Professional Documents
Culture Documents
IQRM Book 2020 Jan 28
IQRM Book 2020 Jan 28
TO
QUANTITATIVE
RESEARCH
METHODS
This coursebook is designed to include sufficient statistical concepts to allow students to make
good sense of the statistical figures and numbers that they are exposed to in daily life. After
reading the book, students should understand the basics of quantitative research and be able
to critically review simple statistical analysis. The book is intended to be self-contained, but we
do not include mathematical proofs.
The examples are intended to be relevant in Hong Kong for a wide range of disciplines.
Most of the following topics and questions will be covered in the course:
1: Research Methods...................................................................................................................14
Learning objectives: understand theory, model, Occam’s law and types of random variable .
14 What is research?.................................................................................................................14
Why use research methods to solve problems?........................................................................14
Research plan............................................................................................................................16
What is the scientific method?..................................................................................................17
What are phenomena?..............................................................................................................17
What is objective?....................................................................................................................18
What is a logical and systematic method?..........................................................................19
What is analysis?......................................................................................................................20
Levels of quantitative analysis.................................................................................................20
What is a theory?................................................................................................................21
What is a model?......................................................................................................................21
2
What is a hypothesis?...............................................................................................................23
Occam’s principle (or law or razor).........................................................................................24
What is proof?.................................................................................................................................26
What is verified?.......................................................................................................................26
What is a constant?.............................................................................................................27
What is a random variable?......................................................................................................27
What is an explanatory variable?.......................................................................................27
What is an independent variable?.......................................................................................28
What is a dependent variable?............................................................................................28
What are extraneous variables?..........................................................................................28
Mechanisms of control.............................................................................................................29
2: Association and causation.......................................................................................................33
Learning objectives: understand association & causation........................................................33
What is association?.................................................................................................................33
What is causation?....................................................................................................................33
What is Grainger causality?.....................................................................................................35
Magnitude and consistency of association...............................................................................36
Experiment versus observation................................................................................................36
3: Research design.......................................................................................................................37
Learning objectives: understand representativeness, probability sampling & non-sampling error
............................................................................................................................................37
What is a population?...............................................................................................................37
What is a sample?.....................................................................................................................37
What are units of analysis?.......................................................................................................37
What is representativeness?...............................................................................................38
What is a probability sample?............................................................................................38
What is a simple random sample?......................................................................................38
What is a cluster sample?...................................................................................................39
What is a stratified sample?...............................................................................................39
What is a systematic sample?.............................................................................................40
What is a network sample?................................................................................................40
What is a distance sample?................................................................................................41
Sampling versus non-sampling errors:.....................................................................................41
Telephone versus face-to-face interviews................................................................................43
Mobile versus fixed-line telephone surveys.............................................................................43
Primary versus secondary data...........................................................................................43
Choosing sample size...............................................................................................................44
Observation versus participation........................................................................................44
Qualitative versus quantitative...........................................................................................44
4: Basics of qualitative research..................................................................................................46
Learning objectives: understand key concepts of qualitative research....................................46
Biography or narrative research...............................................................................................47
Phenomenology...............................................................................................................................47
Grounded theory.......................................................................................................................48
Ethnography.............................................................................................................................49
Case study.................................................................................................................................49
Qualitative sampling.................................................................................................................50
Saturation............................................................................................................................50
Theoretical or purposeful sampling..........................................................................................51
Convenience sampling..............................................................................................................51
Snowball sampling...................................................................................................................52
Observation, interviewing and other means of collecting qualitative data..............................52
Observation...............................................................................................................................52
Interviewing..............................................................................................................................53
Recording and analyzing qualitative data..........................................................................53
Computer-Assisted Qualitative Data Software........................................................................54
5: Measuring instruments......................................................................................................55
Learning objectives: understand three key criteria for a measuring instrument: reliability,
validity & precision, and scales of measurement...................................................................55
What is reliability?...................................................................................................................55
What is validity?.......................................................................................................................56
What is precision?....................................................................................................................57
Making operational choices (how to measure something).......................................................57
Operational choice considerations:...........................................................................................60
Levels/scales of measurement........................................................................................................61
What is nominal scale?.............................................................................................................62
What is ordinal scale?...............................................................................................................62
What is interval scale?........................................................................................................62
What is ratio scale?...................................................................................................................62
Index versus scale measures...............................................................................................63
What is a Likert scale?.......................................................................................................63
What is semantic differential scale?.........................................................................................64
What is Guttman scaling?.........................................................................................................64
6: Probability...............................................................................................................................65
Learning objectives: understand three laws of probability.......................................................65
What is probability?.................................................................................................................65
When do we add probabilities?................................................................................................66
When do we multiply probabilities?........................................................................................66
Binomial distribution................................................................................................................68
What is conditional probability?..............................................................................................69
What is Bayes’ Law? .................................................................................................................. 69
Bayesian updating of evidence.................................................................................................71
7: Statistical Computing..............................................................................................................75
Learning objectives: understand use of computers for statistical analysis...............................75
Computer packages for statistical analysis...............................................................................75
What do we need to understand when using statistical computing?........................................76
8: Summarising data.............................................................................................................77
Learning objectives: understand how best to summarise data...........................................77
Graphical data summary....................................................................................................77
Numerical summaries for center of a distribution....................................................................80
What is the mean?....................................................................................................................80
What is the median?.................................................................................................................80
What is the mode?....................................................................................................................80
Comparing the mean, median and mode..................................................................................81
Numerical summaries for spread/deviation of a distribution...................................................82
What is the variance?...............................................................................................................82
What is the standard deviation?..........................................................................................82
What is the interquartile range?..........................................................................................82
9: Estimation and Hypothesis testing.........................................................................................84
Learning objectives: understand how to test hypotheses and estimate population characteristics
using statistics......................................................................................................................84
Estimating means or proportions........................................................................................84
Testing hypotheses about means or proportions................................................................87
Population to sample................................................................................................................88
Sample to population................................................................................................................91
Sample theory...........................................................................................................................91
Making mistakes/errors............................................................................................................94
More precise statistical formulae..............................................................................................98
One-tailed or two-tailed tests?................................................................................................100
Observed significance level....................................................................................................101
What if the population variance is unknown?........................................................................101
The problem of multiple tests.................................................................................................104
Benjamini–Hochberg procedure.............................................................................................104
Extension of hypothesis testing and confidence intervals to other situations........................107
Paired T-test......................................................................................................................107
Two-sample T-test..................................................................................................................110
Effect size...............................................................................................................................113
Categorical data with more than 2 categories........................................................................113
What is the Pearson’s Chi-squared Goodness of Fit statistic? ..............................................114
10: Relationships between pairs of variables............................................................................117
Learning objectives: understand statistical tools for relationships...................................117
Testing for independence of categorical variables.................................................................119
Use of correlation...................................................................................................................123
Simple (bivariate) linear model..............................................................................................126
Residuals.................................................................................................................................132
Meaning for r2.........................................................................................................................................................................................135
Prediction................................................................................................................................135
11: Multiple Regression............................................................................................................137
Learning objectives: understand statistical models for multiple continuous variables.........137
Diagnostics.............................................................................................................................156
Polynomial Regression...........................................................................................................157
Transformations......................................................................................................................159
Normality Tests......................................................................................................................163
12: Analysis of Variance for categories (Factors or Groups)....................................................164
Learning objectives: understand how to model effect of categorical variables on a continuous
variable....................................................................................................................................164
Multiple Comparisons............................................................................................................168
Contrasts...........................................................................................................................170
Nested vs. Crossed..................................................................................................................171
Lack of Fit & Replicates.........................................................................................................172
Nested.....................................................................................................................................175
Model choice objective.................................................................................................................175
Assumptions...........................................................................................................................176
Transformations again............................................................................................................176
Random effects.......................................................................................................................177
13: General Linear Model.........................................................................................................178
Learning objectives: understand how to model effect of categorical and continuous variables on a
continuous variable..................................................................................................................178
14: Model selection criteria.......................................................................................................180
Learning objectives: understand how to identify the best model...........................................180
15: Experimental Design...........................................................................................................182
Learning objectives: understand how best to design an experiment................................182
Main Effect.............................................................................................................................183
Interaction...............................................................................................................................183
Blocking........................................................................................................................................183
Randomization........................................................................................................................184
Blinding..................................................................................................................................184
Placebo..........................................................................................................................................185
Ethical concerns......................................................................................................................185
Completely Randomized Design............................................................................................186
Randomized Complete Block Design.....................................................................................186
Factorial Design......................................................................................................................187
Full Factorial Design..............................................................................................................187
Fractional Factorial Design....................................................................................................187
Principles of Optimal Design.................................................................................................188
16: Time Series..........................................................................................................................192
Learning objectives: understand how to model when the residuals are correlated over time
sequence............................................................................................................................192
17: Generalized Linear Model...................................................................................................200
Learning objectives: understand how to model when the dependent variable follows a
distribution other than the Normal distribution...................................................................200
18: JMP Basics..........................................................................................................................203
Learning objectives: understand basics of using JMP............................................................203
Data Tables.............................................................................................................................203
Variable Type.........................................................................................................................203
Entering Data....................................................................................................................203
Row Selection...............................................................................................................................204
Graphs and Reports..........................................................................................................204
Manipulating Tables...............................................................................................................204
Formulae for creating variables or generating random data.............................................205
19: Data display with JMP........................................................................................................206
Learning objectives: understand how to display data in JMP..........................................206
Univariate distributions....................................................................................................206
Non-normal data.....................................................................................................................207
Multivariate data...............................................................................................................208
20: Model Building with JMP...................................................................................................209
Learning objectives: understand how to build models in JMP...............................................209
Two groups.............................................................................................................................209
Multiple Groups (One-way ANOVA)..........................................................................................209
Multiple Regression................................................................................................................209
General Linear Models with JMP (Multiple Regression + Multi-way ANOVA combined)
(Analysis of Covariance)........................................................................................................210
Multivariate Linear Model with JMP...............................................................................211
Generalized Linear Model with JMP......................................................................................212
Ordinal Data with JMP.....................................................................................................213
Compositional data with JMP................................................................................................214
Time Series with JMP.......................................................................................................214
21: R Basics.....................................................................................................................................215
Learning objectives: understand basics of using R................................................................215
Installation........................................................................................................................215
Data Tables.............................................................................................................................218
Variable Type.........................................................................................................................218
Entering Data....................................................................................................................218
Saving Data.............................................................................................................................219
Manipulating Tables...............................................................................................................219
Creating variables or generating random data..................................................................219
22: Data display with R.............................................................................................................220
Learning objectives: understand how to display data in R.....................................................220
Univariate distributions....................................................................................................220
Multivariate data..............................................................................................................222
23: Model Building with R.......................................................................................................223
Learning objectives: understand how to build models in R...................................................223
Two groups.............................................................................................................................223
Multiple Groups (One-way ANOVA)...........................................................................................223
Multiple Regression...............................................................................................................223
General Linear Models with R...............................................................................................225
Generalized Linear Model with R..........................................................................................226
Binary, Ordinal and Multinomial Data with R.................................................................227
Compositional data with R.....................................................................................................228
Beyond R Commander...........................................................................................................228
24: Hierarchical Linear Models................................................................................................229
Why HLM:....................................................................................................................................229
Statistical inference approaches for linear hierarchical models.............................................231
Maximum Likelihood (ML)..........................................................................................................232
Bayes......................................................................................................................................234
Comparison of classical, ML and Bayes in simple situation..................................................238
Marginal Posterior Distribution.......................................................................................239
Markov Chain Monte Carlo (MCMC)...................................................................................240
References..............................................................................................................................242
BUGS language......................................................................................................................243
Inhalers (ordered categorical repeated measures)..................................................................257
Other examples.......................................................................................................................262
Related tools...........................................................................................................................262
Stan...................................................................................................................................263
Statistical Advice Centre for Students (STACS)......................................................................264
1: Research Methods
Learning objectives: understand theory, model, Occam’s law and types of
random variable
What is research?
