then explored (Section 18.2).
The chain ladder has a restricted model structure and a number of embellishments
are considered (Section 18.3). These include the incorporation in the model of accident year effects through the use of exposure data (e.g., accident year claim counts)
(Section 18.3.2), and also the incorporation of claim closure data (Section 18.3.3).
A subsequent section considers models that incorporate claim closure data on an
operational time basis (Section 18.3.4). In each of these cases emphasis is placed
on the ease of inclusion of these model features in the GLM structure. All models
in Sections 18.1 to 18.3 relate to conventional claims triangles. These datasets are
aggregate, as opposed to unit record claim datasets that record detail of individual
claims. The chapter closes with a brief introduction to individual claim models (18.4).
On occasion these models use survival analysis as well as GLMs.
18.1 Introduction to Loss Reserving
18.1.1 Meaning and Accounting Significance of Loss Reserving
Typically, property & casualty (P&C) insurance will indemnify the insured against
events that occur within a defined period. From an accounting viewpoint, a liability
accrues to the insurer at the date of occurrence of an indemnified event. From that
date the insurers accounts are therefore required to recognize the quantum of this
liability to the extent that it has not been settled.
However, there will be an inevitable delay between the occurrence of the event and
notification of an associated claim to the insurer, and a further delay between notification and settlement of the claim. These delays can be considerable. For example, in
449
450
casualty classes of business, such as auto bodily injury liability, a typical period from
claim occurrence to settlement might be 3 or 4 years, with some claims remaining
unsettled for 10 to 20 years.
During the period between occurrence and notification, even the existence of the
claim will be unknown to the insurer. Between notification and settlement its existence
will be known, but its ultimate quantum will be unknown because it may depend on
the outcome of negotiation between insurer and insured or on an adversarial legal
process. The insurer is obliged to make accounting provision for the claim despite
these shortfalls in knowledge.
The balance sheet item recognizing the insurers liability is called the loss reserve,
and the determination of this reserve is loss reserving. The reserve will certainly
include the liability for the cost of the claims themselves and depending on the
accounting regime, it may also include allowance for other associated costs such as
loss adjustment expenses.
For long-tail lines of insurance, the loss reserve may amount to several times the
annual premium income, and therefore it may amount to a large multiple of typical
annual profit. As a result, a modest proportionate error in the loss reserve can impinge
heavily on annual profit, and so accuracy of the reserve is a matter of considerable
importance.
The estimation of loss reserves in practice relies on a wide variety of models and
methodologies. It is not feasible to review them all here; for that review, refer to
Taylor (2000) and Wuthrich and Merz (2008). Of all models in use, the so-called
chain ladder is the most commonly encountered. For this reason this chapter begins
with this model (Section 18.2), and then proceeds to examine variations of it as well
as some other models.
18.1.2 Data and Notation
A typical P&C claim is generated by the occurrence of some event. The generating
event may be an accident (e.g., automobile bodily injury insurance), or it may be
something else (e.g., professional liability insurance, where the date of occurrence
is the date on which some lapse of professional conduct occurred). Nonetheless, the
date of this event is referred to generically as the accident date.
For the purpose of data analysis it is common to group claims according to the
calendar periods in which their accident dates fall. These periods may be years,
quarters, months, or any other convenient choice and are referred to as accident
periods. For convenience, the present chapter assumes that they are accident years.
Some lapse of time usually occurs between the generating event and the subsequent
payment of the claim. Time measured from the accident year is referred to as development time, and years measured from the accident year are referred to as development
451
ykj
PAST
OBSERVATIONS
Accident
year .
.
.
FUTURE
K
years. The accident year itself is labeled development year 1, and subsequent years
development years 2, 3, and so on.
Consider the situation at the end of calendar year K. Accident year K will have just
been completed (i.e., development year 1 will have been completed). For accident year
K 1, development years 1 and 2 will have been completed. In general, for accident
year k ( K), development years 1, 2, . . . , K k + 1 will have been completed.
Observations on accident years 1, 2, . . . , K thus form a triangular structure, called
a claims triangle, as illustrated by the upper part of Figure 18.1. Accident years are
represented by rows and labeled k = 1, 2, . . . , K; development years are represented
by columns and labeled j = 1, 2, . . . , K; and ykj denotes the observation on claims
experience in development year j of accident year k. Diagonals in this triangle,
represented by k + j = const, denote contemporaneous years of calendar time and
are referred to subsequently as calendar years.
The precise nature of the observations ykj are left unspecified for the moment. They
may comprise counts of claim notifications or closures, amounts of paid losses or case
estimates, or some alternative. Specific examples are given later in this chapter.
Let the claims triangle in Figure 18.1 be denoted DK = {ykj : 1 k K, 1
j K k + 1}. Future observations on the same accident years, up to and including
c
development year, K, form the lower triangle in Figure 18.1, denoted DK
= {ykj :
2 k K, K k + 1 < j K}. Also let
+
c
DK
= DK DK
.
c
This chapter is concerned with the prediction of DK
on the basis of observed DK .
c
Row sums in DK are of particular interest. These are the quantities
Rk =
K
j =Kk+2
ykj ,
k = 2, 3, . . . , K.
(18.1)
452
j
yki .
(18.2)
i=1
K
(18.3)
j =Kk+2
R=
K
Rk .
(18.4)
k=2
c
Note that R is the sum of the (future) observations in DK
. In the case that the ykj denote
paid losses, R is referred to as the total amount of outstanding losses. Likewise, Rk
denotes the amount of outstanding losses in respect of accident year k. The objective
stated earlier is to forecast the Rk and R.
In regarding R as the total amount of outstanding losses, it is assumed that no
claims activity occurs beyond development year K. If this is not so, estimated outstanding losses will require increase accordingly. Such an increase will not, however,
be obtainable from the available data, which do not involve development years beyond
K, and the details of this issue are beyond the scope of the present chapter. In general
terms, one would need to rely on the following:
(18.5)
which indicates that ykj is a random variable with some distribution Dist defined by
a vector of parameters hkj (b) hkj being some function specific to cell (k, j ) and b
+
being a vector of parameters relating to the entire array DK
rather than to cell (k, j ).
453
The fact that ykj is a random variable in the model definition means that the
model is stochastic. Notwithstanding that this is essential if the stochastic properties
of model forecasts are to be understood, many conventional actuarial loss reserving
models are not of this type (see, e.g., Taylor 2000). One such model is encountered
in Section 18.2.1.
fj =
j = 1, 2, . . . , K 1.
Kj
k=1 xkj
(18.6)
c
:
Then define the forecasts of ykj DK
(18.7)
K
(18.8)
(18.9)
j =Kk+2
R =
K
R k .
(18.10)
k=2
454
455
Response variate
Covariates (categorical)
ykj
Accident years k, development
years j
Log
ODP
None
None
ln k + ln j
yk,j +1 |xkj
Accident years k, development
years j
Log
ODP
Column dependent permitted
ln xkj
ln fj
Link function
Error structure
Weights
Offset
Linear response
456
(18.11)
with the fi (.) predefined parametric functional forms and the i regression coefficients
requiring estimation.
Generalized additive models were introduced into the loss reserving literature by
Verrall (1996). The idea of loss reserving by means of regression with parametric
functions as independent variables can also be found in earlier papers (e.g., De Jong
and Zehnwirth 1983; Frees and Wang 1990; Reid 1978). Some of these approaches
are discussed by Wuthrich and Merz (2008).
It is sometimes appropriate in practice to suppose that the profile of the j over j
follows a gamma-like shape:
j = Aj b exp(cj ),
(18.12)
(18.13)
with a = ln A.
The linear predictor (k, j ) is indeed linear in all unknown parameters, and the
number of those parameters has been reduced from the 2K 1 of the categorical
model to K + 2 (again one parameter will be aliased).
It is of particular interest that the adoption of a smooth functional form f (j ) in
(18.11) ensures both a smooth development year profile and a smooth sequence of
estimated age-to-age factors. It would replace the hand smoothing of the sequence of
fj mentioned at the end of Section 18.2.1. However, the parameters defining f (j )
would be estimated objectively, which may be preferable to ad hoc smoothing.
18.3.1.2 Accident Year Profiles
The same sort of reasoning may be applied to the models accident year structure. If
there were no indication of dependence of the linear predictor on accident year, then
457
the categorical parameters k could simply be dropped from (18.13) and the number
of parameters would be reduced to three.
Alternatively, if the linear response appeared to exhibit a linear trend in k (exponential dependence of observations on k), then the linear predictor would become
(k, j ) = mk + a + b ln j + c(j ),
(18.14)
with m a new unknown parameter. The number of model parameters is now four.
18.3.1.3 Calendar Year Profiles
Inflationary trends operate over calendar years. That is, in the presence of an inflation
factor of e between calendar years k + j and k + j + 1 but no other accident year
trend, the expectation of each observation from calendar year k + j + 1 will be a
multiple e of that of the same development year in calendar year k + j .
A constant rate of inflation would be added to model (18.13), and the accident year
trend omitted, by means of the following linear predictor:
(k, j ) = a + b ln j + c (j ) + (k + j ).
(18.15)
Note that (18.15) is, as required, linear in the unknown parameters a, b, c, and .
If the ykj represent claim amounts, then they may have been corrected for inflation
before modeling in an attempt to eliminate the calendar year trend from the model.
This correction is typically carried out using some standard form of monetary inflation;
for example, adjusting workers compensation claim payments by reference to wage
inflation or adjusting automobile property damage claim amounts by reference to a
mixture of price inflation (vehicle parts) and wage inflation (repair labor). On such a
basis the ykj would be adjusted to common dollar values, such as those relating to the
end of calendar year K, the date of evaluation of the loss reserve.
Claim amounts may sometimes escalate at a rate different from the seemingly
natural inflation index. The difference between the observed rates of inflation and
those associated with the inflation index is called superimposed inflation (SI). In such
cases a calendar year trend is required in the model of claim amounts and is used to
estimate SI.
SI is often found to progress irregularly over time, and so similarly for the total
inflation of claim amounts. In this case a constant rate will not exist as in (18.15),
and it will be necessary for (18.15) to include a more general function of (k + j ). See,
for example, Section 18.4.2.
18.3.2 Use of Exposure Data
For some lines of insurance it is possible to identify a time series {ek } with respect to
accident year such that the claims experience of accident year k can be expected to be
458
proportional to ek . For example, the counts of auto bodily injury claim notifications
of accident year k might be proportional to the average number of vehicles insured
over that year. The counts of workers compensation claim notifications of accident
year k might be proportional to the total payroll of the workforce over that year.
In such cases the quantity ek is referred to as the exposure associated with accident
year k. Since claims experience is postulated to be proportional to it, exposure replaces
k in (18.11) thus:
(k, j ) = ln ek + f (j ).
(18.16)
This substitution may appear trivial. However, it should be noted that ek in (18.16) is
a known quantity, whereas k in (18.11) is an unknown quantity requiring estimation.
In other words, ln ek forms an offset within the GLM, whereas k is a parameter.
Recall that the model under discussion uses a log link, in which case ln E ykj =
(k, j ) and so, by (18.16),
7 8
ln E ykj /ek = f (j ) .
(18.17)
The quantity within the braces represents claim experience per unit exposure. If, for
example, ykj denotes the amount of paid losses in the (k, j ) cell, then the quantity
within braces represents payments per unit exposure. If ek = N k (i.e., the estimated
number of claims incurred in accident year k), then the the quantity within braces
represents payments per claim incurred, as illustrated in section 4.2 of Taylor (2000).
By linearity of the expectation operator, (18.17) yields
ln {E [xkK ] /ek } =
K
f (j ) ,
(18.18)
j =1
and if ykj denotes the claim notification count in the (k, j ) cell, then the quantity
within braces represents ultimate claim frequency for accident year k (e.g., claim
frequency per vehicle insured in the auto bodily injury case mentioned at the start of
this subsection).
The right side of (18.18) is independent of k and so represents the case of claim
frequency that is constant over accident years. Although the exposure term was
introduced into (18.16) as a proxy for the accident year effect k , it will be an
inadequate proxy if claim frequency exhibits a trend over accident years. In this case
it will be necessary to reintroduce k to capture that trend, so that (18.18) becomes
ln {E [xkK ] /ek } = k +
K
j =1
f (j ) ,
(18.19)
459
(18.20)
K
f (j ) ,
(18.21)
E [xkK ] /ek = (exp )k exp
j =1
and claim experience per unit exposure exhibits an exponential trend over accident
year.
18.3.3 Use of Claim Closure Data
The present subsection is specific to the case in which ykj denotes the amount of
paid losses in the (k, j ) cell. Suppose that the number of claim closures in the cell
is available and is denoted ckj . For certain lines of business, it may be reasonable to
assume that ykj will be related to ckj ; specifically, ykj may be modeled as
7
8
ln E ykj /ckj = k + f (j ),
(18.22)
or, in a form parallel to (18.20),
ln E ykj = ln ckj + k + f (j ) ,
(18.23)
where the quantity in the braces in (18.22) is the expected amount of paid losses per
claim closure (PPCC) (see section 4.3 of Taylor 2000) and the term ln ckj in (18.22)
is now an offset. PPCCs are sometimes referred to as payments per claim finalized
(PPCF).
18.3.4 Operational Time
Define
dkj =
j
cki ,
(18.24)
i=1
which is the cumulative number of claims closed up to the end of development year
j of accident year k. Note that dk = Nk , the number of claims incurred in accident
year k, because all of these claims must ultimately close.
460
Accident Year k + 1
%
10
20
50
80
100
$
5,000
10,000
30,000
50,000
100,000
%
20
50
80
100
$
7,500
30,000
50,000
100,000
Now define operational time (OT) associated with accident year k and corresponding to elapsed time j (integral) from the beginning of that accident year as
tk (j ) = dkj /Nk ,
(18.25)
(18.26)
(18.27)
461
where kj , kj are the mean and coefficient of variation, respectively, of ykj and f is
some function requiring specification. By (18.26), the model has a log link, and as
previously, (k, j ) denotes the linear predictor.
Accident year and calendar year effects may be included as necessary, just as in
Section 18.3.1.
18.3.5 Interactions
From Chapter 5, the linear response of a GLM takes the vector form
= X,
or, in component form,
i =
xim m ,
(18.28)
(18.29)
with xim the value of the m-th explanatory variable for the i-th observation.