‘A systematic and unbiased way of solving a problem (by answering questions or supporting
hypotheses) through generating verifiable data.’ This is the fundamental definition we need, so
we need to understand systematic, unbiased, hypotheses and verifiable, all of which we will
examine later.
Other possibilities: rely on authority (parents, supervisor, police, etc.), personal experience
(what happened when I tried to do this before), common sense (apply simple logic), revelation
(rely on my god to tell me) or intuition (rely on my instincts or feelings).
Let us examine some problems to understand how research compares with the alternatives:
1) Should I cross the road at a specific place where there is no pedestrian crossing?
2) What should the HK government do to improve air quality in the next 20 years?
3) Who should I marry?
4) Should I become a Christian (or Buddhist)?
Conclusion: these other methods may all be useful at times, but not good ways to provide
good long-term solutions to important problems.
In all research, it is important that other researchers can try to replicate your findings.
However, some research cannot be repeated (e.g. effect of handover on Hong Kong) because
the conditions of the data collection cannot be repeated. May be able to reanalyse the data, but
not collect a new set.
Hence high quality data collection is a particularly important issue for social science (but
also for others where data collection is very expensive or difficult to repeat, e.g.
astronomers
studying creation of stars or geologists studying volcanoes). It also explains why in the US
and Europe, research funded with public money must share the data and journals often
require public access to any data analysed in a journal paper.
The reality is that we all make mistakes and have false preconceptions. Society (and the state
of knowledge) can only advance if mistakes can be identified and corrected, which is possible
with research, because people can check the findings for mistakes.
Perhaps this is the key reason that democracy works better than the alternatives, because if
the representative you choose is ineffective, you can try another one?
Research plan
After choosing a research topic, a research plan usually contains most of the following
elements:
We will focus on research design, sample and population, operationalization and the
basics of data analysis and presentation of results in this course (practical data analysis is
a separate course). We will now discuss what is called the ‘scientific method’ of research,
including both qualitative and quantitative methods, which are used in the social,
biological and physical sciences.
Note: the key is objectivity in evaluation, i.e. when testing the ideas and reporting the findings,
not when choosing the research problem (if you have no passion for your research topic, life
will truly be boring!)
Note: unethical evaluation which tries to distort the evaluation, clearly causes problems,
here we focus on unintentional errors.
Note: asking questions can alter respondent views and even for quantum experiments,
measurement affects reality.
Example: survey that asks whether elderly people have discussed with family members the
possibility of retiring in the Mainland – if the answer is no, what do you think they will do
after the survey is finished? This is an example of why survey design requires skill to avoid
bias in the process.
Example: during SARS, we did a repeated telephone survey for the government, asking people
living in Amoy Gardens about their personal hygiene behaviour – this was interventional
research, not observational research, as claimed.
Example: there is known bias in the academic process (publication bias) where evidence
that findings that are different to current mainstream thought, may be selected either for or
against. For this reason, in many countries, clinical trials of drugs must now be pre-
registered, so the results are all made public regardless of outcome. There are researchers
who believe that all research should be pre-registered, to avoid this bias.
Deduction: general to specific (generate ways to test theories in new situations, looking for
a situation where the theory fails)
Induction: specific to general (generate theories from observations, creating a new theory,
replacing any failed theories)
Amusing story: a former Physics chair professor in HK gave a public lecture just before he left
HKU – he claimed that physicists are the only true scientists because they rely on deduction,
not making the mistake of using induction! I wanted to ask him how he thought Einstein came
up with the theory of relativity, if not by using induction?
Systematic Method: a procedure for doing things that can be explained to others and is
built on previously existing knowledge. If you cannot explain to others, how can they
understand it or evaluate it?
What is analysis?
Analysis means both qualitative and quantitative methods (i.e. with and without numerical
information) of processing and summarizing information.
Explanatory: understand why things happen as they do, how reliable is that understanding,
or are there other explanations? For example, if the unemployment rate has risen, we want
to understand why – is it because of young people entering the labour market, is it because
restaurants are sacking dishwashers and replacing them with machines? Statistical
models can be invaluable in this situation.
Predictive: need a model of future outcomes, how well does it work? For example, if we
are looking at admissions to universities, can we predict which students will receive
admission offers before they get their public exam results, if we know their performance
in school exams? This is not hard for the group of all school students, but is very hard for
individual students.
What is a theory?
A theory is a generalised synthetic explanatory statement, in other words, an abstract
conceptual explanation of the world. Conceptual explanations are important in order to
generalise our findings across a wide range of situations, but unless they lead to
predictions, we cannot test our theories, so we need models.
What is a model?
A model is a way to generate verifiable predictions based on a theory, although there may be
uncertainty (randomness) involved and also unknown parameters for the model, that need to
be estimated in our research. For example, if our theory states that increased education leads
to increased personal income and our model predicts that personal incomes increase
proportional to the years of education, we still need to estimate the slope and intercept of that
linear relationship before we can make predictions. We need statistical models to estimate
these parameters, given that there is also uncertainty, as these are not exact models. Even
when there is an exact relationship, there is always uncertainty in practice from
measurement error.
A theory is of little value until it is testable, in other words, we should be able to build a model
that can be empirically tested. If that model fails, then we will need to modify our theory. If
the model does not fail, it does not mean the theory is correct, but suggests the model may be
useful, if and until we are able to find a test that the model fails.
What do we mean by “a model fails”? We mean that there is some inconsistency between what
the model predicts and what happens. We will examine later how to use statistical testing to
identify a model failure.
Clearly, a theory that leads to models that allow us to make good predictions is a useful
theory, but we should not assume that it is necessarily a fundamental truth about the world.
For many centuries, people thought that Newton’s laws were a fundamental truth, until
Einstein showed that some of them fail under certain conditions (speeds close to the speed
of light). Note that we still use Newton’s laws every day, despite Einstein’s findings, because
they are so close to true at everyday speeds!
In short, while absolute truth may exist, we will never know whether we have found it yet
using the scientific method! A famous statistician (George EP Box) said, “all models are
wrong, but some are useful”.
What is a hypothesis?
A hypothesis is a statement that can be empirically tested, i.e. translation of theory into a
testable statement. Note that a research question may be as specific as “Do young people
protest more than older people?”, which links to the hypothesis: “Young people protest more
than older people” or may be a broader question, such as “Why do HK people protest?” which
does not link clearly to any hypothesis.
The research (or alternative) hypothesis is a positive statement about what the researcher
expects to find, like “Young people protest more than older people”
The null hypothesis is a statement that a relationship expected in the research hypothesis
does not exist, i.e. that the world is simpler than predicted by theory, like “There is no
difference in how much young and older people protest”.
Example: Consider the proportion of males and females studying undergraduate degrees in the
whole of HKU. The simplest explanation would be that there are equal proportions (50%), so
this might be my null hypothesis, while I might expect the proportion to be unequal; hence this
is my research hypothesis.
Note: research papers do not always explicitly state the hypotheses, but if they are testing
whether there is a relationship between education and income, the implicit null hypothesis
is that there is no relationship and the implicit research hypothesis is that there is a
relationship.
Question: why do I think my hypothesis of equal proportions is simpler? This yields two
questions – what is simple and why do I care about simplicity?
Karl Popper: We prefer simpler theories to more complex ones because their empirical content
is greater and because they are better testable
Richard Swinburne: Either science is irrational in the way it judges theories and predictions
probable or the principle of simplicity is a fundamental synthetic a priori truth.
Ludwig Wittgenstein: Occam's principle is, of course, not an arbitrary rule, nor one
justified by its practical success. It simply says that unnecessary elements in a symbolism
mean nothing.
Signs that serve one purpose are logically equivalent; signs that serve no purpose are logically
meaningless. The procedure of induction consists in accepting as true the simplest law that can
be reconciled with our experiences.
Example: the hypothesis that the world is pre-determined is unbelievably complex and of no
use for predicting the future, so it will be a very low priority in our set of theories.
If statistical models are used, we expect the null hypothesis to be rejected and hence the
alternative (research) hypothesis to be tenable (believable). The null and the alternative are
thus usually complementary.
The null hypothesis is usually the simpler statement, such as there is no effect caused by
something, while the research hypothesis would be that there is an effect, or that there is a
positive effect or that there is a negative effect. If the paper uses statistical methods, there
must be a null hypothesis and a research hypothesis, even if they are not clearly stated.
We can only disprove, rather than prove hypotheses, hence we look for evidence to disprove
the null hypothesis and if we cannot find sufficient evidence, we accept the null hypothesis.
Strictly speaking, statistics cannot usually even disprove a hypothesis, it can at most show
that the outcome we observed was (very) unlikely, if we assume that the null hypothesis is
true. This is because statistics uses probability (chance) statements rather than absolute
statements. Thus we are usually prepared to reject the null hypothesis as unlikely to be true if
there is sufficient empirical evidence against it. We will come back to the question of what
constitutes sufficient evidence.
Example: if I toss a coin 20 times and every time I get a head, would you believe this is a
fair coin (i.e. has 1 head and 1 tail)? The chance is about 1 in a million of observing this
outcome with a fair coin, but a certainty if my coin has 2 heads!
How do we define simplicity? Often it is defined in terms of the number of parameters in our
statistical model or the sample size minus the number of parameters in our statistical model,
which we call the degrees of freedom (we will revisit this), where simplicity means the number
of parameters is small or the degrees of freedom are large.
What is proof?
Proof means that you know the truth for certain.
What is verified?
Verified means that your tests did not disprove the truth.
Example: if I compare your face to your picture in your identity card or passport and they
match, I have verified your identity, but it is not proof because you might have an identical
twin or have obtained a fake document.
If you test many hypotheses resulting from a theory, and none of them are shown to be false,
you may think the theory is true because of numerous verifications, but you still have not
proved it.
It may be that the weaknesses in the theory have not been identified yet (e.g. theory of
relativity and Newton’s laws).
What is a constant?
A constant is something that (it is assumed) does not vary over the study and hence cannot
explain anything that does vary over the study.
Creating a good operational definition is a skill and it is important to look at previous work
(hence the importance of a good literature review). We will discuss operational definitions
later, including evaluation of operational definitions.
Assumed irrelevant: extraneous variables believed not to play any role in the research
Clearly, the research is at risk if there are important uncontrolled variables or if the
variables assumed irrelevant are confounding variables. It is important that in stating our
research problem that we give careful thought (and check the research literature) to allocate
variables correctly. In practice, there are limits to how many variables we can control for,
so we need to ensure that we control for those with the largest potential effect on our
explanatory variables. Controlling is an attempt to exclude the effects of variables that are
not of interest.
Mechanisms of control
One category: fix all subjects to have the same value (category) of a controlling variable
(e.g. only study males). Disadvantage is that the conclusions cannot be generalized to
other values of the controlling variable.
Block control: divide up subjects by value of controlling variable and study separately (e.g.
study males and females separately). Good for simple situations with only a few blocks but
often impractical for multiple controlling variables as the block size gets too small (too few
subjects per block).
Paired control: pair subjects to be similar on several variables and randomly assign to
groups from pair
Statistical control: try to control using a statistical model - this assumes that the statistical
model is good enough to completely remove the effect.
Measurement error: In practice, none of these controlling mechanisms are perfect, because
there may be errors in the measurement of the controlling variables. Example: early passive
smoking studies looked at non-smoking women with smoking husbands, but did not account
for the problem that some people lie about smoking (but not usually lie about not smoking) and
that it is more likely for people to lie if you are married to a smoker, which together led to bias
in estimate of passive smoking effects. This problem can be solved if we can eliminate or
model the measurement errors (in our example, we can check urine or hair samples to check
how much people smoke).
Note that controlling methods that rely on allocation are only feasible if the researcher can
allocate subjects to groups, which may be infeasible or unethical (e.g. smoking for humans).
The control variables may also interact (e.g. effect of income within males may differ from
within females).
Propensity scoring: when the allocation is non-random, we can look at possible bias by
examining the conditional probability that a subject will be treated given the observed
explanatory variables (the propensity score). If we stratify by propensity score, we obtain
estimated treatment effects that are not biased by the measured explanatory variables - see
Rosenbaum and Rubin’s 1983 Biometrika paper: “The central role of the propensity score in
observational studies for causal effects”.
Example: The classic 1950 paper by Doll and Bradford Hill in the British Medical Journal:
“Smoking and Carcinoma of the Lung” examined frequency tables of lung cancer or control
(other hospital patients matched by gender, age and hospital) by gender and smoking or not.
The research hypothesis is that smoking is associated with lung cancer while the null
hypothesis is that there is no association. Lung cancer or not is the dependent variable,
smoking or not is the independent variable, while gender, age and hospital are control
variables. Whether a random variable is an independent variable or control (cannot be both)
may not be obvious without checking the research hypothesis (independent variable is
included, control is not).
Mendelian randomization: Mendel discovered that when he crossed purebred white flower
and purple flower pea plants, the result was not a blend. Rather than being a mix of the two,
the offspring (known as the F1 generation) was purple-flowered. When Mendel self-fertilized
the F1 generation pea plants, he obtained a purple flower to white flower ratio in the F 2
generation of 3 to 1, suggesting that separate genes for separate traits are passed
independently of one another from parents to offspring. We can use the random assignment of
an individual’s genotype from parental genotypes that occurs before conception to make
causal inferences (assuming that the genotype is associated with the exposure of interest and
independence of the genotype from confounding factors and from the outcome given the
exposure and confounding factors). The strength of this is that it does not rely on allocation.
2: Association and causation
Learning objectives: understand association & causation
What is association?
Association is observed linkage, i.e. two outcomes X and Y tend to both occur (positive
association) or tend to occur separately (negative association). This is separate from the
question of whether X causes Y, Y causes X or W causes both X and Y.
What is causation?
X is a cause of Y if when X occurs, Y must occur. This is called X being a sufficient
condition for Y (e.g. breaking your spinal cord is sufficient to cause paralysis). If Y only
occurs if X occurs, then X is a necessary condition for Y (e.g. drinking alcohol is currently
necessary to know if you are an alcoholic). If X is necessary and sufficient for Y, then X is
the one and only true cause of Y. These are stringent conditions that are not often met!
Note: if X occurs (or not) after Y, X cannot be a cause of Y (unless you believe in time
travel!), so let us consider situations where X occurs (or not) before Y occurs (or not).
If X and Y sometimes both occur and sometimes only one of them occurs, we can only make
probabilistic statements such as, when X occurs then Y is more likely to occur (than when X
does not occur). Clearly, this means that something else must also be determining whether Y
occurs, so our model is incomplete.
What we seek in practice is a model that enables us to predict (with high probability) whether
Y occurs, given X occurs (or not).
Strictly speaking, we cannot move from association to causation, but only work the other
way, that is, causation implies association, but not necessarily the reverse. This should be
obvious, as association does not indicate the direction of causation. In short, causation can be
tricky!
Example: visibility and health outcomes – poor visibility is not a cause of poor health, but
poor air quality is a cause of both poor visibility and poor health.
Example: exercise and physical disability – newspaper article claiming that lack of
physical exercise was a cause of physical disability – true in theory, but the association
is mainly a consequence that exercise opportunities are restricted for the physically
disabled
Example: juvenile crime and reading violent comics – social workers believed that reading
violent comics was a cause of juvenile crime, but life histories showed the major linkage was
that juvenile crime led to detention, which led to expulsion from school, which led to
hanging out on the street reading comics with similar youth!
Example: having breakfast and school performance – a newspaper reported that the researcher
claimed that eating breakfast made almost one year’s difference in school performance, with no
mention of other potential underlying causes.
Example: the original study of smoking and lung cancer in doctors by Sir Richard Doll was
very persuasive because he showed that the risk increased with the amount smoked and later
showed that doctors who stopped smoking had decreased cancer risk, closer to non-smokers.
Observatory
What is a sample?
A sample is the respondents selected from population for study
Example: 500 mobile phone users in Hong Kong, selected randomly using their phone number
Example: 1000 average daily ozone levels recorded sequentially at the street level
monitoring station in Causeway Bay
Example: a simple random sample of size 2 of students in a class with 10 students means that
all 45 (1092) possible pairs are equally likely to be selected.
What is a cluster sample?
Cluster sampling means sampling at two or more levels (e.g. school, class, student), usually
require that each individual has the same chance of selection overall.
Example: a sample of 500 secondary students is hard to draw without a list of all students.
However, we can easily obtain a list of secondary schools. We sample schools, classes
within schools and students within classes. In practice, we often sample all the students in
a class, as the additional cost of sampling an extra student is low once we have disturbed a
class. Cluster sampling requires disturbing many less schools than a simple random sample
– which would require obtaining consent from nearly all schools.
Disadvantage: students in the same class are likely to be more similar than students drawn
at random (similar school philosophy, recruitment, teachers, social interaction). Intra-
cluster correlation can be used to adjust the sample size (the effective sample size is
smaller than a simple random sample).
Example: it is very hard to sample drug addicts in general because they are not all easily
identifiable, but they are usually connected in drug supply networks. A network sample can
be drawn by asking an initial set of drug addicts for introductions to their drug-using
friends and then asking a sample of those friends, etc.
What is a distance sample?
Distance sampling involves randomly placing points or lines (transects) on a map and then
measuring distances to objects. There are two main approaches – point transects and line
transects. In both cases, we are trying to estimate the density of objects by counting how many
we can see (from a point or line). The complication is that we need to account for the chance of
detection being different at different distances. Examples would be estimating the number of
people marching along Hennessy Road (usually count the number going under a bridge in
random sample of time periods) or the number of people in Victoria Park (usually look at
density in random sample of small areas).
Note: In the rest of this course, we will assume simple random sampling, for simplicity.
Example: any telephone survey that samples numbers directly from the Chinese language
telephone directory in HK misses many Chinese writing families that prefer not to be easily
found and all families that do not write Chinese so cannot claim to represent the Hong Kong
population.
Example: any household survey that fails to make repeated (at least 5) attempts to
contact households will under-represent poor and single working households (who are
often out at work) and young households without children (out at work or play).
Example: a self-selected Internet survey omits most of the population who are not interested in
the topic, do not see the survey or do not trust the website, so is likely to be of very limited
research value.
Telephone versus face-to-face interviews:
Most interviews now use telephone or face-to-face. Telephone loses visual contact, gains
on lower cost, better supervision, faster collection, ease of computer assistance (CATI
stands for Computer Aided Telephone Interviewing, which automates question display,
telephone number selection etc.). In Hong Kong, fixed line telephone penetration is still
high (80%+) and it can be difficult to gain entrance to private housing estates, although the
compact geography makes face-to-face interviews in households possible. For children,
schools provide a context for face-to-face or self-report.
Example: study that drops envelopes (with or without stamps) in different cities and sees what
proportion of them is posted as an assessment of social behaviour in different cities.
There is a very wide range of methodologies (approaches to collecting and analysing qualitative
data), which include:
Example: Cancer patients reveal their own experiences of being treated for cancer, which offer
insights into different aspects of the care and the people who provide it.
Phenomenology
This is observing what happens in order to gain understanding of what really happens
rather than what people tell you happens. It enables the researcher to gain some
understanding of what it feels like for the subjects to be living a particular experience.
Example: “If you want to know why athletes are willing to take steroids – you need to
understand their lived reality of winning and losing. If you want to help someone through
breast cancer – you need to know how they feel about their body, their self-esteem, and their
future. If you want to understand how you can help motivate struggling students – you need
to know what it is really like for them at the bottom of the class.” (O’Leary, Z. (2010). The
Essential Guide to Doing Your Research Project. London: SAGE, p.119)
Grounded theory
This is a form of analysis constructing new theories from the data, i.e. qualitative induction. It
generally consists of four stages (Glaser, B. & Strauss, A. (1967). The Discovery of
Grounded Theory: Strategies for Qualitative Research. Chicago: Aldine):
1. Observing the data to identify patterns that lead to the emergence of categories, then
identifying the underlying properties of the categories.
2. Integrating categories and their properties: Comparing an incident to the underlying
properties of the category.
3. Delimiting the theory: Fewer and fewer modifications are needed as the categories are
confirmed, then the number of categories can be reduced as further refinements take place.
In this way, the theory begins to solidify.
4. Writing theory: Hypotheses and generalizations emerge from the analysis, as
opposed to starting with and testing hypotheses.
Example: O’Reilly, Paper & Marx (O’Reilly, K., Paper, D. & Marx, S., Demystifying
grounded theory for business research. Organizational Research Methods. 15, 247, 2012)
described a study of business segments that lack effective communication and cooperation.
They wanted
to understand the views, perceptions, and beliefs of front-line employees (FLEs) and how
these perspectives might help improve customer-company interactions. They examined
differences in service levels, service outcomes, and service attitudes of the participant
companies and FLEs who work there and, by comparing the participants’ stories they
identified the constraints within the organizations and their impacts.
Ethnography
This is the anthropological approach of becoming part of the culture in order to understand it.
Example: working in a factory in order to understand what it means to be a worker both from
direct experience and asking/observing others.
Case study
This involves studying a small number of cases in great depth in the expectation that this
gives deep insights into the process
In all cases, the focus is on understanding the full scope of the problem, rather than
quantifying the problem. Arguably, qualitative research is a necessary precursor for
quantitative research unless the scope of the problem is already well understood (e.g. a
study aiming to understand how sports participation has changed over time might not need
qualitative research unless we think the underlying drivers of participation might have
changed).