If one now wishes to add a further effect to the linear response, involving say the
m1 -th and m2 -th explanatory variables, and this effect cannot be expressed as a linear
combination of functions of these variables, then the linear response takes the form
i =
xim m + f (xim1 , xim2 ),
(18.30)
m
(18.32)
Now suppose that this model holds for accident years up to and including 2005. A
legislative change was introduced at the start of 2006 that, with one exception, halved
the expected PPCCs. The exception affected the final 20% of closures. These are, by
and large, the most serious claims, and the legislation did not change the conditions
governing their entitlements. The results are shown in Table 18.3.
462
$
5,000
10,000
30,000
50,000
100,000
$
2,500
5,000
15,000
25,000
100,000
(18.33)
The term (ln 1/2) I (k 2006) results in the halving of all PPCCs for accident years
2006 and later, and the additional factor I (0 tk (j ) < 0.8) ensures that this does not
apply to OTs above 80%. The model covariates here are shown in Table 18.4. Note
that the term added to the linear predictor in (18.32) is a non-linear function of this
set of covariates. It is therefore of the same nature as the final member of (18.30) and
constitutes an interaction.
18.3.6 Forecasts
The purpose of the modeling just described is, as discussed in Section 18.1, ultimately
c
the forecast of DK
on the basis of the model of DK . The general form of a GLM of
Table 18.4. Set of Covariates
in Model (18.33)
I (0 tk (j ) < 0.1)
I (0.1 tk (j ) < 0.2)
I (0.2 tk (j ) < 0.5)
I (0.5 tk (j ) < 0.8)
I (0.8 tk (j ) < 1)
I (tk (j ) 2006)
463
(18.34)
where y denotes the set of ykj DK arranged as a vector, y EDF indicates that
it is sampled from a member of the exponential dispersion family, denotes the
vector of model parameters, X a design matrix, and g a link function that operates
component-wise on its vector argument.
c
Let the set of ykj DK
be arranged in a vector, yc , which may be represented in a
form parallel to (18.34):
c
c
= g 1 X c ,
EDF, with E ymodel
(18.35)
ymodel
where X c is the matrix of covariate values that apply to the future. The notation
c
ymodel
is used to recognize that the right side of (18.35) is an idealized representation
of reality and that observation may not (almost certainly will not) conform with it
c
.
precisely. In other words, yc is distinguished from ymodel
c
c
c
A forecast y of y is given by an estimate of E ymodel from the model, which is
yc = g 1 X c
(18.36)
by (18.35), with an estimate of .
The matrices X and X may be illustrated by the following
example. One
may write y = y11 , y12 , . . . , y1K , y21 , y22 , . . . , y2,K1 , . . . , yK1 with the prime
denoting transposition. Now consider, for example, the ODP cross-classified
model described in Table 18.1,
in which case, g(.) = ln(.), = ln 1 , ln 2 , . . . ,
ln K , ln 1 , ln 2 , . . . , ln K , and
1 0 0
1 0 0 0
1 0 0
0 1 0 0
K rows
..
1 0 0
0 0 0 1
1 0 0 0
0 1 0
0
1
0
0
1
0
0
K 1 rows
X =
(18.37)
..
0 1 0
0 0 0 1
..
..
1 row
0 0 1
1 0 0 0
464
c
Let the set of ykj DK
be arranged in a vector yc = y2K , y3,K1 , y3K , . . . , yK2 ,
yK3 , . . . , yKK . Then
0 1 0 0
0 0 0 0 1
0 1 0
0 1 0
0 0 0 1 0
0 0 0 0 1
Xc =
..
.
0 0 1
0 0 1
..
.
0 0 1
0 1 0 0 0
0 0 1 0 0
0 0 0 0 1
1 row
:
2 rows
..
.
(18.38)
K 1 rows
y
,
c = g 1 X c g 1 X c
model
process error
parameter error
(18.40)
model error
465
The actual observation yc will also contain the random error (i.e., noise) c , referred
to as process error.
Parameter error and process error are independent because the former depends on
only past observations and the latter on only future observations, and all observations
are independently modeled with a GLM. It will be convenient, and perhaps not greatly
inaccurate, to assume that both parameter error and process error are independent of
model error.
If all three components of forecast error are stochastically independent, then (18.40)
yields
c
c
(18.41)
Var c = Var c
model + Var parameter + Var process ,
where the three arguments on the right side are the model, parameter, and process
errors from (18.40).
The quantity Var [ c ] is referred to as the mean squared error of prediction and
is often written as MSEP[ c ]. It is sometimes referred to as the forecast (or prediction) error, and its components referred to as model error, parameter error, and
process error, respectively. Such usage leads to confusion with the earlier definition
of these terms, and so the present chapter retains those earlier definitions, referring
to the components of (18.41) as model MSEP, parameter MSEP, and process MSEP,
respectively.
466
(18.43)
Var y (n)
(n)
(n)
(n) ,
(18.44)
where the last equality is obtained by recognizing that the two members on the right
represent contributions from past data and future noise, respectively, and, as already
mentioned, they are stochastically independent.
Comparison of (18.44) with (18.40) indicates that the two members on the right of
the former equation are estimates of parameter MSEP and process MSEP, respectively.
The bootstrap has thus provided estimates of two of the three components of total
forecast MSEP. A numerical example can be found in chapter 11 of Taylor (2000).
The remaining component, model MSEP, is not obtainable from the bootstrap
because it does not contemplate model designs other than the single one according
to which the forecasts have been made (see (18.34)). The estimation of model error
remains problematic, with little coverage in the existing literature, despite the fact
that in many cases its magnitude will be highly material, possibly even exceeding the
total of parameter and process MSEP. One substantial, albeit subjective, attempt to
address the issue is found in ODowd, Smith, and Hardy (2005).
(18.45)
467
where ckj denotes the number of claim closures (as previously) and pkj denotes the
observed average PPCC. The corresponding forecast is then
ykj = ckj p kj ,
(18.46)
468
r A transaction file. This consists of one record per claim transaction, containing information (often financial information) about that transaction, such as the following:
r policy number;
r date of transaction;
r type of transaction, such as a claim payment or a change of claim status (e.g., from
open to closed);
r amount of claim payment (if there is one);
r type of payment (e.g., income replacement, medical expense).
For modeling purposes, the claims header file, and transaction file are typically be
merged into a single claim data file, containing one record per claim. That one record
may contain details of many transactions.
The claims triangle of Figure 18.1 dates from an age in which modern data handling
facilities were not available and there was a need to summarize data into a convenient
form for analysis. To the extent that one is concerned with the analysis of claims
experience, this motivation of claim triangles no longer exists, and the more natural
dataset for loss reserving may often be the unit record claim data file just described
(though note the comment on IBNR claims in Section 18.4.4).
The use of individual claim information may, of course, involve the manipulation of
large datasets. For example, a portfolio with a 10-year history may contain hundreds
of thousands of claims. These would provide hundreds of thousands of observations for analysis, with each observation accompanied by a possibly large number
of explanatory variables. The claims triangle would summarize this data into just
55 observations, each with just two explanatory variables: accident year and development year. However, this compression of data is liable to eliminate much useful
predictive information.
469
Alternatively, the dataset represented by y might comprise only closed claims, with
yi denoting the total cost of the i-th claim. In this case the regression could be used
to forecast the ultimate sizes at closure of currently open claims. This example is
evidently relevant to loss reserving, and some details of it are pursued next.
Example 18.1 (Australian Automobile Bodily Injury Scheme). Consider the
claims triangles in Table 18.5 and Table 18.6. These are annual summaries of data
that are collected quarterly that summarize selected data from the automobile bodily
injury scheme of one Australian state. The scheme is tort-based but subject to some
controls:
r For accident dates from October 2000 the amounts that could be awarded as plaintiff
legal costs were subject to prescribed ranges that were related to the amounts of general
damages awarded (e.g., no plaintiff costs could be awarded if the amount of general
damages was less than $25,000).
r For accident dates from December 2002 all injuries were awarded a point score under
scheme guidelines, and this score then determined the amount of general damages to
be awarded, diminishing judicial discretion.
The scheme began in October 1994, and so the tables cover almost its entire history
to date. The injury associated with each claim is assigned one of six severities ranging
from 1 (lowest) to 5 (catastrophic injury) and 6 (fatality).
Table 18.5 displays the average claim size of legally represented claims closed in
each cell with injury severity 1 (the lowest severity). All claim sizes are adjusted to
common dollar values according to the wage inflation index of the state concerned.
Average claim sizes beyond development year 11 are not shown because they become
erratic due to sampling error.
Although Table 18.5 presents average claim sizes in the conventional triangular
form, this is for convenience of presentation only. Modeling is carried out in terms of
individual claims. There are more than 125,000 claims.
Table 18.6 displays the OTs of the same subset of claims attained at the ends of the
development years represented by the columns are not shown beyond development
year 6 because they approach 100% and so do not inform the present discussion.
An earlier version of the same dataset appeared in Taylor and McGuire (2004),
together with a detailed discussion of its modeling, and a further discussion was given
by McGuire (2007).
The tables display three distinct and obvious characteristics that require incorporation in any model:
(1) Claim size increases steadily with increasing OT.
(2) There is a discontinuity in experience between accident years 2002 and 2003. Legislation that took effect in December 2002 made fundamental changes to the conditions
470
Table 18.5. Numerical Example: Average Sizes of Closed Claims (Inflation Adjusted)
Development Year
Accident Year
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
10
11
$
14,620
13,161
12,343
14,448
15,082
17,461
11,792
18,013
7,787
4,196
5,222
6,579
8,714
12,681
11,717
23,420
11,633
$
19,184
18,621
20,618
24,954
27,091
29,851
31,633
34,404
11,129
14,540
25,762
39,832
52,607
50,332
53,285
53,609
$
32,661
34,383
43,207
43,874
50,416
53,776
52,060
47,482
27,803
42,393
60,734
74,430
78,564
79,380
81,832
$
57,481
60,636
62,752
67,325
76,817
81,377
64,294
63,724
50,640
80,391
91,310
95,254
92,617
103,572
$
90,989
92,385
97,396
95,421
105,900
101,728
90,423
99,474
90,792
121,179
108,486
130,581
128,433
$
106,046
103,483
132,429
109,523
118,667
120,093
117,821
121,201
113,241
125,790
159,563
155,904
$
134,588
150,741
131,814
108,293
113,189
136,809
131,769
169,461
110,052
147,571
144,770
$
101,227
136,327
115,993
93,789
177,795
161,484
160,808
182,175
160,063
113,933
$
151,094
132,676
155,410
127,846
124,940
193,284
162,300
153,423
125,909
$
144,406
86,991
220,250
151,627
122,901
204,604
259,690
154,628
$
76,189
104,742
92,973
116,563
150,407
126,939
185,004
471
1
%
2
%
3
%
4
%
5
%
6
%
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
7
6
5
6
9
6
4
5
1
1
2
1
2
2
2
1
2
32
36
42
41
43
35
36
31
16
18
20
21
22
22
23
26
59
66
71
68
68
67
64
59
45
49
56
56
54
57
61
80
84
86
82
84
84
84
80
76
76
81
79
77
81
91
92
92
90
92
92
92
92
90
90
91
90
90
96
95
96
95
96
96
96
96
95
95
95
95
of the scheme, that eliminated many smaller claims. This elimination shifted the
development patterns of both OTs and average claim size.
(3) The changes discussed in (2) result in average claim sizes that increase over time.
These sizes were already adjusted for inflation, so a question arises as to whether the
increase amounts to SI. This is discussed later, but a period of very rapid increase is
particularly noticeable over the shaded area of Table 18.5.
The model is a single GLM of average claim size with log link and comprises submodels of the six different injury severities. The model also includes the following
effects:
r
r
r
r
r
472
of its range (in fact, 0.2 OT 0.8). That is, (18.34) for an individual observation
reads
) (OT )
X
+ i ,
(18.47)
yi = exp (OT
mi
imi
with
(OT )
,
=
max
0.2,
min
0.8,
t
Xim
im
i
i
(18.48)
where mi is the severity of the i-th claim, timi is the (severity) mi OT of that claim at
closure calculated on the basis of severity mi claims only (i.e., only severity mi claims
)
included in the numerator and denominator of (18.25)), (OT
is the rate of increase
m
per unit of OT in log average claim size of severity m, and the dots indicate that there
are many other terms in the regression.
Superimposed inflation. It is a feature of this scheme that SI is uniform over
neither time nor injury severity. Claims experience exhibits periods of SI quiescence,
punctuated by short and possibly sharp bursts of SI. These dynamics reflect the
shifting ascendancy of claimants and their legal representatives, on the one hand, and
the schemes cost containment measures, on the other.
When SI does occur, its impact in terms of proportional increase tends to be
greatest for smaller claims. Large claims, involving catastrophic injury, are subject to
reasonably clear precedent and show relatively little increase in real terms over time.
Smaller claims, in contrast, have headroom within which the boundaries of damages
claimed can be tested.
Analysis of the claim sizes underlying Table 18.5 indicates SI of severity 1 claims
over the two-year period 2006Q2 (the second quarter of 2006) to 2008Q1. The rate of
SI decreased linearly with increasing OT, attaining a zero value at OT=100%. Thus
(18.34) for an individual observation reads
) (SI )
X
+ i ,
(18.49)
yi = exp (SI
1
i1
with
(SI )
= (1 ti1 ) max q , min qi , q ,
Xi1
(18.50)
and where qi denotes the calendar quarter of closure of the i-th claim, q , q denote
)
is the rate of SI for severity 1 claims
2006Q2 and 2008Q1, respectively, and (SI
1
at OT zero. Note that the contribution to the linear predictor appearing in (18.49) is
constant for qi < q and, qi > q , indicating no SI over these time intervals.
Claims of higher injury severities were differently affected by SI. Generally, the
effects were less (in proportional terms), and severities 4 to 6 were not affected at all.
The 2002 legislation. Two major effects of the legislation are apparent. First,
Table 18.6 shows a marked reduction in the rate of claim closure under the new
473
legislation, followed by a partial recovery. Second, Table 18.5 shows a marked reduction in the sizes of closed claims, followed by an increase.
It is not immediately obvious whether the changes in claim sizes are genuine or
induced by the changed rates of closure. However, one can, for example, calculate a
cumulative average size of claims closed in the first two development years of accident
year 2002 as 1% $18,013 + (16% 1%) $34,404 = $31,925. This relates to OT
31%.
The corresponding calculation for the first three development years of accident
year 2003 gives a cumulative average claim size of $21,618 at OT 45%. This confirms
a genuine reduction in claim sizes, because the data indicate that average claim size
increases monotonically with increasing OT.
The reductions in both closure rates and claim sizes appear to narrow as one goes
to later development years, suggesting that larger claims have been less affected by
the legislative change.