Example: Silverman (Silverman, D. (2006). What is Qualitative Research?
http://www.sagepub.com/upm-data/44074_Silverman_4e.pdf and http://www.sagepub.com/upm-
data/11254_Silverman_02.pdf) conducted a case study of British cancer clinics to form an
impression of the differences in doctor-patient relationships when the treatment was private or
public. He cautioned that his data could not offer proof of the differences he identified, but
that they provided strong evidence to support them.
Qualitative sampling
Probabilistic notions of sampling are not relevant for some methodologies. It is not always
possible to achieve representative sampling because of the exploratory nature of the
research and the sheer logistics. Depending on the nature of the research, it is sometimes
necessary to select a sample that meets a particular need. Usually, with this kind of
sampling, it is not appropriate to generalize the data to wider populations. The key ideas of
qualitative sampling are:
Saturation
The idea is to collect data until no new perspectives are being obtained. This means that the
sample size cannot always be predetermined.
Example: You are interested to find out how young children learn to write Chinese characters.
You observe children engaged in writing tasks and analyze samples of children’s work.
Eventually you become aware that the same patterns are being repeated and there is
nothing new that has not arisen before. There is no point to collect further data as no new
perspectives are being obtained.
Convenience sampling
This is sampling driven by the feasibility and convenience of the selection process. Some
people criticize that it does not have a place in ‘credible research’ (O’Leary, 2010), but it may
be the only option for a small, low-budget study or a pilot.
Example: A group of recent graduates is invited to volunteer to attend an interview about the
impacts of their undergraduate programme on their professional lives. Only a limited number
of students are able to be contacted or willing to make themselves available, so the researchers
need to utilize those who can be accessed and are willing to participate.
Snowball sampling
Snowball sampling assumes relevant respondents are connected so that we can use
those connections to construct a sample from a small initial sample. In other words, it
involves building a sample through referrals, as each respondent recommends others.
Example: A population of homeless people might not be easy to identify, but a sample can
be built by using referrals (O’Leary, 2010).
Observation
Observation is the collection of existing data. It usually takes place in a real situation, not
a contrived context and captures first-hand what people actually do in the situation as
opposed to telling the researcher about what they do.
Example: A school district introduces a child-centered learning approach and wants to collect
authentic data about how the teachers are actually implementing this approach.
Interviewing
Interviewing can be used to provide rich qualitative data and provides flexibility to explore
tangents (O’Leary, 2010). Good interview questions can elicit data about whom, when, why,
where, what, how and with what results (Hutchison, A., Johnston, L. & Breckon, J. Using
QSR-NVivo to facilitate the development of a grounded theory project: An account of a
worked example. International Journal of Social Research Methodology. 13, 4, 283-302)
Interviews can be structured, semi-structured or unstructured.
Example: What are the perceptions of carers living with people with disability, as regards
their own health needs? (Lacey, A. & Luff, D. (2009). Qualitative Data Analysis. The
NIHR RDS for the East Midlands/Yorkshire & the Humber. http://hk.bing.com/search?q=the
%20NIHR%20RDS%20for%20the%20East%20Midlands/Yorks hire%20%26%20the%20Humber
%202009%20Qualitative%20Data%20Analysis&FORM=AARB LB&PC=MAAR&QS=n )
Example: In the research about the perceptions of carers living with people with learning
disabilities about their own health needs, there are different ways in which the data could
be analyzed depending on what the researcher is interested to explore (from Lacey &
Luff, 2009): Content analysis: Count the number of times a particular word or concept
(e.g. loneliness) appears – categories these quantitatively and do statistical analyses.
Thematic analysis: Find all units of data (e.g. sentences or paragraphs) referring to loneliness,
code them and look for patterns (e.g. certain times and conditions where carers feel lonely).
Theoretical analysis (e.g. grounded theory): This goes further to develop theories from the
patterns in the data. It may include data that contradict the theory. Gradually the theory is
built and tested.
What is reliability?
Reliability in this context (it has other meanings in statistics and in daily life) means
consistency: do we get the same result if measured repeatedly - various types of reliability:
Test-retest: ask respondents again, i.e. measure a second time using the same sample
and same instrument.
Split-half: compare the same sample using different parts of the same instrument (applies to
survey instruments, where we often assess using a combination of questions)
Note that mathematical formulae should be perfectly reliable, as they always give the
same answer.
What is validity?
Validity is about whether our measurement really measures our concept (or something
else)? Is it meaningful as a measurement tool for this concept?
Face validity: does it even look like it is measuring the right thing (e.g. Assess weight
by asking how much money in your pocket!)
Criterion (predictive) validity: does it predict outcomes that we believe relate to the
concept (e.g. If heavy people have more diabetes, does our measure of weight predict
diabetes?)?
Construct validity: does it relate to other variables in the way our theory predicts? (e.g. if
we are measuring marital satisfaction and our theory says that it should be associated with
marital fidelity, is that true?)
Content validity: does it cover the range of meanings contained in the concept (e.g. the
concept prejudice contains prejudice on grounds of race, minority, gender etc.)
Note that the word validity is also used in evaluating a research design – does it have
internal validity (are the conclusions about the independent variables causing the
dependent variables correct, see the earlier discussion on association versus causation) and
does it have external validity (can the findings be generalized outside the conditions of the
study).
What is precision?
How precise is the measurement - e.g. is age measured to the nearest decade, year, or month.
There is little point in having precision that is not needed, but if the precision is too low,
cannot fix it later. Depends on situation, so age in months is too broad for newborn babies, too
narrow for the elderly! Often has little impact on cost, unless it is sensitive data which people
are reluctant to reveal (e.g. exact age or income). However, very precise measurement may be
impractical, e.g. exact income may be hard to calculate for people who do not have a fixed
monthly income.
Note that there may be some trade-off between reliability and validity. Highly reliable
measures are likely to be factual and give little insight and may not have high validity (they
are often quantitative). Measures that have high validity may be very personalised and
have relatively low reliability (they are often qualitative). Personalised does not necessarily
mean it is not objective, it can just mean that the measuring instrument does not work the
same way for all sample elements. Objective assessment will almost always be more
reliable, but we should not assume that it must be more valid than subjective assessment
(consider measuring quality of culture or a user interface). A good strategy is often to use
multiple measuring tools simultaneously to measure different dimensions.
Possible operationalizations:
Migration rates – reliability: in HK and Singapore we know population flows quite accurately,
but when someone leaves, we do not know if is temporary or permanent, while in Europe, do
not know population flows accurately because of weak national border controls, making the
figures imprecise – validity: positive is that if it is a good place to live, we expect people to
come and if it is a bad place, we expect people to leave, negative is that it can be hard to get
approval
to move countries legally unless you have skills, money or family reunion so it may be
measuring connections and abilities of the individual instead and also the US and parts
of Europe have many uncounted illegal immigrants, so migration rates are consistently
underestimated.
Poll people in HK, Singapore and a 3rd place – reliability: depends heavily on media reports,
if they change then the poll will change – validity: is it based on personal experience as a
resident or tourist, family, friends or media reports, all of which may be biased
Suicide rates – reliability: death rates are very reliable, but not all suicides leave notes –
validity: suicide can be seen as only way to escape terrible life, but it also can be seen as
untreated mental illness, so arguably both elements imply very poor quality of life for some
people
Air pollution levels – reliability: quite high as based on physical/chemical processes – validity:
high for people who are asthmatic, but clearly this is only one element in quality of life
Ask experts to rate each place on a set of living characteristics – reliability: probably high
if they are experts on those characteristics – validity: how do we know if those
characteristics are relevant to people considering living in that place?
Note: quality of life has a strong personal element, so it is possible that for so people Singapore
gives higher quality of life (those who value air quality?) while for other people Hong Kong
gives a higher quality of life (those who value freedom of speech?)
There are many other important operational issues for questionnaires including clarity of
questions, ensuring that respondents are competent to answer, questions should be relevant,
keep questions short, avoid negative items, avoid biased items (wording is critical, positive
vs. negative words). Questionnaire design is an art as much as a science and it is essential
to do a pilot (pre-test) to check that the questionnaire works as designed (pre-tests usually
involve qualitative analysis)
(Questionnaire design is worthy of a course itself and the Faculty of Social Science runs
a course each summer)
Detail: how much detail is useful and collectable, not just units (as in precision), but detail of
categories
Example: Single, married, others - do you need to break down others? Interesting study of
mental health – for men, deterioration after divorce or widowhood, while for women,
deterioration after widowhood, but not divorce!
Dimensions: a concept may have many dimensions, which are of interest in the research?
Example: corruption: how much, what causes it, what should be done, what would they do
personally etc.
Example: when assessing outcomes from medical treatment: life/death, survival time after
treatment, quality of life, pain, blood pressure, white cell count etc.
Levels/scales of measurement:
Attributes must normally be mutually exclusive (no overlap) and exhaustive (cover all
possibilities)
Example: underemployed, employed or unemployed, cannot fit into more than one category
and must fit into at least one category if economically active.
Index counts the number of positive (or negative) responses, while a scale allows for
intensities (levels of response). When creating an index, we want the underlying items to
be measuring the same thing - usually weight items equally (i.e. just add them up). Often
do item analysis (reliability check if each item is consistent with the rest of the items) and
external validation of complete scale. In Hong Kong, often take scale from other
cultures/languages, but the translated scales may not work in HK (concepts may not
translate or attitudes may be different).
Note: Likert scale does not guarantee that the difference between strongly disagree and
disagree is the same as between agree and strongly agree, so should not treat it as
interval scale without checking if this is valid
What is semantic differential scale?
Semantic differential scale means asking people to choose positions between two polar
opposites. Examples: love to hate, simple to complex
Example: everyone who agrees to abortion on demand should logically also agree to abortion
after rape
6: Probability
Learning objectives: understand three laws of probability
What is probability?
Probability is the chance (long-run relative frequency) that something will happen. In other
words, if you perform an experiment many times, what proportion of the time does a particular
outcome occur?
Example: toss a coin, if the coin is fair then heads and tails are equally likely, so after a large
number of tosses the relative frequencies should be close to one half, meaning long-run relative
frequency should be one half for heads and for tails.
While statistical formulae are largely irrelevant to users, the language of statistics
(probability) is important and some knowledge is helpful. Anyone who gambles or invests
should know the basics of probability – it is easy to show that failure to follow the rules of
probability guarantees that you will lose on average when gambling with someone who does
follow the rules.
What are the rules that help us work out the chances for combinations of outcomes?
Start with a set of outcomes that are mutually exclusive and exhaustive (i.e. they are unique
and contain all possible outcomes) (e.g. if we toss a coin, 2 possible outcomes, Head or
Tail), we call this set the sample space.
All probabilities must be in the range between zero and one (obvious in terms of relative
frequency).
The sum of the probabilities of all the mutually exclusive and exhaustive outcomes must add
up to 1, because one of the outcomes must happen.
e.g. Pr(H)+Pr(T)=1
Pr(HH)=Pr(H) Pr(H)
For this double experiment, there are 4 outcomes, HH, HT, TH, and TT
Pr(HH)+Pr(HT)+Pr(TH)+Pr(TT)
=Pr(H)Pr(H)+Pr(H)Pr(T)+Pr(T)Pr(H)+Pr(T)Pr(T)=Pr(H)(Pr(H)+Pr(T))+Pr(T)(Pr(H)+Pr(T))=1
So, this is quite simple except for we usually are interested in the number of occurrences
rather than the sequence. For this double experiment, we can get 0,1 or 2 heads in 2 tosses.
So if p=chance of a head (and 1-p is then chance of tail), then
Pr(2 Heads)=Pr(HH)=p2
Pr(1 Head)=Pr(HT or TH)=2p(1-
p) Pr(0 Heads)=(1-p)2
This case is easy, but in general, we must calculate the number of combinations, e.g. if we toss
a coin 10 times, what is the chance of 9 heads (and 1 tail)? Need to be able to work out the
number of different ways that this can happen (10, because the tail could occur on each of the
10 tosses). Harder would be the chance of 8 heads (and 2 tails). Now the number of ways
is 10 x 9 /2= 45, which is written mathematically as
2=10!(8!2!)
10C
and so on, but the formula is not important as we can use tables, calculator or a computer.
Binomial distribution
This set of probabilities for how many successes we get is called the binomial distribution and
is very important in statistics. The usual notation is B(n,p), where n is the number of trials
(tosses) and p is the chance of a success in one trial.
We will see how to use the Binomial distribution later on to check if a coin is “fair”.
What is conditional probability?
The idea is to calculate the chance of an event A, if we already know that event B occurred.
Example: if I tossed a coin 5 times and you happen to see that the last toss was a head,
how should that affect your estimate of how likely it was that I got all 5 heads in 5
tosses?
A)Pr(A)/Pr(B)
Pr(A and
B)=Pr(A)Pr(B) In our
example, we get:
Pr(HHHHH|last is H) = Pr(HHHHH)/Pr(last is H)
The important thing for this course is to “get” the idea of conditional probability and know of
Bayes’ Law (even if you need to look it up!)
Note: this law is very useful if you are a card player, whether bridge, poker, blackjack or the
like as it lets you calculate the probability of an event given some indirect information.
Example: if I have a pack of well-shuffled cards and the first card I deal is an Ace, what is the
chance that the second card is also an Ace if use the same pack without shuffling?
Bayes’ Law is also the key idea behind many spam filters. They look at the probability of an
email having these characteristics if it is or is not spam and then calculate the probability of it
being spam given these characteristics. They update the probabilities as new emails come in
(this is called the training process). That is why they are often called Bayesian filters (see
https://en.wikipedia.org/wiki/Naive_Bayes_spam_filtering for details)
When there are two possible outcomes (Head/Tails, Win/Lose etc.), it is often easier to think of
odds instead of probabilities, where
We will see that Bayes’ law is easier in terms of odds when there are 2 possible outcomes.
So if we are looking at A or not A given B
Example: you are evaluating whether it is a good idea to implement urine drug screening on
all secondary school students in Hong Kong (about 400k) (this is loosely based on the real
situation in Hong Kong, although the HKSARG seemed unaware of the need for a quantitative
assessment like this and was unaware of the need for obtaining performance data for the
selected screening tool, which was a relatively cheap drug test that could be done in the school,
rather than in a laboratory):
Assume that before screening, the odds that someone has taken a specific drug are 1 in 1,000.
If the person has taken the drug, the probability of a positive screen is 0.98 (false negatives of
2%) and if they have not taken the drug, the probability of a positive screen is 0.02 (false
positives of 2%) (these conditional probabilities were apparently not known by the HKSARG,
so they are taken from data I found in an Australian study)
What are the odds an individual student has taken the drug after a positive or a negative
screen?
Pr(D)/Pr(not D)=10-3
Pr(S|D)=0.98
Pr(not S|D)=0.02
Pr(S|not D)=0.02
Pr(not S|not
D)=0.98
so
Pr(D|S)/Pr(not D|S) =Pr(S|D)/(Pr(S|not D) x Pr(D)/Pr(not D) =0.98/0.02 x 10-
3 (statistical jargon is posterior odds = likelihood ratio x prior odds)
Note: The government defended their position on the basis that all positive would be double
checked in the laboratory using much better tests. However, the testing scheme would all
know that a student tested positive on the screen, which would do great damage to the
school if a false positive was leaked to the media.
7: Statistical Computing
IBM SPSS - very popular for desktop use, but expensive and not very flexible (in terms of
adding new procedures). Free to use for current HKU staff and students:
https://www.its.hku.hk/services/procurement/software/license#spss
SAS - popular in finance sector, very flexible but less user-friendly than SPSS
JMP - user-friendly, with a free demo version at http://www.jmp.com. You can purchase an
educational license for the full version at http://www.onthehub.com/jmp/ for US$49.95 for a 12-
month license (US$29.95 for 6 months).
R – open source, very popular with statisticians, but not so easy to use.
All of these have PC and Mac versions.
Many other software tools can also be used to do statistical calculations, although statistics is
not their primary focus (e.g. Excel, Matlab etc.)
Please note that thanks to computers and statistical packages, we rarely need to do statistical
computations by hand and hence the formulae are largely irrelevant.
Stem-and-leaf: display sorted list of numbers, where for 2 digit numbers, the tens digit is the
“stem” and the ones digit is the “leaf” to illustrate the data frequencies (the row length)
without losing sight of the underlying data. Stem-and leaf can be used for any number of
digits, but it is simplest with 2 digits. We need to divide up the digits into those used for the
stem and those used for the leaves. We can scale the numbers and note the scale below (for
example with 1 decimal place, multiply by 10 so that the numbers are integers). Easy to do by
hand with paper and pencil!
Notes: stems and leaves must be sorted, must include duplicates, stems with no leaves must
be included.
Variations include splitting the stems into 2 stems if the number of stems is small, i.e. put
leaves 0-4 on one stem, 5-9 on another, so the previous plot becomes
4|669
4|02
3|
3|
2|
2|04
1|6
1|24
Bar chart: chart with height of bars proportional to frequency (standard chart in Excel)
Histogram: bar chart with area of bar proportional to frequency (area is equivalent to height
if we have equal widths)
Example: We choose (equal) group widths of 5, so the bar chart and histogram are same:
Size
Box Plot: box shows central half of the data, with a line showing the middle of the data
and with whiskers and dots showing more extreme data.
Size
Note that this does not show the separation of the data.
Numerical summaries for center of a distribution:
What is the mean?
The mean is the average value (i.e. Add all the values up and divide by the number of values).
We often use the Greek symbol mu: for the population mean, often write the sample mean as
a X with a bar above it, X¯ and the formula for the sample mean as xi/n, which means adding
up all the values and dividing by n, the sample size, so sample mean=309/10=30.9 for our
example.
If n is the sample size and is odd, the median is the (n+1)/2 biggest (or smallest) value, if n is
even number of values, the median is the average of the 2 middle values (the n/2 and n/2+1
values). The median can easily be read from the stem-and-leaf plot, as the data points are
already sorted. Hint: you can check your answer by counting up and counting down to check
that you get the same answer, which is the average of the 5 th and 6th values, i.e. (24+40)/2=32
for our example.
The mode is not always unique and is not robust (changing one data value slightly can cause a
large change), so usually use the mean or median.
Mean is easier to deal with mathematically, so most theory is based on the mean, but
the median is more robust (reliable).
Example:
Size
Median 32
Mean 30.9
Mode 46
Median 32
Mean 75
Mode 46
The mean changes substantially, but in this case, the median does not change at all. Also
transforming the data usually transforms the median, so the median of log(income) is the log of
the median of income (because log function does not change the order of the data values, the
log of the middle value is the middle value of the log values), but this isn’t true for the mean.
Numerical summaries for spread/deviation of a
distribution: What is the variance?
The variance is the average squared distance of data values from the mean. This seems
complicated, but it gives a simpler theory. Population variance is denoted 2 (sigma squared)
and sample variance is denoted as s 2. The formula for the sample variance is written as (xi-
mean)2/n.
Note: if using sample mean instead of population mean, we adjust the divisor to be (n-1)
because one data point is “used up” by the sample mean)
The standard deviation is the square root of the variance, i.e. measure of spread in the original
units. Often use Greek symbol sigma, for the population standard deviation and s for the
sample standard deviation.
Example:
Size:
Standard deviation 15.0
Variance 225
Interquartile Range 30.5
9: Estimation and Hypothesis testing
Learning objectives: understand how to test hypotheses and
estimate population characteristics using statistics
Note: this idea only works if the sample is representative, i.e. all individuals have equal chance
of selection. If the individuals all have known (non-zero) chance, we can adjust (weight)
respondents and still estimate population proportions.
Our best estimate is 520/1000=0.52, i.e. the sample proportion is our best estimate of the
population proportion.
We will simplify things slightly for now, by assuming that the total no. of children in our
population is much greater than 1000 (e.g., if only 1000 in the population, then our
estimate would be exact). It is rare that our sample size is more than 1% of the population
size.
Note: If the sample size is more than 1% of the population, we should make an adjustment to
the following calculation, as the estimate will be more accurate than we estimate here.
The jargon we use for describing the accuracy of our estimate is that we say that the
approximate 95% Confidence Interval (C.I.) for the population proportion is:
p +/- 2 √( p x (1-p) / n)
where p is the sample proportion and n is the sample size. This approximation is good as
long as the sample size is at least 30 and the sample proportion is not too close to either 0
or 1.
Otherwise, we need to use statistical tables.
In percentage terms, our interval contains values between 48.8% and 55.2%
We call this interval the 95% C.I., which means that, on average, we would expect that in 95
out of 100 samples, our confidence interval would contain the population proportion. This is
already a hard concept as it is not a probability statement about our specific interval, but
instead what would happen if we repeat our sampling.
We will come back to these ideas in a more careful way later and explain where the formula
comes from.
Testing hypotheses about means or proportions
If our concern was whether there was a sex imbalance in the schools, our hypotheses would
1 for now
Question: is there evidence that our null hypothesis is false (relative to our alternative)?
Idea: find a good summary of the data (called our test statistic) that best summarizes the
evidence for choosing between the hypotheses. In this case, the sample proportion is the
best test statistic. If the test statistic has a value that is unlikely to occur if the null
hypothesis is true, then we reject the null hypothesis (strictly speaking we mean that this
value or more extreme values are unlikely, rather than that one value is unlikely).
Implementation: find the probability of observing the test statistic (or a more extreme value) if
the null hypothesis is true. If this probability is below a cut-off value (the cut-off is called the
significance level), we reject the null hypothesis, on the grounds that what we observed is not
consistent with the null hypothesis.
In this case, the probability of observing a sample proportion of 0.52 or a more extreme value
(i.e. More than 0.52 or less than 0.48) is about 0.2 if the population proportion is truly 0.5 (this
probability will usually be given in the output from our statistical package).
The usual significance levels used are 0.05 (5%) or 0.01 (1%). This means that we choose to
reject the null hypothesis when it is really true 5% (1 in 20) or 1% (1 in 100) of the time.
Which significance level we use as cut-off values depends on what risk we accept for making
the mistake of rejecting a true null hypothesis (we will return to this topic more carefully later
as well).