Such calculations are tedious, and the legislative effect on claim sizes is much
better quantified by the inclusion of an accident year effect in the GLM. The inclusion
of an accident year effect does indeed confirm a confinement of the effect to smaller
claims, which was indeed the intention of the legislation.
Further investigation of the data indicates several simultaneous effects of the legislation on claims experience:
r Some claims were eliminated from experience entirely. This was not as obvious as
might be expected because claim frequency was declining anyway, both before and
after 2003.
r Naturally, the eliminated claims were typically those that would have been small in
the absence of the legislation. This elimination of claims occurred only for injury
severity 1.
r These eliminations meant that the relation between OT and average claim size differed
between pre-2003 and post-2002 accident years. Claims that would have been closed at
OT t in the older accident years were closed at OT h (t) in the later ones, where h (.) is
a contraction distortion in the sense that h (0) = 0, h (1) = 1, h (t) t for 0 < t < 1.
r The degree of claims elimination varied as the effect of the legislation matured. As a
result the distortion function varied over the accident periods 2003 and later, before
eventually stabilizing.
r There were genuine reductions in claim size superimposed on the above changes. These
were observed in only the lower injury severities (but not only for severity 1). For
the severities affected, the proportionate changes varied by severity and diminished
with increasing OT, with zero reduction above some OT threshold. Thus (18.34) for an
individual observation reads
(Lh)
yi = exp (Lh)
X
+ i ,
mi
imi
(18.51)
474
with
(Lh)
Xim
= max 0, hki ,mi Tmi hki ,mi timi ,
i
(18.52)
and where ki denotes the accident quarter to which the i-th claim belongs, hki ,mi the
distortion function for severity mi in accident quarter ki (= identity function for
mi = 1), and Tmi the OT threshold for severity mi .
Discussion. The model just described contains 73 parameters. It is unlikely that the
information contained in the 17 17 triangles of which Table 18.5 and Table 18.6
are excerpts (153 observations per triangle) would be sufficient for the reliable estimation of all parameters. Possibly the information would be insufficient even for the
formulation of some of the effects described.
This situation would be considerably eased by the use of quarterly data, which
are available because insurers lodge data quarterly. This would convert the 17 17
triangles to 68 68 triangles (2,346 observations per triangle). Even in this case,
however, fine gradations of OT would be lacking. As can be seen from Table 18.6,
OT increases by 2030% per annum in the early development years, which implies
increases of about 57% per quarter. The modeling described earlier was based on
2% divisions of OT.
In short, the use of individual claim data (>125,000 observations per triangle)
creates a richness of information that facilitates both the detection of data features
and their subsequent modeling.
18.4.3 Stochastic Case Estimation
Stochastic case estimation (or stochastic case estimate, as the context requires) (SCE)
consists of the formulation of the estimated ultimate claim cost of individual claims
on the basis of a predictive model applied to information specific to those claims. In
other words, it is a mapping, S : Ii ci , where Ii denotes the total information in
respect of the i-th claim and ci the estimated ultimate claim cost of that claim.
The predictive model is stochastic. By definition, the model underlying SCE is an
individual claims model. In fact, there is nothing in the definition that is any more
specific than the individual claims models discussed in Section 18.4. However, the
name SCE is usually reserved for models that employ detailed data specific to the
individual claim.
The SCEs are intended to function as case estimates (i.e., as substitutes for manual
or physical case estimates, estimates of ultimate cost made by claims experts after
consideration of all the facts and circumstances of the claims).
The more data specific to a claim used as input to the SCE (subject to significance
and the generally sound construction of the predictive model), the more accurate those
475
estimates can be made and the greater will be the potential dispersion of the SCEs,
taken over the whole body of claims. An SCE that predicts the same claim size for
every claim is of little value.
Examples of these SCE models are found in Brookes and Prevett (2004), Taylor and
Campbell (2002), and Taylor, McGuire, and Sullivan (2006). The first two of these
relate to Australian workers compensation insurance and the last to U.S. medical
malpractice.
Taylor et al. (2006) classify explanatory variables into a number of categories:
r Static variables constant over the life of a claim (e.g., date of birth of claimant);
r Dynamic variables: liable to change over the life of a claim, which consist of two
subcategories:
1. Time variables: directly related to just the passage of time, whose future values are
therefore fully predictable (e.g., development period);
2. Unpredictable variables: future changes not predictable with certainty (e.g., development period of claim closure).
The last category, unpredictable variables, requires special treatment. If they are
included in a model, then any forecast of the model will require forecasts of these
explanatory variables. The advantages and disadvantages of the inclusion of unpredictable variables were discussed in relation to the timing of claim closures in Section 18.3.9.
Example 18.2 (Australian Workers Compensation Insurance). Workers compensation claims are characterized by periods of incapacity, during which the
insurance provides income maintenance, and spells of activity, when income
maintenance is not required.
In Taylor and Campbell (2002), for example, income maintenance is modeled on
the basis of a stochastic process of claim status (incapacitated or active). A spell of
incapacity is modeled as a survival process in which the hazard is recovery to an
active status. Likewise, a spell of activity is modeled as a survival process in which
the hazard is relapse to incapacity.
The hazard rates of the survival processes depend on many explanatory variables
specific to the claim. Brookes and Prevett (2004) categorize these as follows:
r Claimant characteristics: age, gender, occupation, marital and dependant status, wage
rate, etc.
r Employer characteristics: industry, wages, location, etc.
r Claim status: claim is open/closed/reopened/disputed, work status, etc.
r Claim characteristics: injury nature, location, etc.
r Claim history: payments and rates of payment, time lost, etc.
476
The hazard rates of Taylor and Campbell (2002) are modeled by means of survival
analysis (Kalbfleisch and Prentice 1980), specifically proportional hazards survival
models in which the probability of survival of the i-th claim to time t + dt, conditional
on survival to t, is, 1 i (t) dt where the hazard rate i (t) takes the form
i (t) = 0 (t) exp( xiT ).
(18.53)
Yes
I=0?
477
No
Model 1
F(U|I=0,U>0)
Model 2
F(R|I>0,U>0)
Estimate of
F(U|I=0)
Estimate of
F(R|I>0)
Model 3
Prob[U=0]
478
The dataset includes a history of case estimates for each of its claims, so that
each closed claim generates multiple observations of the development of the case
estimate from one epoch to the ultimate claim cost. These, however, are unlikely to
be stochastically independent, as assumed by a GLM, and so the dataset modeled
includes just a single observation from each claim, randomly selected from those
available.
(18.54)
The ultimate number of claims mk notified in respect of each accident year, and
hence the number of IBNR claims, is often estimated by means of the chain ladder
(Section 18.2) with ykj denoting the number of notifications in cell (k, j ) or some
k , n k respectively, then m
k = xk,Kk+1 + n k .
variant of it. If mk , nk are estimated by m
Various approaches to the estimation of the average claim size in (18.55) are
possible and a couple are discussed next.
18.4.4.1 Estimation by Reference to Aggregate Models
One simple possibility consists of the application of an aggregate model to form an
estimate u k of the ultimate claims cost uk in respect of each accident year k. This
might, for example, be one of the models discussed in Sections 18.2 and 18.3. Then,
ak is estimated by
a k =
u k
.
k
m
(18.55)
References
479
The estimate of uk can now be discarded and only n k , a k retained for insertion in
(18.55). This approach has the advantage of simplicity, but may be involve an implicit
assumption that the average size of IBNR claims is the same as for claims already
notified. This may or may not be true.
18.4.4.2 Estimation by Reference to Individual Models
The possibly unwarranted assumption just referred to may be eliminated by a more
detailed approach, but at the cost of an increase in complexity.
Let the number of IBNR claims in respect of accident year k be estimated by
n k just as earlier, but now assume that its components ykj , j = K k + 2, . . . , J
c
, draw a random sample of Nkj claims from past
are available. For each (k, j ) DK
accident years that were notified in development year j .
The number of past accident years would need to be sufficiently large to yield a
sample of at least Nkj . Subject to this, more recent accident years would usually be
preferred over earlier ones as being more relevant to forecasts.
Let skj i , i = 1, 2, . . . , Nj k denote the ultimate claim size of the i-th of these Nkj
claims, estimated by skj i , comprising all claim payments up to the valuation date and
the SCE of any remaining liability. Some of these claims may have been closed so
that their sizes may be treated as known with certainty (skj i = skj i ), provided that the
likelihood of their reopening is small.
The average size of claims of accident year k notified in development year j may
then be estimated by
Nkj
skj i
a kj = i=1
,
(18.56)
Nkj
and the quantity ik in (18.55) is estimated by
ik =
J
ykj a kj .
(18.57)
j =Kk+2
References
Brookes, R. and M. Prevett (2004). Statistical case estimation modelling an overview of the
NSW WorkCover model. Institute of Actuaries of Australia Xth Accident compensation
Seminar. http://www.actuaries.asn.au/Library/Accident
\%20Compensation\%20Seminar\%20SCE\%20Paper\%20RGB\%2022Nov04
\%20final.pdf.
Cox, D. R. (1972). Regression models and life tables. Journal of the Royal Statistical Society,
B 34, 187220.
De Jong, P. and B. Zehnwirth (1983). Claims reserving state space models and the Kalman
filter. Journal of the Institute of Actuaries 110, 157181.
480
19
Survival Models
Jim Robinson
Chapter Preview. Survival modeling focuses on the estimation of failure time distributions from observed data. Failure time random variables are defined on the
non-negative real numbers and might represent time to death, time to policy termination, or hospital length of stay. There are two defining aspects to survival modeling.
First, it is not unusual to encounter distributions incorporating both parametric and
nonparametric components, as is seen with proportional hazard models. Second, the
estimation techniques accommodate incomplete data (i.e., data that are only observed
for a portion of the time exposed as a result of censoring or truncation). In this chapter,
we apply Rs survival modeling objects and methods to complete and incomplete data
to estimate the distributional characteristics of the underlying failure time process. We
explore parametric, nonparametric, and semi-parametric models; isolate the impact
of fixed and time-varying covariates; and analyze model residuals.
So, the distribution of y can be defined in terms of the density function, the survival
function, or the hazard function. These functions may be continuous, discrete, or a
mixture. They may also be parametric, nonparametric, or a mixture.
481
482
Survival Models
Two aspects of the 1999 NNHS have been simplified to provide tractable examples of survival modeling. First, the
stratified sampling structure of the NNHS implies that each resident has a different probability of being sampled.
The survey provides weights inversely proportional to the sampling probability for the resident. To eliminate the
need to work with weighted observations, we extracted a new sample of records from the NNHS current and
483
All current resident observations are right-censored at the nursing home duration
recorded at the end of the one-year survey period. Current residents who are more than
one year since admission are also left-truncated at the censoring duration minus 365
days. Current residents who are within one year of admission are not left-truncated.
Discharged residents are not right-censored. They are, however, left-truncated if the
simulated entry duration is greater than zero. The population diagram in Figure 19.1
summarizes the survey observation period and four possible censoring/truncation
situations.
r At the start of the study year, Resident A is 0.5 years from admission and remains
in the nursing home for an additional 0.4 years. This is an example of a lefttruncated uncensored observation. The contribution to the likelihood function should
be h(0.9)S(0.9)/S(0.5).
r Resident B enters the study at duration 0.3 years and remains in the nursing home
throughout the study year. This is an example of a left-truncated right-censored observation with a likelihood contribution equal to S(1.3)/S(0.3).
discharged resident files in which the probability of selecting a record is proportional to the survey weight. The
resulting resampled records can be considered equally weighted in that each record represents the same number of
nursing home residents (or discharges) in the general population. Second, the discharge records in the NNHS do
not indicate the date of discharge in the year leading up to the survey. The duration since admission is provided
(in days), but we do not know when each discharged resident entered the one-year study period. All we know is
that the duration at which each residents exposure started was within 365 days prior to the recorded duration at
discharge. To create a more conventional left-truncation example for this chapter, we randomly generated a starting
duration for each discharge record from a uniform distribution on the interval from the greater of zero and the
discharge duration minus 365 days to the discharge duration. Building estimation techniques appropriate for the
original NNHS data (i.e., for weighted samples with interval-censored entry durations) from basic principles, would
unnecessarily complicate the presentation of survival modeling concepts.
484
Survival Models
r Resident C is admitted to the nursing home 0.3 years after the start of the study
and is discharged 0.4 years later. This is a complete (uncensored without truncation)
observation with a likelihood contribution of h(0.4)S(0.4).
r Resident D is admitted 0.6 years into the study and remains through the remainder of
the study period. This is a nontruncated right-censored observation with a likelihood
contribution of S(0.4).
Throughout the remainder of this chapter, we use R to fit and assess a variety of
survival models to this nursing home length of stay data. The resampled NNHS data
were assembled using SAS and imported into an R data object named nnhs using the
following code.
> library(foreign)
> nnhs<-read.xport("transport")
Next, we loaded Rs OISurv object library, set selected NNHS variables as categorical using Rs factor method, and created filtering variables for later use.
> library(OIsurv)
> attach(nnhs)
>
>
>
>
>
>
>
>
>
>
>
nnhs$CERT<-factor(CERT)
nnhs$FACID<-factor(FACID)
nnhs$RESID<-factor(RESID)
nnhs$LIVLOC<-factor(LIVLOC)
nnhs$LIVWITH<-factor(LIVWITH)
nnhs$LOC<-factor(LOC)
nnhs$MARRIED<-factor(MARRIED)
nnhs$MSA<-factor(MSA)
nnhs$OWNER<-factor(OWNER)
nnhs$PAYOR<-factor(PAYOR)
nnhs$SEX<-factor(SEX)
>
>
>
>
smpl<-(runif(length(nnhs$ENTRY))<0.01)
nnhs$SMPL<-smpl
nnhs$ADM<-(ENTRY==0)
nnhs$T18<-(PAYOR==4)
> detach(nnhs)
> attach(nnhs)
485
ADM is another vector of indicators for observations that are observed from admission (i.e., records that are not left-truncated).
T18 is a vector of indicators for observations where the subject was a Medicare
resident before discharge or on the survey date.
Rs Surv method creates a survival data object formatted for use by other survival
modeling methods. The code below creates the survival data object nnh surv, using
variables ENTRY, DAYS, and DISCHG.
> nnhs_surv<-Surv(ENTRY,DAYS,(DISCHG!=0))
ENTRY is the duration at which exposure starts. DAYS is the duration at which
]
]
]
]
]
]
]
( 2,16
( 2,16
( 0, 3
( 0,14
( 0, 5
( 0, 5
(17,32
]
]
]
]
]
]
]
( 2,16
( 0, 3
( 0, 3
( 0,14
( 0, 5
(17,32
( 0,13
]
]
]
]
]
]
]
( 2,16
( 0, 3
( 0,14
( 0,14
( 0, 5
(17,32
( 0,13
]
]
]
]
]
]
]
( 2,16
( 0, 3
( 0,14
( 0,14
( 0, 5
(17,32
( 0,13
]
]
]
]
]
]
]
We see that only 2 of the first 100 observations are right-censored. Applying the
summary method, we obtain some quick sample statistics.