Population to sample
Given some knowledge of a population, what can we say about samples from that population?
The average of the sample means tends towards the population mean as the sample size
increases as long as we have a probability sample. In fact, the mean of all possible
sample means is the population mean (we call this property unbiasedness, meaning that
we get the correct answer on average).
The sample variance is a good estimator of the population variance for sample sizes greater
than about 30.
The standard error (the standard deviation of the distribution of the sample mean for repeated
samples) is the population standard deviation divided by the square root of the sample size.
Note: this assumes that the population size is much bigger than the sample size, otherwise we
need to multiply by a correction factor = (1- sample size/population size). This is intuitive in
that a random sample (without replacement) the same size as the population must have the
same mean as the population! However, as long as the population size is much larger than the
sample size, it does not have much effect on the accuracy of a random sample.
The distribution of the sample mean is approximately Normal (a known distribution) for
sample sizes greater than about 30, (almost) regardless of the original population
distribution2.
None of these statements require us to know the full distribution of values in the population –
however, the population distribution is important if we deal with small sample sizes (less than
30), when we cannot use some of our approximations without careful checking.
Even with 20 tosses, histogram is bell-shaped; as sample size increases, spread (width)
decreases, centre of the distribution close to 0.5. Now unfair coin (Pr(H)=0.7)
Toss 1 Toss 10 Toss 20 Toss 80
As sample size increases, get symmetrical histogram, decreasing spread, centre close to 0.7.
Sample to population
As the sample mean gets very close to the population mean for large samples, we say that the
sample mean is a good estimator of the population mean.
In all cases, we are using our knowledge about how the population relates to the sample to
make reverse statements about the population from a selected sample.
Sample theory
The 95% Confidence Interval (C.I.) for the population mean using the Normal distribution for
the sample mean is:
2) The 99% C.I. is: Sample mean +/- 2.65 x standard error
We will explain where the 1.96 and 2.65 come from later
To test hypotheses about the population mean, we use the sample mean as an estimate of
the population mean, where we know how reliable an estimate we have by looking at the
standard error (small standard error means little spread in our estimates, i.e. good estimate).
If our sample is big enough that we can assume the Normal approximation (i.e. sample size
of at least 30), then we can find the chance of observing this data or more extreme data, if
the null hypothesis is true. For example, to test whether the population mean has a
particular value against the two-sided alternative (i.e. the population mean does not have
that value), our test statistic, assuming a sample size of at least 30 is:
which we call the z-statistic and compare against the standard Normal distribution (mean=0,
standard deviation=1), i.e. we reject at significance level 5% if the value is less than -1.96 or
greater than +1.96, we reject at level 1% if the value is less than -2.65 or greater than +2.65. In
other words, for a two-tailed test, we reject if the confidence interval does NOT contain the
hypothesized population mean.
The complication is that we need to know the standard error, which is equal to the population
standard deviation divided by the square root of the sample size. This means we need to know
the population standard deviation. In practice, we often do not know this, so we have to
estimate this from the sample, using the sample standard deviation. In order for this estimate
of the population standard deviation to be good, the sample size needs to be at least 30;
otherwise we need to account for this using the Student’s T-distribution instead of the Normal
distribution (we’ll look at how this is different later)
For example, if we toss a coin and the thing we are measuring is the proportion of heads,
The standard error is the population standard deviation divided by the square root of the
sample size.
In the special case of a proportion, the population standard deviation has a simple formula:
As p x (1-p) is always less or equal to 1/4, then the standard error will always be less than
1/√(4 x sample size).
In fact, as we do not know p, we often use this simple limit for the standard error when
calculating CIs for sample proportions, except when the sample size is large, when we can
substitute the sample proportion in the formula. For hypothesis testing, we usually will know
p, so we can use the exact formula. For small samples or p very close to 0 or 1, we should use
Binomial tables, otherwise, the Normal approximation is easier to use and good enough as an
approximation.
Making mistakes/errors
When doing a hypothesis test, we can make 2 types of mistake (error). We can reject the null,
when it is true (called Type I error). We know that the chance of this type of error depends on
the significance level we use to reject the null hypothesis. If we reject at p=0.01, this means
that we have a 1% chance of rejecting a null hypothesis that is really true.
The other type of error is to fail to reject the null hypothesis when it is false (called
Type II error). This is harder to work out, because it depends on the alternative, which
often is not specific.
Example: if we have an unfair coin, this means that the chance of heads, p is not 1/2, but does
not tell us the value of p
How easy it is to detect that it is not fair depends on 2 things:
Consider the most extreme case of a 2-headed coin: the sample proportion is 1, so we need
only 8 tosses to reject fair coin hypothesis at p= 0.01 using a 2-tailed test (2/256=0.0078<0.01)
However, if the population proportion is closer to 1/2, then we need more tosses (bigger
sample size).
Example: This is a simulation that illustrates the effect of changing how fair the coin is and
how big the sample is. We will look at the results for sample size from 1 to 200 and with
Pr(H)=0.1,0.2..0.5 and show the sample proportion for each sample and the 99% C.I. if the
coin is fair, so we can see if we would reject the null hypothesis (fair) for each sample. This
means that if the marker is outside the confidence interval, we would reject the null hypothesis
of a fair coin. The graph shows that it is easier to detect an unfair coin if the sample size is
larger or the coin is more unfair.
For Pr(H)=0.1 (red circles), a sample of 15 is plenty, for Pr(H)=0.2 (green +), need a sample of
about 30, for Pr(H)=0.3 (blue diamond), need a sample of about 100, while for Pr(H)=0.4
(brown x), even 200 is barely sufficient. Note that for the fair coin (green triangle), there are 2
or 3 cases where we would reject the null hypothesis, which is reasonable as we have a chance
of
0.01 of making this mistake and 200 experiments, so expect about 2 mistakes on average.
The related question is how likely are we to reject the null when it is
false? This is called the power of our test and is 1 - the chance of a Type
II error.
This depends on what the true population mean is, but doing this calculation for an assumed
true population mean allows us to check whether our sample is likely to be useful or not. If our
sample size will not allow us to reject the null, even when the true value is quite different, then
our sample is of little use.
If you are seeking funding for a piece of research where the cost of doing the research is
high, it is likely that the funding agency would expect you to show that your sample size is
such that the chance of being able to reject the null is reasonable (usually require at least
80%) given the likely value of the population mean and variance (based on a pilot study or
a literature review).
Remember our coin example – we can choose a sample size that is likely to reject if the
coin has at least a given bias.
For power calculations, use tables or specialized software (webpages for simple cases can
be found on the Internet): free software called GPower for example.
More precise statistical formulae
In general, we may not know the population variance, so we may not be able to do
confidence intervals or hypothesis tests directly. However, for large sample sizes, we can
just use the sample variance as a replacement, while for smaller sample sizes, we need to
also replace the normal distribution (bell-shape) coefficients with ones using the Student’s T
distribution. That allows for the fact that we are using the sample variance instead of the
population variance, so we cannot make such precise statements about the mean (i.e. the
confidence interval will be wider).
For large samples (n>30), the 95% Confidence Interval for the population mean is:
and we substitute the sample standard deviation for the population standard deviation, if
necessary.
1.96 is called the Z-value. To be more precise it is Z0.025 because it is the value such that
there is a 2.5%=0.025 chance of a bigger value (the other 2.5% is the chance of a value
below -Z). For a 99% C.I., we replace it by 2.65 which is Z0.005 (we can get these values
from statistical tables or calculator). Z is what we call a quantile for the standardized
Normal distribution, where standardized means a Normal distribution with zero mean and
unit variance.
The way the formula above works is that the sample mean approximately follows a Normal
distribution with mean equal to the population mean and variance equal to the population
variance divided by the sample size.
Equivalently,
Hence, we get our confidence interval formula – the coverage of the confidence interval in this
case is 95%.
For hypothesis testing, we are essentially doing the opposite. Instead of asking what are the
likely values for the population mean, we ask whether the hypothesised value for the
population mean is reasonable, or in other words, does it lie inside a confidence interval. If it
does not lie in the 95% C.I., we say that we reject the null hypothesis at significance level of
5% (i.e. 100%-95%).
An equivalent way of expressing the hypothesis test is that we reject the null at 5% if:
In this version, we assume that values on either side of the hypothesised value are
equally strong evidence against the null hypothesis. This is called a two-sided test.
If we were sure that only values on one side of the hypothesised value were evidence against
the null hypothesis, our confidence interval and hypothesis test would be one-sided.
For example, if we believed that only population means greater than our hypothesised value
were plausible, our 95% C.I. would cover all values up to the sample mean + 1.645 x standard
error, where 1.645=Z0.05
Again, if this interval does not include the hypothesised value for the population mean,
we would reject at 5% using a one-sided test. Note that it is easier to reject the null
hypothesis
with a one-sided test, in that the upper limit of the interval is lower. However, the decision
to choose one tail should be a priori (beforehand), not after looking at the results!
follows a standard t distribution with (n-1) degrees of freedom where n is the sample size
So, instead of +/-1.96 as interval, we will get a slightly larger number and hence a wider
interval than when using the Z-value, unless the sample size is large.
Note that we can apply all this theory to proportions in just the same way, using the sample
proportion as our estimate for the population proportion and using either the exact formula
for the population variance (=p x (1-p)) or the conservative limit (=1/4). For small samples,
we can use the exact tables for the Binomial distribution to find the confidence interval and
do the hypothesis test, instead of using the Normal approximation.
Example: We have a sample of the heights of 63 12-year old children and show the 95%
C.I. and the test all 3 alternatives against a null hypothesis that the mean population height
is 60
The problem of multiple tests
Note that if you are doing many significance tests, which are independent, the chance of
making at least one false rejection of a null increases proportional to the number of tests
you do (known as p-hacking if you fail to adjust). This can be adjusted for using the
Bonferroni correction - the effective significance level is the stated significance level
multiplied by the number of hypotheses tested. However, it is then very hard to detect
effects when the number of tests (m) is large (the power is low, meaning we are unlikely to
detect that the null hypothesis is false). However, we can instead control the false
discovery rate (proportion of discoveries which are false, where discovery means the null
hypothesis is rejected) – this assumes that we care about the proportion of false rejections,
rather than the number of false rejections.
Benjamini–Hochberg procedure
The Benjamini–Hochberg procedure (BH step-up procedure) controls the False Discovery Rate
at level alpha. It works as follows:
1. We have m null hypotheses tested and we list the m p-values in ascending order and
denote them by P(1)-P(m).
2. For a given alpha, find the largest k such that P(k) is less than k x alpha/m
3. Reject the null hypothesis (i.e., declare discoveries) for i=1 to k
Geometrically, this corresponds to plotting P (k) vs. k/m (on the y and x axes respectively),
drawing the line through the origin with slope alpha, and declaring discoveries for all points on
the left up to and including the last point that is below the line. This procedure is valid when
the m tests are independent, and also in various scenarios of dependence, but is not
universally valid.
We examine a simple example below with 9 tests and alpha =0.05. Bonferroni would suggest
the unadjusted p-value should be 0.0055, while BH looks at the line with slope 0.05 and
rejecting all tests until there is one above the line (i.e. reject the first 3).
Extension of hypothesis testing and confidence intervals to
other situations
These ideas can be extended to other situations. The most common is testing for change in
population mean. In other words, is the mean of population 1 the same as the mean of
population 2 (i.e. the null hypothesis is that the difference between the means is zero)?
It is important to distinguish between two situations: namely whether the two samples
contain the same individuals or not. If they do contain the same individuals, then we can
analyse the change in measurements on the same individuals and get much more sensitive
results. For example, to assess the effect of training, testing the same individuals before and
after is a much more sensitive way of finding a change than testing a different set of
individuals. This is because we are able to exclude the variability across individuals and
focus on variability within individuals.
Paired T-test
Consider 64 individuals with average score of 50 and standard deviation of 8.
If we improve the score of half the individuals by 1 and half by 2, then we now have a new
average score of 51.5. If we match individuals and look at the change, we have 64 differences
with mean 1.5 and standard deviation 0.5, so the 99% confidence interval for the mean change
is (1.33,1.67), so we are quite certain that there is a positive change (even if it is not big enough
to be useful!).
In other words, by looking at the differences for each individual, we are just treating the
change as our variable, and then applying the same statistical procedures to the changes to get
confidence intervals and hypothesis tests for the difference in the population means. We call
this a paired t-test (t-test, because we almost always do not know the population standard
deviation for change and need to estimate it from the sample).
Example: we have 24 measurements of the temperature inside and outside a box. We create a
new variable, diff, which is the difference between inside and outside temperature. We show
the 95% C.I. and the test of the (obvious) null hypothesis of no difference, i.e. the mean
difference is zero.
Two-sample T-test
If the populations are different, or we do not use the same sample, what then?
Not surprisingly, the difference in sample means is still the best estimator of the difference
in population means, while the standard error of the difference is √(population variance
1/sample size 1+population variance 2/sample size2), so we can use similar ideas for
producing confidence intervals and hypothesis tests for the difference in the population
means. This is called the two-sample t-test because the population standard deviations are
not usually known and need to be estimated from the samples
For a large sample, the 95% confidence interval for the difference in means is:
(Sample mean1-sample mean2)+/-1.96 x standard error
Hence, for our example above, if we do not match individuals (treat them as a new sample),
then the mean difference is 1, but the standard error is
so our 95% C.I. is 1.5 +/- 1.96 x √2 and 99% C.I. is now 1.5 +/- 2.65 x √2, which includes 0,
so we fail to reject the null (i.e. no change).
Example: we look again at the heights of the 63 children, this time we examine whether the
boys and girls have different heights. The results are shown twice for the difference in heights,
with and without the assumption that the variance is the same in the two populations.
Effect size
The (standardized) effect size is a standardized measure of the strength of a relationship. In
this case, it is just the standardized mean difference between the two groups. This is
important if we are trying to calculate the power for a two-sample case with different
sample sizes for an assumed effect size.
The Goodness of Fit test can be applied to all situations where we have count data and the
alternative hypothesis is the 2-tailed form (e.g. for 3 categories, the null hypothesis is that the
population proportions are specified (for example, that they are equal) and the alternative is
that at least one of the proportions is something different).
Example: if we have data for how many people in our sample live in HK Island, Kowloon
or NT, we can test the hypothesis that the proportions are consistent with the proportions in
the 2011 census data.
The usual test statistic is called the (Pearson’s) Chi-squared (or X 2) Goodness of Fit statistic.
Where Observed means the actual count for that category and Expected means the count we
would expect for that category if the null hypothesis is true. X2 sums over all categories.
You can see that this measure is a way of summarizing how close the observed data is to
the data that is most likely if the null hypothesis is true, hence the name goodness of fit.
If k is the number of categories, then we compare X 2 against tables for the Chi-squared
distribution with k-1 degrees of freedom (This is the distribution for the sum of k-1
independent squared standard Normal variables)
The calculation of X2 and of the probability of observing this value of X2 or a value more
extreme (observed significance level) can be easily found using a computer package.
Strictly speaking, this assumes the Normal distribution, so the probability distribution for X2
is only approximate, but this is a good approximation as long as the expected count is at least 5
for each category.
There is another statistic used in this situation, called the G2 Likelihood ratio statistic,
which is essentially just another approximation to a chi-squared statistic.
Note: we are assuming that the data is nominal scale, not ordinal. If the data is ordinal
scale, then there are more sensitive statistical methods that should be used.
Example: We have simulated data for the mother tongue of students in an international school:
English, Cantonese or Putonghua with Pr(E)=1/3, Pr(C)=1/2,Pr(P)=1/6. The first simulation of
1000 experiments uses a sample size of 30, the second uses a sample size of 60. We test a null
hypothesis that the proportions are equal. If the counts are E=18,C=31,P=11 for a sample size
of 60, then under the null, the expected values are all 20, so X 2 = ((18-20)2+(31-20)2+(11-
20)2)/20 = (4+121+81)/20=10.3.
Sample of 30 Sample of 60
If we use statistical tables for Chi-squared with 2 degrees of freedom, Pr(X 2>5.99) = 0.05, so
the power for a sample of 30 is about 50%, sample of 60 is about 85% (based on the proportion
of values greater than 5.99).
10: Relationships between pairs of variables
If we do not know what type of relationship there is between a pair of variables, the best
starting point is usually a graphical display or a table. For variables that are categorical,
we usually use a table of counts (can use mosaic plot); otherwise we use a graphical
display called a scatterplot.
Let’s look at an example with two nominal scale variables and then another example with two
interval scale variables.
Car poll: This shows that there is association between preferred car type and marital status
and the mosaic plot shows the key difference is Singles prefer a Sporty car to a Family car.
Car poll: This shows that there is association between preferred car type and marital status
and the mosaic plot shows the key difference is Singles prefer a Sporty car to a Family car.
Testing for independence of categorical variables
A common situation with categorical data is that we have 2 nominal scale variables with a
research hypothesis of association between the variables while the null hypothesis is
independence (which means that probabilities multiply).
In the example above, the question would be whether we can reject the hypothesis that
the preferred car type and marital status are independent in our population. The
alternative hypothesis is that there is some (unstated) association between them.
It turns out that we can use the same concept (GoF) and formula as we used for 1 variable with
multiple categories above.
In the case of our example, the null hypothesis implies that the proportion of married people
who prefer family cars should be the same as the proportion of singles who prefer family cars
or equivalently, the proportion of those preferring family cars who are single should be the
same as the proportion of those preferring sporty cars who are single. In other words the
probability of preferring a type of car is independent of the probability of being married.
Given that independent probabilities multiply, this means that the expected number of married
people preferring a family car is the total sample size times the sample proportion of married
people overall times the sample proportion of those preferring family cars overall; similarly
for the other five combinations of marital status and car preference.
We can then use these expected and observed counts in the X 2 statistic and the degrees of
freedom will be (k-1) x (m-1) where k is the number of categories for one variable and m is the
number of categories for the other variable, so in this 3 x 2 case, the degrees of freedom are 2.
The expected count for the first cell (i.e. married man prefers family car) is: 303 x (196/303) x
(155/303) = 100.3 so the X2 contribution for the first cell is (119-100.3)2/100.3 = 3.5. Similarly
for the other 5 cells. Note: Need expected count of at least 5 in each cell.
Note: there are more powerful methods if the variables are not both nominal scale, such as
Spearman rank correlation if both are ordinal scale and Kruskal-Wallis test if one is ordinal
scale and one is nominal scale.
Growth of babies: This plot shows a strong non-linear pattern to the relationship
between growth ratio and age.
Once we log transform both growth ratio and age, we see an approximate linear relationship
Those examples show that the relationships can have many forms, but they often are
approximately linear (possibly after simple transformation, e.g. taking logs), which allows
us to build simple models.
Use of correlation
We can summarize the linear relationship using correlation (r). Correlation
measures association in terms of how strong the linear relationship is, where a
correlation of:
r by definition is:
Covariance(X,Y)/√(Var(X)xVar(Y))
Where:
Covariance(X,Y)=E((X-Mean(X))(Y-Mean(Y)); Var(X)=E((X-Mean(X))2)
i.e. correlation is the joint variability of X and Y scaled by the variability of each of X and Y.
r=∑((Y-Mean(Y))(X-Mean(X))/√(∑ (X-Mean(X))2x∑(Y-Mean(Y))2)
It is straightforward then to show that if Y=X, r=1 and if Y=-X, then r=-1
Obvious hypothesis test is if the population correlation is zero, but this test requires us
to assume that both variables (X and Y) follow a Normal distribution.
The test relies on the fact that if X and Y follow a Normal distribution with zero correlation,
then:
r/√((1-r2)/(n-2)) follows a Student’s t distribution with (n-2) degrees of freedom. If the sample
size is at least 30, the t distribution is very close to a Standard Normal distribution
(otherwise the tails of the distribution are “thicker” to account for our estimating the
population variance using the sample variance)
If X and Y both follow a Normal distribution, then it turns out that independence is
identical to zero correlation, but this is not necessarily true in other situations (zero
correlation just means no linear relationship, remember, while independence means no
relationship at all, linear or non-linear).
However, note that r2 does not tell us about the direction of causation. It is possible
that X => Y
Y => X
There is a whole field of statistical methodology that looks at correlations for large sets of
variables (factor analysis or principal component analysis), but the mainstream approach
is to build models for linear relationships.
Simple (bivariate) linear model
We write the simple linear model as:
Y=A + B X + ε
where ε is the random error, A is the intercept, B is the slope, X is the independent variable, Y
is the dependent variable and we assume that random errors from different observations are
independent and have constant variance.
This is like elementary algebra for a line, except for the random error
There is a very general mechanism for fitting statistical models with constant error
variability. This method is called “least squares”. It minimizes the squared error, or in other
words, it finds estimates for A and B (we will call them a and b) that minimize the squared
error for Y:
This can also be written as ∑ri2, where ri is called the residual for the ith data point, being
the difference between the observed and fitted y for that point.
Hence the name least squares. It is possible to prove that this is the best way to fit a
linear model if the error variance is constant.
Also, the error variance can be estimated from the sample as:
s2 =∑ri2/(n-2)
(we divide by n-2 because we “use up” 2 observations to account for the 2 parameters we
estimate – remember that we can always fit a line perfectly to any 2 points)
b=∑(Y-Mean(Y))(X-Mean(X))/ ∑(X-Mean(X))(X-Mean(X))
=Cov(X,Y)/Var(X)
a=Mean(Y)-b x Mean(X)
For those of you with good mathematical skills, it is easy to show that these estimates
are optimal by using calculus to find the values of a and b that minimize s2
Numbers (such as a,b and s 2) alone (without graphics) may not be a good summary of what is
going on.
If we can also assume that the errors follow a Normal (bell-shaped distribution) (in addition to
our assumption of independent errors with constant (unknown) variability), then we can
develop statistical inference for the slope. The key hypothesis to test is whether B=0 as this
simplifies our model to mean that X and Y are independent. This turns out to be the same test
as testing r=0, even though here we are only assuming that the errors for Y given X are
Normal, not that X and Y are both Normal.