> summary(nnhs_surv)
start
stop
Min.
:
0.0
Min.
:
1.0
1st Qu.:
0.0
1st Qu.:
20.0
Median :
0.0
Median : 102.0
Mean
: 373.4
Mean
: 519.6
3rd Qu.: 304.0
3rd Qu.: 607.0
Max.
:11695.0
Max.
:12060.0
status
Min.
:0.0000
1st Qu.:0.0000
Median :1.0000
Mean
:0.6053
3rd Qu.:1.0000
Max.
:1.0000
We see that roughly 61% of the observations are discharges and 39% are censored.
We also note that more than half of the subjects are observed from admission.
486
Survival Models
1
9
tj t
dj
Rj
t < t1 .
t t1
Greenwood (1926) derived the formula for the variance of the product-limit
estimator:
dj
2
.
V
ar(S(t))
= (S(t))
Rj (Rj dj )
t t
j
Rs survfit method takes a survival data object and creates a new object containing the Kaplan-Meier estimate of the survival function along with confidence
intervals. The code below demonstrates this process using NNHS subjects observed
from admission. Note that the Surv method is applied without the ENTRY field when
the observations are not left-truncated. The ~1 tells the method to fit a single combined
survival function. The subsequent print method provides summary statistics for the
fitted survival curve.
> nnhs_KM<-survfit(Surv(DAYS,(DISCHG!=0))~ 1, subset=ADM, data=nnhs)
> print(nnhs_KM,rmean="common")
Call: survfit(formula = Surv(DAYS, (DISCHG != 0)) ~ 1,
data = nnhs, subset = ADM)
records
9016.00
n.max
9016.00
n.start
9016.00
events
6474.00
*se(rmean)
median
0.95LCL
0.95UCL
1.56
30.00
29.00
32.00
* restricted mean with upper limit = 365
*rmean
103.95
487
0.6
0.4
0.0
0.2
survival function
0.8
1.0
50
The results include the restricted mean of 104 days (i.e., the area under the fitted
survival curve from admission through 365 days). The median length of stay is 30
days, significantly less than the mean and consistent with a skewed distribution.
The plot method is used to graph the fitted survival function, including the 95%
confidence interval at each duration.
> plot(nnhs_KM, main="Kaplan-Meier estimate - Admissions only",
xlab="days", ylab="survival function")
The resulting graph in Figure 19.2 indicates that about 17% of admissions remain
in the nursing home for at least one year. We next expand the sample to include
left-truncated residents.
> nnhs_KM<-survfit(Surv(ENTRY,DAYS,(DISCHG!=0))~ 1, data=nnhs)
> print(nnhs_KM,rmean="common")
Call: survfit(formula = Surv(ENTRY, DAYS, (DISCHG != 0)) ~ 1,
data = nnhs)
records
15127.00
n.max
9016.00
n.start
0.00
events
9157.00
*se(rmean)
median
0.95LCL
0.95UCL
5.59
30.00
29.00
32.00
* restricted mean with upper limit = 12060
*rmean
229.09
Survival Models
0.6
0.4
0.0
0.2
survival function
0.8
1.0
488
500
0.6
0.4
0.0
0.2
survival function
0.8
1.0
The restricted mean is now 229 days, representing the area under the survival curve
from admission to 12,060 days (the largest observed duration). The median remains at
30 days, because much of the additional left-truncated information relates to resident
survival beyond the first year. Figure 19.3 displays the fitted survival curve through
3,000 days, along with a 95% confidence interval.
50
100
150
200
250
300
350
days
489
We next fit the survival curve to a 1% sample of those observed from admission.
n.max
106.0
n.start
106.0
events
77.0
*rmean
94.2
*se(rmean)
median
0.95LCL
0.95UCL
13.5
33.0
20.0
43.0
* restricted mean with upper limit = 343
> plot(nnhs_KM, main="Kaplan-Meier estimate - Admission Sample",
xlab="days", xmax=400, ylab="survival function", conf.int=TRUE)
In Figure 19.4 we see that, although the confidence interval has significantly
increased, the estimated survival curve and restricted mean are quite similar to their
counterparts from the complete sample.
490
Survival Models
Aalen 1978) estimate of the cumulative hazard function to obtain an estimate of the
survival function. The estimated cumulative hazard starts at zero and is incremented
at each observed event duration by the number of events divided by the number at
risk. The second alternative (type=fh2) handles ties, in essence, by assuming
that multiple events at the same duration occur in some arbitrary order. So, if 3 of
10 nursing home residents are discharged at the same duration, rather than increment
the cumulative hazard by 3/10, we increment by 1/10 + 1/9 + 1/8 at that duration.
Note that when this ordering of tied results is applied to the Kaplan-Meier estimate
of the survival function it no effect. In that case, we multiply the survival function
by 7/10 (allowing for ties) or by 9/10 8/9 7/8 = 7/10 (arbitrarily ordering the
tied discharges). The following code and Figure 19.5 plot the estimated cumulative
hazard function for all three options.
Even with the small sample, we can see in Figure 19.5 that the three options
yield similar estimates. Note that the plot also displays the 95% confidence interval for the Kaplan-Meier estimate based on Greenwoods estimate of the survival
variance, taking into account the log transformation needed to obtain the cumulative
hazard function. This confidence interval is a pointwise estimate specific to each
duration along the curve. We cannot say that the corridor formed by the interval
will capture the entire hazard curve 95% of the time. To obtain such a simultaneous
confidence band, we employ the confBands method. In the following example, we
plot both the pointwise and simultaneous 95% confidence intervals for the survival
function.
491
1.0
0.5
KaplanMeier
NelsonAalen
FlemingHarrington
0.0
Cumulative Hazard
1.5
50
100
150
200
250
300
350
Days
As expected, Figure 19.6 shows that the simultaneous confidence interval is wider
than the pointwise counterpart. Also note that we have used the log-log transformation option to construct the confidence intervals, which assumes that the log
of minus the log of the survival function estimate is approximately normal. Note
that the confBands method requires the OIsurv package and does not work with
left-truncated data.
We next compare the nonparametric estimates of survival for subsets of nursing
home residents (e.g., by gender, type of care, and payer). The survfit method
allows us to stratify the analysis by specifying one or more categorical variables after
the survival object. The first example constructs the Kaplan-Meier estimates of the
survival curves for male versus females residents based on those admitted during the
year prior to the survey.
Survival Models
1.0
492
0.6
0.4
0.0
0.2
Survival Function
0.8
KM Estimate
Pointwise Conf. Interval
Simultaneous Conf. Interval
50
100
150
200
250
300
350
Days
SEX=1
SEX=2
0.95LCL 0.95UCL
SEX=1
23
26
SEX=2
33
38
* restricted mean with upper limit =
365
The print method provides sex-distinct estimates of the median and restricted
mean time to discharge (using a common upper limit of 365 days for both males
and females). The plot method graphs the Kaplan-Meier survival curves. From
493
Figure 19.7 it is evident that females are more likely than males to remain in the
nursing home for at least one year.
The following example shows survival estimates for combinations of care level
(skilled and intermediate care) and payment source (Medicare and Medicaid), again
for new admissions during the year prior to the survey.
> nnhs_KM<-survfit(Surv(DAYS,(DISCHG!=0))~ PAYOR+LOC,
subset=(ADM&((PAYOR==4)|(PAYOR==5))&(LOC!=3)&(LOC!=4)),
data=nnhs)
> print(nnhs_KM,rmean="common")
Call: survfit(formula = Surv(DAYS, (DISCHG != 0)) ~ PAYOR + LOC,
data = nnhs, subset = (ADM & ((PAYOR == 4) | (PAYOR == 5))
& (LOC != 3) & (LOC != 4)))
PAYOR=4,
PAYOR=4,
PAYOR=5,
PAYOR=5,
LOC=1
LOC=2
LOC=1
LOC=2
364
> sum(nnhs_KM$time*(nnhs_KM$n.event+nnhs_KM$n.censor))
[1] 407843
> sum(nnhs_KM$n.event)
[1] 4465
> plot(nnhs_KM, main="KM Estimates by Payor and Level of Care
All Admissions", xlab="Days", ylab="Survival Function",
lty=1:4, mark.time=FALSE)
> legend("topright",c("Medicare-Skilled","Medicare-Interm",
"Medicaid-Skilled","Medicaid-Interm"),lty=1:4)
We see in Figure 19.8 that first-year discharge rates vary significantly. Medicare
residents have shorter nursing home stays than Medicaid residents, and within each of
these payment sources, skilled-care residents have shorter stays than intermediate-care
residents.
Survival Models
1.0
494
0.6
0.4
0.0
0.2
Survival Function
0.8
Male
Female
50
1.0
Plotting the log-log Kaplan-Meier survival function (or the log of the cumulative
hazard function) versus the log of time can be used to assess two assumptions. First,
if discharges follow an exponential distribution, the curve should be linear with unit
slope. Second, if discharges for different resident populations satisfy the proportional
hazards assumption (to be discussed later in this chapter), then the curves will be
parallel. The next example focuses on new admissions of Medicare residents.
0.6
0.4
0.0
0.2
Survival Function
0.8
MedicareSkilled
MedicareInterm
MedicaidSkilled
MedicaidInterm
50
495
1
2
3
Ln(H(t))
Skilled Care
Intermediate Care
Residential Care
10
20
50 100 200
The results in Figure 19.9 indicate that neither the exponential assumption (i.e., constant hazard rate) nor the proportional hazards assumption appears to be appropriate
for this population.
496
Survival Models
The log-rank p-value of 49% indicates that we cannot reject the null hypothesis
that males and females have the same discharge rates. By rerunning the example with
rho=1, the reader can confirm that p-value of the Wilcoxon test is 55%.
With these tests, we can explore subpopulations of residents based on, say, combinations of gender, age, level of care, payment source, and ADLs impaired (i.e., the
number of activities of daily living for which assistance is required). The party
library offers an automated tool to perform this search: the conditional inference tree,
using the ctree method. This method does not work with left-truncated data. The
following example demonstrates the results for new admissions in the year prior to the
nursing home survey. The plot output in Figure 19.10 provides a convenient visual
display of the results.
> library(party)
> nnhs_KM_ctree <- ctree(Surv(DAYS,(DISCHG!=0))~ SEX+LOC+PAYOR
+AGE+ADL, subset=ADM&(LOC!=4)&((PAYOR==4)|(PAYOR==5)),
data=nnhs)
> #Define a larger graphics page
> windows(width=25,height=10,pointsize=8)
> plot(nnhs_KM_ctree, main="Conditional Inference Tree
- All Admissions")
The method looks for the most significant split first, in this case based on Medicare
versus Medicaid payment. Each population is then split again. The process continues
until no significant split remains. This results in a partitioning of residents showing
the estimated survival curve and the number of residents in each grouping. Note that
the p-values are specific to each split decision (node) and that the probability of a
Type 1 (or Type 2) error accumulates throughout the process.
497
1
PAYOR
p < 0.001
4
16
ADL
p = 0.006
20
SEX
p < 0.001
2
LOC
p < 0.001
1
3
ADL
p < 0.001
0
4
SEX
p = 0.007
2
5
AGE
p = 0.045
{2, 3}
>0
1
21
AGE
p = 0.012
9
SEX
p < 0.001
11
AGE
p = 0.04
89
77 > 77
19
LOC
p < 0.001
{1, 3}
27
AGE
p < 0.001
24
AGE
p < 0.001
2 >2
72 > 72
> 89
42 > 42
13
ADL
p = 0.018
74 > 74
4 >4
Node 7 (n = 127)
Node 10 (n = 1262)
Node 14 (n = 277)
Node 17 (n = 118)
Node 22 (n = 62)
Node 25 (n = 218)
Node 28 (n = 315)
Node 6 (n = 77)
Node 8 (n = 127)
Node 12 (n = 1654)
Node 15 (n = 41)
Node 18 (n = 272)
Node 23 (n = 485)
Node 26 (n = 559)
Node 29 (n = 67)
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
n
9
i=1
n
9
n
9
i=1
i=1
Left-truncation simply requires division by the survival function at the truncation point for each individual. This leads to a difference in the baseline cumulative
hazard in the last term. Conventional maximum likelihood estimators are not available because the baseline hazard function is not parametrically defined. Cox (1972),
Breslow (1974), Efron (1977), and Kalbfleisch and Prentice (2002) offer a variety of
estimation techniques designed to maximize this likelihood. The techniques vary in
how discrete baseline probabilities are handled, especially with regard to ties in the
observed event times. In all cases, estimations of the baseline hazard function and
the regression coefficients are performed separately. Inspection of the likelihood as
a function of the baseline hazard function (fixing the regression coefficients) suggests minimizing the function at all durations except those at which failure events are
498
Survival Models
observed. The terms of the log-likelihood can be arranged to isolate each discrete jump
in the baseline cumulative hazard function. The components can be differentiated and
set to zero to find optimal choices for each discrete baseline jump (as functions of the
regression coefficients). These optimal jump values can then be substituted into the
likelihood to obtain a partial likelihood function that depends only on the regression
coefficients. Breslows technique leads to the following partial likelihood function:
%1i
$
n
"
exp(xi )
LP () =
j R(yi ) exp(xi )
i=1
where R(y) is the risk set at duration y (i.e., those individuals exposed at duration y).
With regression coefficients that maximize the partial likelihood, a variety of techniques have been suggested to estimate the baseline hazard function. Many of these
baseline estimators can be considered variations of the product-limit or Nelson-Aalen
estimators in which each exposed individual is assigned a weight proportional to his
or her hazard multiplier. Again, the variations reflect different approaches to handling
ties and discrete baseline jumps.
The coxph method allows the user to specify one of three methods to estimate
the regression coefficients; Breslow (method=breslow), Efron (method =
efron), and an exact method assuming a discrete logistic form for the hazard
function (method=exact). The default is the Efron method and is used for all
of the examples in this chapter.
The following example fits a proportional hazard model to all the Medicare residents
against gender, level of care, and ADL impairment count.
> nnhs_PH <- coxph(Surv(ENTRY,DAYS,(DISCHG!=0))~ SEX + LOC + ADL,
subset=T18, data=nnhs, weights=, x=TRUE, method="efron")
> summary(nnhs_PH)
Call:
coxph(formula = Surv(ENTRY, DAYS, (DISCHG != 0)) ~ SEX + LOC +
ADL, data = nnhs, subset = T18, method = "efron", x = TRUE)
n= 4938, number of events= 4088
coef exp(coef) se(coef)
z Pr(>|z|)
SEX2 -0.17458
0.83981 0.03239 -5.389 7.07e-08
LOC2 -0.71314
0.49010 0.05647 -12.628 < 2e-16
LOC3 -0.86208
0.42228 0.18500 -4.660 3.16e-06
LOC4 0.16646
1.18112 0.21470
0.775
0.438
ADL -0.05809
0.94356 0.01139 -5.099 3.42e-07
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1
***
***
***
***
1
SEX2
LOC2
LOC3
LOC4
ADL
499
p=0
p=0
p=0
The summary method displays the regression coefficient estimates, standard errors,
and p-values. The model indicates that female discharge rates are 84% of male rates;
that intermediate-care and residential-care discharge rates are 49% and 42% of skilledcare rates, respectively; and that discharge rates decrease roughly 6% for each ADL
impaired. This assumes that the proportional hazard and other model assumptions are
correct. All three of the tests for model significance reject the null hypothesis that all
the regression coefficients are zero.
To test the proportional hazards assumption, we can use the cox.zph method to
assess the constancy of the coefficients over time. The method employs rescaled
Schoenfeld residuals to generate and plot local regression coefficient estimates over
time. If the proportional hazards assumption is correct, plots of residuals over
time should be horizontal. Correlation coefficients between the residuals and observation time are used to compute test statistics for each coefficient. The following example shows results for the proportional hazards model fit in the previous
example.
500
Survival Models
Medicare Stays ADL Beta(t)
3.7 14 25 61
Time
10
3.7 14 25 61
Time
3.7 14 25 61
Time
We see that the p-value for the null hypothesis that the female gender effect is
uncorrelated with time is 16%, supporting the proportional hazards assumption. In
contrast, there is a significant correlation of the ADL effect with time because the
p-value is less than 0.1%. The plots in Figure 19.11 show the residual patterns over
time for the female gender, intermediate-care, and ADL effects.
501
Each plotted point can be considered the contribution of an observed failure (discharge) to the estimated regression coefficient. The fitted moving average and confidence interval provide a visual indication of the varying effect of each factor over time.
We see that ADL impairment at early durations is associated with reduced discharge
rates, whereas the reverse is true at later durations. This may be due to the increasing
role of mortality as the reason for discharge at later nursing home stay durations.
502
Survival Models
3
2
1
Cumulative Hazard
Medicare Stays
Male Cumulative Hazard
Baseline (Skilled
Skilled ADL=1
Skilled ADL=5
Intermediate ADL
Intermediate ADL
500
1000
1500
2000
Days
impairment is associated with lower discharge rates. Recall, however, that the early
ADL effect was found to reverse at later durations when we allowed the regression
coefficients to vary over time.
19.5.2 Residuals
A variety of residuals are available to assess the proportional hazards model assumptions. The Cox-Snell residual for an observation is the estimated difference in the
cumulative hazard function from entry to exit. If the model is correctly specified,
these residuals will behave like a censored/truncated sample from a unit exponential
survival process. So, a plot of the Nelson-Aalen estimator of the cumulative hazard
function for the Cox-Snell residuals versus the residuals should be a straight line with
unit slope.
The residuals method applied to a coxph object allows the user to select several
residual types. Although the Cox-Snell residuals are not immediately available, a
closely related residual, the martingale residual, is an option. The martingale residual
is defined as the difference between an event indicator and the Cox-Snell residual
for each observation. If the model is correctly specified, the martingale residuals can
be considered the difference over time of the actual number of failures minus the
expected number of failures. As with conventional regression residuals, plots of the
martingale residuals versus a covariate can be used to determine the appropriateness
of the functional form used in the hazard multiplier.
+
+
+
+
+
+
+
+
Martingale Residual
2
1
0
+
+
++
+
++
+
+
+
+
+
+
+
+
+++
+
+
+
++ + ++
++
+++
++
+
+++ +
++
++
+
+++
+++
++
+++
+++
+
+
+ +
+++++
+
+
+
++
++
+
+
+
++
+
+
+
+
++
++++
++ +
++
+
+
+++
+ ++
+ +
++ +
+ +
+
+
+
++ + +
+
++
+ +++ +
+
++
+
+
+
+
+
+++
+
++
+ +
+
+
+
+
+
+
+
++
+
+
+++ + +
+
+ ++
+++
++
++++
+
+
+
+
+
++ ++ ++
+ +++ +
++
+
+
+ + + +
+
+
+
+
+
+
+
++
+
+
+ ++++
+++
+
+
++ ++
+
+
++
+
+
+
+
+
+
+
++
++
+
+
+++
+
+
+
+
+++
+
++
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
++
+
+
+
++
+
+
+
+
++
+
+
+
++
++
++
++
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
++ +
+
++ +
+ +
+ ++
+
+ +
+ ++
+
+
+
+ +
+
+
+ +
+ +
+
+
+ +
+
+
+
+ +
+
+
++
+
++
+
+
+
++
+
+
+
+
++
+
++
++
++
+
+
+
++
+
+
++
++
+
++
+
++
+
+
++
++
+++
+
+++
+ +
++
+++
+
+ +
+++
+
+
+
+
+
+
+++
+
++
+
+
++
++
+
+
+
+ +
+++
++
++
++
+
++
++
++
+
+
+++
+
+
+ +
+
+
+
+
+
+
+
++
+
++
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
++
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
++
+
++
++
+
+
+
+
+
++
++
+
+
+
+
+
+
+
+
+
+
+
++
+
++
+
++
+
+
+
++
++
+
++
+
+
+
+
+
+
+
+
+
+
+
+
+
++
+
+
+++
++
+
+
+
+
80
+++
+
+
+
+ +
++
++
++
+
+ +
++
+
++
+
++
+
+
+
+
+
++
+
++
+
++
+
+
+
+++
++
+
+
++
+
+
++
++
+++
+++
++
++ +
+
+
+ +
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
++
+
+
+ +
+ ++
+
+
+
+
+
++
+
+
+
+
+
+
+
+
+
+
++++
++
++
++
+
+
+++
++
+++
+
+
+++
++
++
+
++
+
++
+
+
+
+
++
++
+++
+
+
++
+
++
+++
+
+
+ +
+
++++
++++
+++
+
+++
+
+
+ +
++
+
+
+
+ ++
+++
+
+++
+
+
+
+
++
+
++++
+
+
+ ++
+
++
++
+
++
+
+ ++
+
+ +
+++
++
++
+
+
+++
++
+ +
+
+
+
+
+
+
+
+
+
+
++++
++
++
+
+
+
+ ++
++ +
+ ++
++
+ +
+
+
++++
+
+
+
+
60
+
++
+
+++
++
+
+
++
++
++
+
++
+
+++
++++
+++
++
+
++
++
++
+
+
++
+
+
+
+++
++
+++
++
+++
++
++
+
++
+
++
+
++
++
+++
++
+
+++
++
+++
++
+
+
+
+
++ +
++++
+ +
+
+
++
+
+
+
+ +
+
+++
++
+
++ +
++++
+
++++
+
+
++
+++
++++
++
+
+
+
++
++
+ ++
++ +
+ ++
+
++
+++
+
++
++
+
++
+
++
+
++++
+
+
+
+ +
+
++++
++
+
+ +
++
+
+
+
+
+ +
+
++
+
+
+
40
+
+
+
++
+
+
+
++
+
++
++
+
+
+
++
+
+
++
+
+
+
++
+
+
+
+
+
+
+
+
++
++
+
+
++
++
++
+
+
++
+
+
+
+
++
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
++
++
++
+
+
+
+++
+
++
+
++
++
+
+
++
++
+
++
++
+
+
++
++
++
+
+
+
+ +
+ +
+
++
+
++
+
+++
+
+
+++
+
++
++
++
+
+
++
++
+
+
++
++
+++
+
+
+
+ +
503
+
++
+
+
100
Age
The following example plots martingale residuals for Medicare data versus age, a
variable not included in the model.
The plot in Figure 19.13 indicates that actual discharges at very young and very old
ages tend to be somewhat less than expected by the fitted model. The data are sparse
at these extremes, however. So, adding age to the model appears to be unnecessary.
While the martingale residuals have mean zero if the model is correctly specified,
the distribution about zero is highly skewed, making outlier detection difficult. A
standardized version of the martingale residual, the deviance residual, attempts to
504
Survival Models
3
+
+
+ +
+
+
+
++
+ ++
++
+
++
+++
+++ +
+++++
++
+
++
+
++
++++ +
+
+
+++
+
+
+
+
+
+
+
+
+++
+ +++
+ +
+
+
+++
+
+
+
+
+
+
+
+
+
+++ ++ +++
+
++
+ +
+
+
+++++
++
+
+
+
+
+
+
+
+
++
+
+
+ ++
+
+
+ +
+
++ ++ +
+
+ +
+
+
+ + ++
++ + +
+++
++
+++
++
+++
++ + +
+
+ +++ +
++
+
+
+ + +
++
+ ++
+
+
+
+
++++
+++
+++
+
+
+
+
+
+
+ + ++
+ + +
+
+
+
+
+ + ++++ +
+
+ +
++
+ +
+
++
+ + +
+ +
+ +
+ ++
+ + +++
+
++++
++
+ +++
+ +
+++++
++++ + +
+
+++
+ +
++
+
++
+
+
+
+
+
+
+
+
+
+
+
++++
+++
+++++
+ ++
+ +++
+
+
+
+
++ ++++ +
++ + ++ +
+ ++
+ + ++
++
+++
+++
+++
++++++
++
+ ++
+
++++
++
+++ ++ ++
+
++ + +++
+
+++
+
++
++
++
+
++
+++
++
+ +
+ ++
+
+ +
+ +
+ + +
+
+++
++
++
+ ++
+
+ +
+
+++
++++
++
+
++++ +
++
+
++
+++
+++
+
+
++
++
++
++
+++
++
++
+
++
++
+++++
+++
+++
++
+
+
++
+++ +
+ +
+ +
+++
+
+
+
+
+
+
+
+
++
+
+
+
+
+
+
+
+
+
+
++++ ++
++ ++++++ ++ +
++
+
++
+
++
+
++
+++
++ +
++
++
++++
+
+ ++++
+ +
+++
++
++
+++
+++
++
++ +
++++
+
+
++
+++
+ ++++++
+
+++
++ +
++
+
+ ++
+
+
++
+
++
+
++++
+ +
++
++
++ +++
+
+++
+
+++
++
++
+
+++++ +++
++ ++ +
+++
++
++ ++ +++
+
+
+ +
++
+++
++
+ +
+
++
++++
+ + +
+
+++
+++++++
++ + + ++++
+++
++++
+
+
+
++++
+
+ +
++++
+++
+++ +++
++
+++
+ +
+
+
+
+
+
+
++ ++++++ ++
++
+++
+++++++++++
+ +
++
+++
+ +
+ +
+
+++
+++
+ +
+
+ +
++++
+
+
+
+
+
+
+++
++++
+
+
+
+
+
+ ++ +
+
+
++
+++++
++ +
++
+++++
+ +
++
+
+ +++
++
+
+ ++
+++
++
++
+
++
++
++ +
+
++
++
+
++
+ ++
+
++
++
+++
+++++++
++ +
+
+
+
++
++
++++ ++++
+ +
++
++ ++
++++
+
+++
+
++ + +++
+ ++
+++ +
+
+ +
+ +
+
+++
+
++
+ ++
++
++
++
++
+++
++++
+
+++
+ +
+
+
+
++
+++
+++
+ ++
++++
+
++++
++
+
++
+++
++
++ +
++ +
++
++
+
++++
+
+
+
+
+
+
+
+
+
+
+
+
+
+
++++++++++ +
+
++ +
++
++
+
++
++
++
++
++
++++
++++++
++
+ +
++
+
+
+
+
+
+
+
+
+
+
+
+
+
++
+
+
++
++ ++++++++
+
+ ++
+++
+++++
+
+
++++
++ +++ +
++++ +
++
+ ++++
++ ++
++
+
+++
+
+
++
+
+
+++
+
++++
+ +++
+ + ++
++
++
+ +
+
+
+++
+++
++ ++
+++ +
++
+ + +
++ + +
+
+++ +
+ +
+++
+
++
++
++
+++++++
+ + +
+
+
+
++++
++
+
+
+
+ +++++
+
++
+
++++ +
++ + +
+ +
++++
+
++ +++++
+
++
+ ++
+ +++
+ ++
+++ +
++
++
++
+
++
+
+
+
+
+
+
+
+
+
+
+ +
+
+
+
+
++++++
+ ++
+
+
++ ++++++++
++ ++
++ +
++
+
+++
+
+ +
+
+++ + ++++
+++
+ + +
++
+
+ + +
++
+
++ +
+ +
+++
++
++
+
+
+ +
++
+ +
++
+
+
+ +
+
+ + ++ ++
+
+
+
+
+
+ + +
+
++++
+
+
+
+ ++
+
+ +
+
+++
+ +
++ +
+
+
++
+++
+
+ ++ ++
+
++
+ +
+ +
+
+
+++
+++ +
+
+ +
+
+
+
++ +++++++
+
+
+
++
+ +
+
+
+ +
+
+
+
+
+
+
+++
+
+
+
+
+
+
+
+
++
+
+
+
+
+
+
+
++
+
+
+
++
+
+
+
+
+
+
+
+
Deviance Residual
+
+
+
+
+
+
+
+
+
++
+
+
+
+
+
40
60
80
+
+
100
Age
In Figure 19.14 we can see more clearly with deviance residuals than with martingale residuals which observations are outliers in both directions.
Several other residuals are available from the residuals method, including
scores residuals (score, derived from the observations contribution to the score
vector), rescaled score residuals (dfbeta, which estimate the impact of removing an
observation on each regression coefficient), rescaled dfbeta residuals (dfbetas),
Schoenfeld residuals (schoenfeld, a part of the score residual based on the actual
less expected values of covariates at observed failure times), and rescaled Schoenfeld
residuals (scaledsch, which are used in coxph.zph to generate local estimates of
the regression coefficients).
505
SEX2
LOC2
LOC3
LOC4
ADL
***
***
***
***
1
p=0
p=0
p=0
506
Survival Models
> beta=nnhs_PH$coeff
> va=nnhs_PH$var
>
> #Test ADL beta
> C <- matrix(c(0, 0, 0, 0, 1), nrow=1, byrow=TRUE)
> d <- rep(0, 1)
> t1 <- C %*% beta - d
> t2 <- C %*% va %*% t(C)
> XW2 <- c(t(t1) %*% solve(t2) %*% t1)
> pchisq(XW2, 1, lower.tail=FALSE)
[1] 3.418314e-07
We see that this results in the same p-value for ADLs as generated by the summary
method. To test the significance of all of the level of care coefficients, we apply the
following code.
> C <- matrix(c(0, 1, 0, 0, 0,
+
0, 0, 1, 0, 0,
+
0, 0, 0, 1, 0), nrow=3, byrow=TRUE)
> d <- rep(0, 3)
> t1 <- C %*% beta - d
> t2 <- C %*% va %*% t(C)
> XW2 <- c(t(t1) %*% solve(t2) %*% t1)
> pchisq(XW2, 3, lower.tail=FALSE)
[1] 2.072037e-38
So, we easily reject the null hypothesis that all of the level of care coefficients are
zero.
507
Call:
coxph(formula = Surv(ENTRY, DAYS, (DISCHG != 0)) ~ SEX + LOC +
ADL + tt(ADL), data = nnhs, subset = T18, method = "efron",
x = TRUE, tt = function(x, t, ...) (t * x))
n= 4938, number of events= 4088
coef exp(coef)
SEX2
-1.770e-01 8.378e-01
LOC2
-7.110e-01 4.911e-01
LOC3
-8.157e-01 4.423e-01
LOC4
1.533e-01 1.166e+00
ADL
-7.182e-02 9.307e-01
tt(ADL) 1.955e-04 1.000e+00
--Signif. codes: 0 *** 0.001 **
SEX2
LOC2
LOC3
LOC4
ADL
tt(ADL)
se(coef)
z Pr(>|z|)
3.240e-02 -5.462 4.70e-08 ***
5.650e-02 -12.585 < 2e-16 ***
1.850e-01 -4.410 1.03e-05 ***
2.148e-01
0.714 0.475490
1.196e-02 -6.007 1.89e-09 ***
5.677e-05
3.444 0.000573 ***
0.01 * 0.05 . 0.1
p=0
p=0
p=0
We see that the coefficient of the ADL-time interaction term is statistically significant. It is also evident that the ADL effect reverses sign after 367 days.
508
Survival Models
Unfortunately, the survreg method does not allow for left-truncated data. So, in
the example that follows, we limit the data to Medicare admissions in the year prior to
the nursing home survey. To assess the adequacy of each of the candidate parametric
distributions, we first compute and plot the Kaplan-Meier estimate of the cumulative
hazard function. The code that follows then fits each of the parametric distributions
and plots the cumulative hazard function on the same graph.
509
2.0
1.5
1.0
KaplanMeier
Exponential
Weibull
Logistic
LogLogistic
LogNormal
0.0
0.5
Cumulative Hazard
2.5
3.0
50
100
150
200
250
300
350
Days
In Figure 19.15 all of the parametric fits fail to match the concavity of the KaplanMeier hazard function. The log-logistic distribution does fit reasonably well over the
first 100 days and shows more concavity at later durations. So, we adopt this form for
the next example.
Prior proportional hazards modeling clearly indicates that gender, level of care, and
ADL impairment all have an impact of nursing home discharge rates. To incorporate
these effects into the log-logistic model, we employ a transformation of the time axis.
Specifically, we define S(t|z) = S0 (te z ). If z is large and positive, the survival
function is increased and the hazard rates are deflated. This is known as an accelerated
failure-time model.
The following example uses the survreg method to fit a log-logistic accelerated
failure-time model to new Medicare admission data with effects for gender, level of
care, and ADL impairment.
> nnhs_par_loglog2 <- survreg(Surv(DAYS,(DISCHG!=0))~ SEX +
LOC + ADL, data=nnhs, subset=(T18&ADM), dist="loglogistic")
> summary(nnhs_par_loglog2)
Call:
survreg(formula = Surv(DAYS, (DISCHG != 0)) ~ SEX + LOC + ADL,
data = nnhs, subset = (T18 & ADM), dist = "loglogistic")
Value Std. Error
z
p
(Intercept) 2.6244
0.0530 49.514 0.00e+00
510
SEX2
LOC2
LOC3
LOC4
ADL
Log(scale)
Survival Models
0.2098
0.8532
0.9713
0.2426
0.0781
-0.3570
0.0402
5.213 1.86e-07
0.0734 11.617 3.36e-31
0.2584
3.759 1.70e-04
0.4031
0.602 5.47e-01
0.0136
5.752 8.84e-09
0.0143 -24.899 7.60e-137
Scale= 0.7
Log logistic distribution
Loglik(model)= -15546
Loglik(intercept only)= -15652.9
Chisq= 213.71 on 5 degrees of freedom, p= 0
Number of Newton-Raphson Iterations: 3
n= 3967
We see that the signs of the regression coefficients are positive rather than negative (as they were in the proportional hazards model), because they act to increase
survival and decrease discharge rates. To compare the resulting fitted model to the
comparable proportional hazards model, we apply the following code. This plots the
fitted cumulative hazard functions for females using skilled care and having three ADL
impairments.
> xvals=data.frame(SEX=2,LOC=1,ADL=3);
xvals$SEX=as.factor(xvals$SEX);
xvals$LOC=as.factor(xvals$LOC)
> plot(survfit(coxph(Surv(DAYS,(DISCHG!=0))~ SEX+LOC+ADL,
data=nnhs, subset=(T18&ADM), weights=), newdata=xvals),
lty=1, fun="cumhaz", conf.int=FALSE, mark.time=FALSE,
main="Medicare Admissions - Log-Logistic Regression
Female Skilled Care ADL=3", xlab="Days",
ylab="Cumulative Hazard")
> lines(predict(nnhs_par_loglog2,newdata=xvals,type="quantile",
p=seq(.001,.999,by=.001)),-log(seq(.999,.001,by=-.001)),
lty=2)
> legend("bottomright",c("Cox Model","Log-Logistic Model"),
lty=1:2)
In Figure 19.16 we see that the resulting cumulative hazard functions are close for
the first 100 days. After 100 days, the log-logistic model generates consistently lower
survival rates.
511
Cumulative Hazard
19.8 Exercises
Cox Model
LogLogistic Model
50
100
150
200
250
300
350
Days
19.8 Exercises
Exercise 19.1. Use R to explore the contents of the nnhs data object. The Appendix
provides a brief description of the NNHS variables.
Exercise 19.2. Explore the structure of nnhs KM. Use R to compute the exposure in
resident-days for this sample using the time, n.event, and n.censor components
of the nnhs KM object.
Exercise 19.3. Compare the first-year survival functions of new admissions by level
of care for Skilled Care (LOC=1), Intermediate Care (LOC=2), and Residential Care (LOC=3) residents.
512
Survival Models
Exercise 19.6. Repeat the log-log survival plots for all Medicare residents (i.e., not
limited to new admissions) over the first 3,000 days in the nursing home.
513
Exercise 19.7. Confirm that the log-rank and Wilcoxon tests for the difference in
discharge survival for a 1% sample of skilled-care and intermediate-care admissions
both have p-values less than 0.5%.
Exercise 19.8. Use the following code to generate a boxplot of the Martingale
residuals versus ADL impairment count.
Exercise 19.9. Use the following code to generate the rescaled Schoenfeld residuals.
Also, verify that the correlation coefficients of the residuals with time are the same
as those generated by the cox.zph method in testing the constancy of the regression
coefficients.
514
Survival Models
Variable
Description
RESID
SEX
AGE
MARRIED
LIVLOC
LIVWITH
DAYS
LOC
PAYOR
MSA
OWNER
BEDS
CERT
DISCHG
FACID
ENTRY
ADL
SMPL
ADM
T18
References
Aalen, O. (1978). Nonparametric inference for a family of counting processes. Annals of
Statistics 6(4), 701726.
Breslow, N. (1974). Covariance analysis of censored survival data. Biometrics, 8999.
Cox, D. R. (1972). Regression models and life-tables (with discussion). Journal of the Royal
Statistical Society. Series B (Methodological), 187220.
Efron, B. (1977). The efficiency of Coxs likelihood function for censored data. Journal of
the American statistical Association 72(359), 557565.
Frees, E. W. (2010). Regression Modeling with Actuarial and Financial Applications.
Cambridge University Press, Cambridge.
Greenwood, M. (1926). The errors of sampling of the survivorship tables. In Reports on
Public Health and Statistical Subjects, Volume 33. Her Majestys Stationary Office,
London.
Kalbfleisch, J. D. and R. L. Prentice (2002). The statistical analysis of failure time data. John
Wiley & Sons, New York.
Kaplan, E. L. and P. Meier (1958). Nonparametric estimation from incomplete observations.
Journal of the American Statistical Association 53(282), 457481.
Moeschberger, M. L. and J. P. Klein (2005). Survival Analysis: Techniques for Censored and
Truncated Data. Springer-Verlag, New York.
20
Transition Modeling
Bruce Jones and Weijia Wu
Chapter Preview. This chapter provides an introduction to transition modeling. Consider a situation where an individual or entity is, at any time, in one of several states
and may from time to time move from one state to another. The state may, for example,
indicate the health status of an individual, the status of an individual under the terms
of an insurance policy, or even the state of the economy. The changes of state are
called transitions. There is often uncertainty associated with how much time will be
spent in each state and which state will be entered on each transition. This uncertainty
can be modeled using a multistate stochastic model. Such a model may be described in
terms of the rates of transition from one state to another. Transition modeling involves
the estimation of these rates from data.
Actuaries often work with contracts involving several states and financial implications associated with presence in a state or transition between states. A life insurance
policy is a simple example. A multistate stochastic model provides a valuable tool
to help the actuary analyze the cash flow structure of a given contract. Transition
modeling is essential to the creation of this tool.
This chapter is intended for practitioners, actuaries, or analysts who are faced with
a multistate setup and need to estimate the rates of transition from available data.
The assumed knowledge only basic probability and statistics as well as life contingencies is minimal. After a discussion of some relevant actuarial applications
in Section 20.1, this chapter gives a brief introduction to multistate stochastic models in Section 20.2. We follow the notation, terminology, and approach of (Dickson
et al. 2009, chapter 8). Although our introduction is intended to stand on its own, the
reader is encouraged to consult Dickson et al. (2009) for a more detailed presentation. After we introduce multistate models, we discuss how to estimate the rates of
transition under different assumptions about the model and the data in Sections 20.3
and 20.4.
515
516
Transition Modeling
0
(x) alive
1
- (x) dead
(x)alive
(y)alive
PP
PP
PP
PP
PP
(x) dead
(y) alive
PP
PP
(x) alive
(y) dead
PP
P
Pq
?
(x) dead
(y) dead
517
0
(x) insured
2
cause 1
m
cause m
cause 2
ZZ
~
lives to die at the same time. This is one way of introducing positive dependence of
the two lifetimes. Another way is to assume that the rates of transition from state 2 to
state 3 are greater than the rates of transition from state 0 to state 1 and that the rates
of transition from state 1 to state 3 are greater than the rates of transition from state 0
to state 2. That is, mortality rates are higher after ones spouse dies than before.
A third example of a multistate model is the multiple decrement model shown in
Figure 20.3. The process begins in state 0 and may move to one of m absorbing states.
Such a model can be used, for example, to represent mortality by cause of death or
the mode of termination of an insurance policy.
The three-state illness-death model shown in Figure 20.4 is another example of a
multistate model that is useful in actuarial practice. Methods for multistate models
are particularly helpful in this situation, because individuals may return from the sick
state to the healthy state. Therefore, probabilities associated with the process being
in a given state at a given future time cannot be calculated directly. However, the
methods presented in Section 20.2.1 allow us to determine these probabilities.
Finally, the claims payment process can be modeled using a multistate model as
shown in Figure 20.5. In practice, we must allow for partial payment of the claim
and the possibility that payment is required by a legal decision after it is denied. This
more complicated situation may still be represented as a multistate model.
0
(x) healthy
(x) sick
J
J
^
(x) dead
518
Transition Modeling
2
modified
0
unreported
1
- reported
not paid
J
J
J
J
J
J
J
J
J
^
3
paid
4
denied
519
process using the splitting of states technique described by Haberman and Pitacco
(1999).
For i, j = 0, 1, . . . , s 1 and z, t 0, we define the transition probability functions
ij
t pz
(20.1)
Notice that this is an extension of the life contingencies notation t pz , which represents
the probability that an individual will survive to age z + t given that the individual
has survived to age z. In this context, it is common to view age z as the time origin.
In our multistate modeling context, time 0 is the time origin, and z is just a number of
years since time 0. This is important to recognize when our multistate model describes
something different from the progression of an individual through a number of states.
For example, consider the two-life model shown in Figure 20.2. At time 0, we have
a life aged x and a life aged y. Suppose a joint life insurance policy is issued at this
time, and we wish to value the policy 10 years later, when one life is aged x + 10 and
00
.
the other is aged y + 10. Our calculations then involve probabilities denoted by t p10
It is also useful to define the probabilities,
ii
t pz
In words, t pzii simply represents the probability that a process in state i at time z will
remain in state i until at least time z + t. The probability t pzii differs from (and is
no larger than) the probability t pzii , because the latter includes the probability that a
process in state i at time z will leave state i before time z + t and return to state i by
time z + t (for multistate processes in which this can occur). For certain multistate
ij
processes, the probabilities t pzii will help us in developing expressions for the t pz
probabilities.
The transition intensity functions or, simply, transition intensities are also very
useful. For i, j = 0, 1, . . . , s 1 with i = j , and z > 0, define
ijz
= lim+
h0
ij
h pz
where the right-hand limit is used because h represents a time period and therefore
must be positive. The transition intensities represent the instantaneous rates of transition from one state to another at a given time. They are analogous to the force of
mortality and are sometimes referred to as the forces of transition. It is convenient to
characterize a multistate model in terms of the transition intensities due to their interpretability. We often have a fairly clear idea of how these functions should behave,
whereas the behavior of the transition probability functions may be less obvious.
520
Transition Modeling
s1
t
ij
ii
(20.2)
z+w dw .
t pz = exp
0
j =0,j =i
This makes sense because t pzii is the probability of remaining in state i (rather than
ij
remaining alive), and s1
j =0,j =i z+w is the total force of transition out of state i
(rather than the force of mortality).
= t pz00 ,
11
t pz
= t pz11 ,
22
t pz
= t pz22 ,
and each of these functions can be determined using (20.2). To obtain an expression
for t pz01 , we integrate over the possible times of transition from state 0 to state 1. We
have
t
01
00 01
11
t pz =
w pz z+w tw pz+w dw.
0
The idea here is that the process remains in state 0 from time z until time z + w,
then makes a transition to state 1 during the (infinitesimally small) time period from
z + w to z + w + dw, and then remains in state 1 until time z + t. The integral then
521
sums over all of the small time intervals of length dw during which the transition may
occur. Similarly,
t
02
00 02
22
t pz =
w pz z+w tw pz+w dw.
0
Finally,
03
t pz
= 1 t pz11 ,
23
t pz
= 1 t pz22 .
and
Note that it will often be necessary to evaluate the above integrals numerically. Dickson
et al. (2009) suggest using repeated Simpsons rule for this.
An alternative approach to determining the transition probability functions from the
transition intensities involves solving the Kolmogorov forward differential equations.
This approach can be used even when repeat visits to states are possible. For i, j =
0, 1, . . . , s 1 and z, t 0, the Kolmogorov forward differential equations are given
by
s1
d ij
kj
jk
(t pzik z+t t pzij z+t ).
t pz =
dt
k=0,k=j
(20.3)
These equations can be solved numerically using an approach called Eulers method.
We begin by noting that, if (20.3) holds, then for small h
s1
ij
t pz
kj
jk
ij
t+h pz
and therefore,
ij
ij
t+h pz t pz + h
s1
kj
k=0,k=j
ij
= 1 for i = 0, 1, . . . , s 1,
and
ij
0 pz
jk
= 0 for i = j,
(20.4)
522
Transition Modeling
ij
choosing a small value of h (e.g., h = 1/12), and using (20.4) to calculate h pz values,
ij
2h pz values, and so on. See Dickson et al. (2009) for examples of these calculations
as well as a derivation of the Kolmogorov forward differential equations.
Since we can determine the probabilities we need from the transition intensities, it
is sufficient to describe a multistate model by specifying the transition intensities. We
next turn to the main focus of this chapter: estimating the transition intensities from
data.
20.3.2 Data
For the purpose of estimating transition intensities, the ideal data would consist of a
large number of complete realizations or outcomes of the process Y . That is, for each
outcome of the process, we would know the value of Y (t) from t = 0 until the process
enters an absorbing state or until the end of the time horizon of interest. It is rarely
the case that we have ideal data. In this chapter, we consider two possibilities in terms
of the available data. In the first case, the process is observed continuously over some
interval of time for each realization. In the second case, the process is observed only
at distinct time points for each realization, as is the case with panel data. The latter
case is discussed in Section 20.4.
Suppose that we are interested in estimating the transition intensity from state i
ij
to state j ; that is, z for a specific i and j . Now assume that our data consist of
n outcomes of the process Y observed continuously over the time interval [0, ].
We denote these outcomes by y1 , y2 , . . . , yn , where for outcome , the collection of
observed states is {y (t), 0 t }. Also, suppose that there are m distinct times at
523
(20.5)
The NA estimator is associated with survival analysis, which is discussed in Chapter 19. Standard textbooks on the subject include Kalbfleisch and Prentice (2002) and
Lawless (2003), and a thorough treatment of the methods applied to multistate models
is provided by Andersen (1993). The NA estimator was proposed by Nelson (1969)
524
Transition Modeling
y1 (t)
6
2
1
s
0
c
3
10
- t
y2 (t)
6
2
1
s
0
s - t
10
y3 (t)
6
2
1
s
0
c
2
s
3
c
6
10
- t
y4 (t)
6
2
1
s
0
c
5
s
9
s - t
10
as a graphical method, and its properties were investigated by Aalen (1978). The NA
estimator is also discussed by Klugman et al. (2012).
To illustrate the method, we introduce a simple example.
Example 20.1. Consider the illness-death process shown in Figure 20.4 and imagine
that we observe the four outcomes given in Figure 20.6. Suppose we are interested in
the transition intensity from state 0 to state 1.
525
tw
2
3
5
6
Rw
4
3
3
2
dw
1
1
1
1
0,
t <2
,
2t <3
7
, 3t <5 .
= 14 + 13 = 12
M(t)
+ 13 = 11
, 5t <6
12
12
11 + 1 = 17 , 6 t < 10
12
2
12
The computations are easily performed using functions in the survival package in
R. The following commands from an R session produce the plot in Figure 20.7. Notice
that the survival data object contains six observations, even though we have observed
just four outcomes of the process. This is because the four outcomes include a total of
six visits to state 0. Our survival data object requires an observation for each interval
of time spent in state 0 in order to correctly calculate the Rw and dw values for each
transition time tw . The survfit function determines the estimates.
> library(survival) # load survival package
Loading required package: splines
> # Create the survival data object.
> start <- c(0,0,0,3,0,9) # interval start times
> end <- c(3,10,2,6,5,10) # interval end times
> event <- c(1,0,1,1,1,0) # indicator of 0->1 transitions
> data.surv <- Surv(start,end,event)
> data.surv
[1] (0, 3] (0,10+] (0, 2] (3, 6] (0, 5] (9,10+]
> # Calculate and plot NA estimates.
> fit.NA <- survfit(data.surv~ 1,type="fl")
> plot(fit.NA, fun="cumhaz",xlab="t",ylab="M(t)",
+ main="NA Estimates",conf.int=FALSE)
M(t)
Transition Modeling
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
526
10
Fig. 20.7. Nelson-Aalen estimates of the cumulative transition intensity from state 0
to state 1 in the illness-death model example.
Example 20.2. In this example, we consider real data from a pension plan. Plan
members are observed over two consecutive calendar years. For each individual who
was a member of the plan at any time during a given calendar year, we observe
the date of birth, the date of hire, the date of termination (if terminated), the sex of
the individual, and the annual salary during the year. The dataset contains a total of
725 observations, and 418 unique members were observed many were observed in
both years.
Suppose we are interested in estimating the intensity of termination for reasons
other than death. We can gain an understanding of the behavior of this function by
plotting the NA estimates of the cumulative intensity of termination (as a function
of years of service) separately for males and females. These estimates are shown
in Figure 20.8 along with approximate 95% confidence intervals. The confidence
intervals are based on the variance estimates obtained from (20.6) along with a
normal distribution assumption.
The graph suggests that males have a significantly lower cumulative intensity of
termination than females. We can justify this conclusion by noting that we are unable
to hypothesize a single cumulative intensity function for both genders that lies mostly
between the confidence limits for males and females. Also, the concave behavior
of the estimated cumulative intensity of termination suggests that the intensity of
termination is a decreasing function (of years of service).
Note that, in using the NA estimator in this example, we must assume that termination for reasons other than death occurs independently of other types of termination
(e.g., death). This assumption is required more generally competing transitions must
occur independently of one another.
527
females
males
10 15 20 25 30 35
years of service
n
"
L ( )
=1
n
"
(20.7)
=1
where is a vector of parameters to be estimated. Note that the probability in expression (20.7) is actually a probability density; the probability of a specific outcome of
the process is 0. The calculation of the required probability densities is illustrated in
the next example.
Standard asymptotic likelihood theory can be used to make inferences about and
functions of , provided that our sample of outcomes is sufficiently large. An overview
528
Transition Modeling
and
11 10
00
L4 ( ) = 5 p000 01
5 4 p5 9 1 p9 .
The vector depends on the parametric models we have chosen for our transition
intensities. In the simplest case, each transition intensity is assumed to be constant.
ij
That is, z = ij and we have the parameter vector = (01 , 02 , 10 , 12 ). Using
equation 20.2, the likelihood function is, after simplification,
L() = exp{24(01 + 02 ) 14(10 + 12 )} (01 )4 (10 )2 12 .
(20.8)
The maximum likelihood estimates are obtained in the usual way by taking the log
of the likelihood function, differentiating the log-likelihood with respect to each
parameter, and setting these derivatives to 0. We have
01 =
4
1
= ,
24
6
10 =
2
1
= ,
14
7
12 =
1
.
14
The MLE of 02 cannot be found using this approach, because the likelihood is
a strictly decreasing function of 02 . The likelihood is then maximized when this
parameter equals 0, the left endpoint of the range of possible values. Thus,
02 = 0.
529
Also, reasoning intuitively, we observe the process in state 0, but no transitions from
state 0 to state 2 are observed. So a transition intensity estimate of 0 makes sense. It is
interesting to note that, when the transition intensities are constant as in this example,
ij =
ij
(20.9)
530
Transition Modeling
during the calendar year. The vector end.time gives the time since hire at which we
stop observing each member during the calendar year, and the vector end.event
contains a 1 for each member who terminated at the time given in end.time and
a 0 for each member who did not. The three vectors start.time, end.time,
and end.event are used to create a survival data object corresponding to the 725
observations.
> library(survival)
Loading required package: splines
> term.surv <- Surv(start.time,end.time,end.event)
The values of the covariates we consider are contained in the vectors hire.age
(the age at hire), sex (F or M), salary (the annual salary during the calendar year),
and year (0 for the first calendar year and 1 for the second calendar year). We we
use R to estimate in the proportional hazards model in (20.9), with these covariates
making up the components of the vector x(t) for each observation.
Note that the annual salary of the member is a time-varying covariate. However, it
is assumed to be constant within each calendar year. The data are structured in such
a way that there is an observation for each member in each of the two calendar years.
So, for the interval of time represented by each observation, the salary is constant and
can therefore be treated as a fixed covariate, as can the other covariates.
Estimation of the components of is performed by the coxph function, and
its output is summarized by the summary function. Only part of the information
produced by the summary function is displayed next. In specifying the model, we
take the logarithm of the salaries because the salary distribution is highly skewed and
we obtain poor results if the salary values themselves are used.
> term.coxph1 <- coxph(term.survhire.age+sex+log(salary)+year)
> summary(term.coxph1)
Call:
coxph(formula = term.surv ~ hire.age + sex + log(salary) + year)
n= 725, number of events= 96
coef
exp(coef) se(coef)
z
hire.age
-0.003913
0.996095
0.011859
-0.330
sexM
-0.120873
0.886146
0.265267
-0.456
log(salary) -0.737475
0.478320
0.231664
-3.183
year
0.108676
1.114801
0.208062
0.522
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Pr(>|z|)
0.74145
0.64863
0.00146**
0.60144
531
The results show that 96 of the 725 observation intervals ended with a termination. In the table provided by the summary function, we see the estimates of the
components corresponding to the four covariates in the column headed coef (for
coefficient), as well as the exponentiated values of these estimates and the standard
errors associated with the estimates. The final two columns provide the information
necessary to perform individual tests of the hypotheses that the components are 0,
implying that the covariate is not statistically significant in modeling the intensity of
termination. The column headed z gives the values of the coefficients divided by their
standard errors. Under the hypothesis that a given coefficient is 0, the corresponding
value of this test statistic is distributed (asymptotically) standard normal. Under this
assumption, the final column of the table then gives the p-values of the hypotheses.
The information in the above table suggests that only log(salary) has a significant
effect on the intensity of termination. This seems to contradict our earlier observation
that this intensity function is different for males and females. Further analysis reveals
a relationship between sex and salary: the male salaries are generally higher, and it
appears that sex is not an important variable when the effect of salary is reflected.
When we fit a proportional hazards model with just the covariate log(salary), we
obtain the following.
> term.coxph2 <- coxph(term.surv~ log(salary))
> summary(term.coxph2)
Call:
coxph(formula = term.surv ~ log(salary))
n= 725, number of events= 96
coef exp(coef) se(coef)
z Pr(>|z|)
log(salary) -0.8146
0.4428
0.1815 -4.489 7.16e-06 ***
--Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
We can perform a likelihood ratio test of the hypothesis that the coefficients corresponding to age at hire, sex, and year are all 0. The maximum log-likelihood obtained
with all four variables in the model is 381.9689 (obtained from term.coxph),
and the maximum log-likelihood obtained with just log(salary) in the model is
382.2580 (obtained from term.coxph2). The likelihood ratio statistic is then
2(382.2580 381.9689) = 0.5781. When comparing this with the quantiles of a chisquared distribution with 3 degrees of freedom, we conclude that there is no evidence
against the hypothesis (p-value = 0.9014).
When including covariates in a model, the model may be improved by using some
function of the covariate value, rather than the value itself. Since log(salary) was
532
Transition Modeling
chosen rather arbitrarily to address the fact that the salary distribution is skewed, we
explore various other functions of salary. We find that the log-likelihood is larger
(equal to 378.8770) when we use an indicator that the salary is greater than 36,200,
rather than log(salary). We obtain the following in this case
533
we can write the following nonstandard expression for the transition intensity:
D
log t u(x(t))
log t u(x(t))
ij
tS0
,
(20.10)
t|x(t) = f0
where S0 () is a fully specified survival function of a random variable that can take
any real value, f0 () is the corresponding probability density function, u(x(t)) =
0 + x(t), and > 0 is a parameter to be estimated, as are 0 and = (1 , . . . , q ) .
This specification of the model allows for time-dependent covariates, which are not
generally used with the accelerated failure-time model. In fact, the accelerated failuretime interpretation is lost when time-dependent covariates are present. Cox and Oakes
(1984) elaborate on this.
Common choices for S0 () include the the survival functions of the extreme value,
logistic, and normal distributions. These lead to Weibull, loglogistic, and lognormal
transition intensities, respectively.
Example 20.5. We fit a number of fully parametric regression models to the pension
plan termination data. We show results for two of them in this example.
The first is a proportional hazards model with a parametric form specified for
the baseline intensity of termination. Since we expect the intensity of termination to
decrease with increasing service, we consider an exponentially decreasing baseline
intensity. We refer to this as a Gompertz intensity, even though it is normally increasing. The single covariate in the model is I (salary > 36,200), which is 1 when the
salary is greater than 36,200 and 0 otherwise. Thus we have the model
t|salary = exp{0 + 1 I (salary > 36,200)}ct ,
where t is the time since hire and 0 , 1 , and c are parameters to be estimated, with
< 0 , 1 < , and 0 < c < 1, respectively. Constructing the likelihood, and
0 = 0.4997,
maximizing with respect to the three parameters, gives the estimates
c = 0.8425. The resulting cumulative intensity of termination is
1 = 1.0531, and
shown in Figure 20.9, which indicates that the fit is rather poor.
The second parametric model we discuss is a lognormal accelerated failure-time
model with the same covariate. For this model,
log t0 1 I (salary>36,200)
,
t|salary =
t 1 log t0 1 I (salary>36,200)
where () and () are the standard normal pdf and cdf, respectively, and 0 , 1 ,
and are parameters, with < 0 , 1 < , and > 0. The maximum likelihood
1 = 1.2997, and
0 = 0.4669,
= 1.1788. The
estimates of these parameters are
resulting cumulative intensity of termination is shown in Figure 20.9 and appears
to fit somewhat better than that of the proportional hazards model with Gompertz
Transition Modeling
log(cumulative force of termination)
4 3 2 1 0 1
534
low salary
high salary
Fig. 20.9. Fitted cumulative forces of termination for low-salary and high-salary
employees based on a proportional hazards model with Gompertz baseline force
(dashed) and a lognormal accelerated failure-time model (dotted).
baseline intensity for the early months of service, but the lognormal model may not
be satisfactory.
Further analysis is required to determine a suitable parametric model for the force
of termination for this pension plan.
A variety of graphical methods and formal statistical tests can be used in selecting
an appropriate parametric model and checking its fit. These are discussed by Lawless
(2003) and Klugman et al. (2012).
535
y (0)
y (6)
y (10)
0
0
0
0
1
0
1
1
1
0
2
0
536
Transition Modeling
{0, 6, 10}
{0, 2, 4, 6, 8, 10}
{0, 1, 2, . . . , 10}
[0, 10]
0.2425
0
0.1554
0.0555
0.2599
0
0.2442
0.0664
0.1995
0
0.1781
0.0689
0.1667
0
0.1429
0.0714
and
12 = 0.0555.
These estimates are quite different from those obtained when we observe each
outcome continuously from time 0 to time 10. Table 20.2 shows the MLEs obtained
with several sets of observation times. It helps us better understand the impact of the
information loss associated with distinct observation times.
The variability associated with the maximum likelihood estimates can be assessed
using standard asymptotic likelihood theory. If one maximizes the log-likelihood
by using the optim function in R with the optional argument hessian = TRUE,
then the output provides the observed information matrix evaluated at the maximum
likelihood estimates. This can be used to obtain variance and covariance estimates
for the maximum likelihood estimators. A good discussion of asymptotic likelihood
theory is provided by Lawless (2003). Asymptotic methods are not suitable for our
simple example involving just four outcomes of the multistate process.
References
Aalen, O. (1978). Nonparametric inference for a family of counting processes. Annals of
Statistics 6(4), 701726.
Andersen, P. K. (1993). Statistical Models Based on Counting Processes. Springer,
New York.
Andersen, P. K. and N. Keiding (2002). Multi-state models for event history analysis.
Statistical Methods in Medical Research 11(2), 91115.
Cox, D. D. R. and D. Oakes (1984). Analysis of Survival Data, Volume 21. CRC Press, Boca
Raton, FL.
Dickson, D. C., M. R. Hardy, and H. R. Waters (2009). Actuarial Mathematics for Life
Contingent Risks. Cambridge University Press, Cambridge.
Haberman, S. and E. Pitacco (1999). Actuarial Models for Disability Insurance. CRC Press,
Boca Raton, FL.
References
537
Hoem, J. M. (1969). Markov chain models in life insurance. Blatter der DGVFM 9(2),
91107.
Kalbfleisch, J. D. and R. L. Prentice (2002). The Statistical Analysis of Failure Time Data
(2nd ed.). John Wiley & Sons, New York.
Klugman, S. A., H. H. Panjer, and G. E. Willmot (2012). Loss Models: From Data to
Decisions (4th ed.), Volume 715. John Wiley & Sons, New York.
Lawless, J. F. (2003). Statistical Models and Methods for Lifetime Data (2nd ed.). John
Wiley & Sons, New York.
Macdonald, A. (1996a). An actuarial survey of statistical models for decrement and
transition data-i: Multiple state, Poisson and binomial models. British Actuarial
Journal 2(1), 129155.
Macdonald, A. (1996b). An actuarial survey of statistical models for decrement and
transition data: II: Competing risks, non-parametric and regression models. British
Actuarial Journal, 429448.
Macdonald, A. S., H. R. Waters, and C. T. Wekwete (2005a). A model for coronary heart
disease and stroke with applications to critical illness insurance underwriting I: The
model. North American Actuarial Journal 9(1), 1340.
Macdonald, A. S., H. R. Waters, and C. T. Wekwete (2005b). A model for coronary heart
disease and stroke with applications to critical illness insurance underwriting II:
Applications. North American Actuarial Journal 9(1), 4156.
Nelson, W. (1969). Hazard plotting for incomplete failure data. Journal of Quality
Technology 1(1).
Norberg, R. (1995). Differential equations for moments of present values in life insurance.
Insurance: Mathematics and Economics 17(2), 171180.
Ramlau-Hansen, H. (1988). The emergence of profit in life insurance. Insurance:
Mathematics and Economics 7(4), 225236.
Waters, H. R. (1984). An approach to the study of multiple state models. Journal of the
Institute of Actuaries 111(2), 363374.
Index
robust, 229
shrinkage estimator, 198
data
clustered, 180, 183
count, 87
cross-classified, 190
cross-sectional, 183
grouped, 146
hierarchical, 183, 188
individual, 146
longitudinal, 167, 219
missing, 180
multilevel, 180, 183
nested, see also hierarchical
non-Gaussian, 399
non-nested, see cross-classified
panel, 167, 183, 534
spatial, 261
survival, 482
time series, 427
transitions, 515, 522
two-part, 141
unbalanced, 171, 180
data example
Annual Flow of the River Nile, 429
Australian Automobile Bodily Injury Scheme, 469
Australian Workers Compensation Insurance, 475
Automobile PIP Questionable Claims Data, 284
California Auto Assigned Risk (CAARP) Data,
284
Credit Insurance Data, 190, 212
Cross Selling, 122
German Car Insurance Data, 262, 269, 276
Group Term Life, 175
Hachemeister Data, 187, 206, 218, 224
Heart Disease Data, 373, 390, 394
Losses and Expenses, 142
Massachusetts Automobile Claims, 152, 242
Medical Expenditure Panel Survey (MEPS), 15, 247,
254
Mexican Central Bank, 346, 353, 356
539
540
data example (cont.)
Mexico Insurance Market, 342, 361
Monthly Global Temperature, 429
Motor Insurance, 69, 74, 80, 83
National Nursing Home Survey (NNHS), 482, 486,
513
pension plan terminations, 526, 529, 533
Personal Automobile Claims, 316
Quarterly Earnings of Johnson & Johnson, 429
Sales Data from BoxJenkins, 429
Singapore Automobile Claims, 91
Texas Closed Claim Data, 284
U.S. Medical Malpractice Insurance, 475, 476
U.S. Workers Compensation Insurance, 187, 190,
201, 412
dependent lifetimes
common shock, 516
dimension reduction, 282
distribution tail
criteria of subexponential, 238
left tail, 238
right tail, 237
tail heaviness, 237
distributions
F , 244
asymmetric Laplace, 254
Bernoulli, 66
beta prime, see generalized beta of the second kind
(GB2)
binomial, 88, 112, 133
bivariate normal, 170
Burr XII, 245
exponential, 507
exponential GB2, 245
exponential family, 110, 146, 161, 238, 241
extreme value, 533
fat-tailed distributions, 237
gamma, 100, 112, 134, 146, 237, 242, 245
generalized beta, see generalized beta of the second
kind (GB2)
generalized beta of the second kind (GB2), 244
generalized gamma, 245
geometric, 94
inverse-Gaussian, 100, 112, 242
Laplace, 252
location-scale, 227
log location-scale family, 244
log-normal, 100, 245
logistic, 507, 533
negative binomial, 93
negative binomial p (NBp), 93
normal, 21, 112, 237, 533
over-dispersed Poisson (ODP), 454
Pareto, 161, 237, 244, 245
Poisson, 87, 111
Poisson-inverse Gaussian, 94
Poisson-lognormal, 95
Tweedie, 128, 150
Weibull, 228, 237, 245, 507, 533
zero-inflated, 102
zero-inflated Poisson (ZIP), 102
Index
estimation
corrected adaptively truncated likelihood (CATL),
228
generalized least squares (GLS), 196
Laplace method, 404
least absolute deviations (LAD), 251
least squares, 21, 241
maximum likelihood, 7, 70, 72, 90, 93, 95, 110, 118,
135, 149, 177, 241, 527, 528, 534
minimum bias technique, 6, 87
penalized least-square, 387
penalized quasi-likelihood (PQL), 406
pseudo-likelihood (PL), 406
restricted maximum likelihood (REML), 199, 219,
228
restricted pseudo-likelihood (REPL), 406
robust, 228
exchangeability, 355, 356
expected utility, 336
exposure, 102, 144, 187, 221
false negative, 75
false positive, 75
Fisher information matrix, 246, 256
full conditional distribution, 265
Gelman-Rubin statistic, 322
generalized estimating equation (GEE), 399
generalized linear mixed model (GLMM), 400
generalized linear model (GLM), 7, 60, 107, 141, 238,
241, 338, 369, 399, 455, 462
gamma regression, 117, 126, 150, 241, 249,
275276
linear model, 338
link function, 68, 90, 109, 142, 369, 399
logistic regression, 67, 149, 345
mixed, 179
offset, 124, 142
Poisson regression, 89, 150, 241, 274, 352
probit regression, 78
systematic component, 369
Tweedie, 128, 150
goodness-of-fit statistic, 75, 243
generalized cross-validation (GCV), 383
adjusted R 2 , 32, 33, 42
mean squared error, 381
predictive squared error, 382
pseudo-R squared, 76
R 2 , 25, 30, 33, 52
hypothesis testing, 316
influential point
Cooks distance, 40, 41, 42
leverage, 34, 36
outlier, 34, 36, 37, 38, 40
information statistic
Akaike information criterion, AI C, 32, 33, 42, 52,
98, 132, 250, 383
Akaike information criterion, corrected, AICC, 132
Bayesian information criterion, BIC, 98, 132, 250
Index
deviance information criterion, DIC, 277, 342, 348,
393
Kullback-Leibler, 241
Schwarz Bayesian criterion
see Akaike information criterion, AIC, 132
Kolmogorov forward differential equations, 521
numerical solution using Eulers method, 521
learning
supervised, 280
unsupervised, 281
life annuity and insurance, 516
log-odds, 68
logistic regression, see generalized linear model (GLM)
logit, 68, 390
loss reserving, 3, 399, 450
chain ladder algorithm, 453
generalized linear model (GLM), 455
models of individual claims, 467
over-dispersed Poisson (ODP) cross-classified model,
454, 463
over-dispersed Poisson (ODP) Mack model, 454
MSE, see residual variance
Markov chain Monte Carlo, 320
Gibbs sampler, 320
Metropolis-Hastings sampler, 320
Markov chain Monte Carlo methods, 267, 276
Markov process, 518
Markov random field, 265
median regression, 360
medoid, 301
meta-analysis, 356
model
GB2 regression, 244, 245, 249
accelerated failure times, 532, 533
Bayesian approach for median regression, 252
Bayesian quantile regression, 254
Burr XII regression, 249
cluster analysis, 283
complete pooling, 183, 201
conditional logit, 81, 84
cross-sectional, 182
cumulative logit, 79
factor analysis, 283
frequency-severity, 148
generalized additive model (GAM), 60, 242, 369,
388, 456
generalized logit, 81
hierarchical, 185, 273, 355
inverse-Gaussian regression, 249
linear fixed effects, 173, 183
linear mixed (LMM), 182, 183, 184, 192, 220
linear probability, 67
linear random effects, 173, 183
MCMC, see Markov chain Monte Carlo
mean regression, 236
median regression, 238, 250
mixed, 354
multilevel, 184, 185
541
542
Index
plot
LOESS curve, 38
ggplot2 package, 182
boxplot, 19
gains chart, 58
Gelman-Rubin, 324
icycle, 308
multiple time series, 218
quantile-quantile (QQ), 39, 247, 249
receiver operating characteristic (ROC) curve, 77
residual, 36
rug, 243
scatterplot, 16
scatterplot matrix, 49
scree, 293
time series, 427, 429
trace, 324
trellis, 175, 187
Poisson regression, see generalized linear model (GLM)
Polya tree, 344
posterior distribution, 253, 317
posterior predictive, 337, 349, 350
checks, 349, 350
prior
conjugate, 253, 337, 340
degrees of belief, 335
diffuse, 337, 341, 346, 348, 350
hierarchical, 334, 355, 356
improper uninformative, 253
informative, 335, 357
nonparametric, 344, 359
Polya tree, 359, 362
vague, 253, 338, 341, 343, 346, 348, 350, 353
prior distribution, 253, 317
conjugate, 318
elicitation, 326
Jeffreys, 328
non-informative, 327
probit regression model, see generalized linear model
(GLM)
proper distribution, 267, 274
proportional hazards
assumption, 494, 497
baseline survival function, 501
model, 497
model residuals, 502
time-varying covariates, 506
proportional hazards model, 476
pure premium, 127, 155
random effect
spatial, 274
random effects
binary outcomes, 177
crossed random effects, 196
generalized linear mixed model
see generalized linear model (GLM) 180
intercepts, 173
multiple random effects per level, 195
nested random effects, 196
single random effect per level, 195
Index
score, 95
Wald, 95, 200
Wilcoxon, 495
threshold, 74
time series inference
autocorrelation, 431, 437
autocovariance, 431
back casting, 435
Kalman filter, 445
prediction, 428, 442
time series model
generalized autoregressive conditional
heteroskedasticity (GARCH), 447
autoregressive of order 1 (AR(1)), 434
autoregressive, integrated, moving average
(ARIMA), 440
autoregressive, moving average (ARMA),
434
autoregressive, moving average of order (p, q),
ARMA(p, q) , 435
Box-Jenkins, 440
geometric random walk, 433
LeeCarter, 447
moving average model of order q (MA(q)), 434
random walk, 432
seasonal ARIMA, 443
state space, 445
stochastic volatility, 446
structural, 443
white noise, 432
time series model properties
invertibility, 437
stationarity, 429, 431, 434, 437
transformation, 238
Box-Cox, 239, 371
inverse hyperbolic sine, 239
log-log, 491
logarithmic, 238
modified modulus, 240
modulus, 239
nonlinear, 239
signed power, 239
time series differencing, 430, 433, 440
time series lag operator, 435, 436, 440
variance stabilizing, 239
transition intensity function, 519, 520, 522
cumulative intensity function, 523
transition probability function, 178, 519, 520
two-part distribution, 148
underwriting, 3
value-at-risk, 237
variables
binary, 17, 66, 67
categorical, 17, 65, 66, 281
components, 282
dummy, 17
dynamic, 475
factor, 282
indicator, 17
Likert scale, 283
ordinal categorical, 79
scale, 281, 283
static, 475
unpredictable, 475
543