Essentially, we are asking what values of b we might expect if there is really no linear
relationship between X and Y, and the errors follow a Normal distribution with constant
variance, and rejecting B=0 if the value we observe for b is too far away from 0 to be likely
under the null.
This, of course, also provides the basis for finding a 100(1-α)% confidence interval for B of
b+/-sb x tα/2(n-2), where tα/2(n-2) is the cutoff value of a t distribution with (n-2) degrees of
freedom
If the sample size is above 30, then the t distribution is very close to a Standard Normal
distribution.
(We could also test separately for A=0 in a similar way, which would mean we are testing if
Y is proportional to X, which is rarely meaningful).
Residuals
As with any tool, diagnostics are important. The simplest diagnostics for regression are the
residuals.
2 key ideas: look for particular data points that look to fit badly (possible error?) and for
patterns that suggest model errors
1-∑ri2/∑(y-Mean(y))2
This shows clearly that a perfect fit will give r2=1 and if the fitted line is no better than
a straight horizontal line through the Mean of y (i.e. yfit=Mean(y)), then r2=0
Prediction
How can we predict y for a value of x?
We can just substitute into our equation, using our estimates, a and b:
Y predicted=a + b × X predictor
We can also estimate the prediction variability, given that the distribution of Y predicted,
given an X predictor is:
(Y predicted-true Y)/spred follows a t distribution with n-2 degrees of freedom, where
This can be used for hypothesis testing or constructing confidence intervals for the mean of Y
given X.
Note that spred is smallest for Xpred=Mean(X) and increases as Xpred moves away from
Mean(X)
(look at example of confidence intervals in JMP – show mean confidence interval versus
individual confidence interval, which is the prediction for a new individual measured)
However, note that this assumes that our linear model is correct for that value of x. This
may be safe if x is within the range of data we observed, but what if x is much greater (or
smaller) than our dataset? In that case, we have no real support from the data for believing
that we can extrapolate like this. It should be obvious that a wider range of x values will
produce more precise results both for estimating B and for prediction.
11: Multiple Regression
All the ideas we have used can be easily extended to handle more than 1 independent
where B0 is now the intercept (previously=A for simple regression) and B1 to B(p-1) are
the slopes (previously=B)
The advantage of writing it this way is that we can also write this in matrix and vector
notation as:
Y=X B + E
One reason for writing it in matrix notation is that we can show via calculus that the
least squares solution has a very simple form in matrix notation:
Yfitted = Xb = HY
where H=X(XTX)-1XT is a projection matrix (called the Hat matrix) that “projects” from n
dimensional space onto p dimensional space to obtain the closest fit to Y (in terms of squared
errors), where n is the sample size and p is the number of parameters in the model.
For example, the null model which has p=1 (i.e. Yfitted =Mean(Y)) has H that projects Y onto
the Mean(Y)
The simple linear model with p=2 has H that projects Y onto a 2 dimensional space, i.e. the
plane that goes through Mean(Y) and X, because we already saw that the fitted line goes
through Mean(Y) and has equation =a+bX
For projection, you can consider the analogy of a video projector, which casts a 2D shadow
(projection) on the screen of a 3D object (e.g. your hand). This is a projection from 3D onto 2D
(like a data set of 3 points when we fit a simple regression line with just one X variable)
We can show using calculus that this is the optimal solution under quite general
conditions. You do not need to know this formula, it is just provided to assist
understanding of what we are doing when fitting a multiple regression model using least
squares.
The divisor is n-p to account for the p parameters that we have fitted – we can fit p points
exactly (compare to 2 points fit exactly for the simple regression)
For testing Bi=0 or finding a confidence interval for B i, the idea is similar to simple linear
regression:
And reject Bi=0 at 5% (two-tailed test) if bi is > t.025 SEBi or bi<- t.025 SEBi, where
SEBi=√(XTX)- 1ii s
This example is about predicting the auction price of grandfather clocks, given the age
and numbers of bidders – clearly both independent variables are important.
This example is predicting the oxygen update for runners, given their age, weight, run time,
resting pulse, running pulse and maximum pulse. Clearly run time is very important and age
and running pulse matter.
This leads to the obvious question: how to select the optimal set of independent variables in a
regression?
We need to think carefully first - variables may be included in a model for 2 reasons:
a) theoretically they are important or statistical controls
b) they can be shown statistically to be important (i.e. B=0 is rejected)
Theoretically important variables that we need to statistically control for (by including them in
the model) are called confounders. In this case, we need to include variables even if they are
not statistically significant.
Ideally, we would randomize our experiment and avoid the problem of confounders, but this
is often not possible unless we have a designed experiment (we often have observational data
or retrospective data).
The variability is measured in terms of sums of squared deviations (SS is the abbreviation
for Sums of Squares, which means the sums of squared deviations). We try to partition
(split up) the variability into different components due to different variables.
The total variability is essentially just the variance for Y times (n-1): Total SS=∑(Y-
Mean(Y))2 We then divide this up - what is left after fitting the model is called the residual:
Residual SS=∑(Y-Yfitted)2=∑ri2, (i.e. the remaining part that is assumed to be random
The difference (meaning the part “explained” by our model) is: Model SS= ∑(Y-Mean(Y))2-∑(Y-
Yfitted)2
For each SS, we also have the degrees of freedom (d.o.f.), which is the dimension of the space
spanned by this part, so d.o.f. for the Total SS is (n-1) as we have n data points, but have
subtracted off the mean.
For the simple linear model, the Model d.o.f. is 1 (1 independent variable) and the Residual
d.o.f. is the difference, i.e. n-2
In the multiple regression model the Model d.o.f. is p-1 (the number of independent
variables) and the Residual d.o.f. is the difference, i.e. n-p
We create an intermediate stage called the Mean Square (MS), which is the SS divided by
the d.o.f.
So this explains why our test statistic is: F=Model MS/Residual MS,
which is called the F statistic, with Model d.o.f. and Residual d.o.f. degrees of freedom and we
reject the null hypothesis if F is large enough.
We can easily look up the distribution of F under the null hypothesis (i.e. that all the B i=0)
For the simple linear model, the F statistic is mathematically equivalent to the t statistic (it
is actually the square of the t statistic) for the slope of X, so it yields the same p value as
the two- tailed t-test in that case.
For the multiple regression model, the F statistic is the overall test that combines all the
individual t tests into one omnibus test. If we cannot reject the omnibus test, we do not
need to spend time trying to identify the best model as it means that all the variables
combined do not provide a better fit than just using the mean for Y.
We can also try to divide up the Model SS amongst the variables, but this is tricky, unless the
Xs are uncorrelated (possible in designed experiments), as otherwise, the SS for each variable
depends on the order we add it in.
We can also expand the idea of r2 to what is sometimes often called R2 (with a capital R) to
distinguish that it is the proportion of variation “explained” by a set of variables (versus
just one variable for r2).
At first glance, it looks sensible to choose the model with the largest R 2. However, this is
flawed, because each time you add a new variable, R 2 will increase, even if the new variable is
unrelated to Y. This means R2 is only useful in comparing models with the same number of
variables. For comparing models with different numbers of variables, we should use what is
called adjusted R2
This adjusts for the bias in R2 towards models with more variables – maximizing Adjusted R 2
is equivalent to minimizing s2.
In fact, adjusted R2 is close to R2 if p is small compared to n, so (n-p)/(n-1) is close to 1, i.e.
if the number of parameters is small compared to the sample size.
So, how do we find and select the best model (i.e. best combination of the independent (x)
variables)?
It used to be popular to do stepwise regression. The idea was that you started with a null
model, then added in the one variable that yielded the highest t statistic, then choose the
variable that gave the highest t statistic when added to the first variable and so on, until
there were no more variables to add that have a significant slope. This is called forwards
stepwise regression.
The problem is that this process has 2 flaws:
1) it tends to yield models that are too large because you are doing many tests, and
choosing the most significant from a range of tests. This means that the true
significance level of the model is much higher than claimed, based on the
significance of each test (i.e. the risk of falsely rejecting the null is much higher
than claimed)
2) it also may exclude some good models because some variables may only be good
if included together with another variable, but we are only adding one at a time.
Another approach is to include all the possible variables to start, check if that model is
significant, and then drop the least significant variables one at a time, until they are all
significant (the t statistic checks for significance conditional on all the other variables already
in the model). This is called backwards stepwise regression. This approach has similar
problems, plus it may also have an additional problem that we call collinearity.
Collinearity means that different sets of independent variables provide almost the same
information. Collinear means - in a line. It means we can predict at least one of the
independent variables well from the others. This is quite common if there are a large number
of variables. In the case of collinearity, there are many possible coefficients for the Xs that are
very close to the best. This is often a problem with observational data. For an experiment, we
can ensure that the x values are chosen such that this will not happen (we will return to this
when we consider experimental designs).
Collinearity is a problem in two senses:
We look at a famous dataset due to Longley – we first look at predicting y individually from
X1 to X6
However, when we try to predict from X1-X6 together, we find something strange – some of
the coefficients that are statistically significant change sign and some variables (X1,X2,X5) are
not useful.
Why does this happen? We need to look at the correlation between the X’s
As can be seen, X1,X2,X5 and X6 are all very strongly positively correlated – this is collinearity
– it means that these 4 variables have a lot of common information about Y and it turns out
that once we know X6 (which is the year), the other 3 variables do not add anything useful
about Y.
Another alternative is mixed stepwise, which combines backwards and forwards steps.
One of the reasons stepwise methods were popular was that the computer power available was
small, which used to be an issue for large models. Nowadays, that is usually not a problem, so
we may use what is called best subsets regression. This means finding the best models for
each value of the dimension of our model, i.e. for any given number of variables. We can then
review the models to see which ones make sense in terms of our theory.
To decide which model to choose, there are many different approaches. One approach is to
start by comparing the simplest and the most complex model.
If this is not significant, then stop, as that suggests that everything we observe, other than
the variability linked to variables in the simplest model, is just random, so we should just
use that simplest model, which is usually the null model with no independent variables.
If it is significant, then add variables in to the simplest model and choose the simplest model
such that the F statistic comparing that model against the most complex model is not
significant. This is called the lattice approach (construct a tree of all possible models).
Arguably, this applies the philosophy of Occam’s razor. This philosophical principle (prefer
simple explanations when explanations predict equally well) says that you should choose the
simplest explanation that is adequate.
Mixed stepwise regression is still a useful tool, as long as we are aware of the risks,
particularly of wrongly rejecting the null, so we should reduce the apparent significance level if
we have a large number of tests.
Minimizing AIC or BIC are more sophisticated criteria, which are more trustworthy than
repeated significance tests (the detailed statistical argument is too complex to explain here)
which means choosing s and p to be small together. Note that these criteria need
adjustment if we consider models with different transformations, while adjusted R2 can
still be used.
Stepwise regression does not always yield sensible answers using the default settings. We
show here the results when we try to predict the oxygen uptake in the previous example
from 30 randomly generated variables (so they have NO relationship with the dependent
variable). This is an extreme example as the number of variables is almost as large as the
sample size.
This happens because the default p-value for adding/dropping a variable is very high (0.25/0.1).
If we reduce this down to 0.01/0.01 we get a null model as the most significant variable has a
p-value of about 0.03, which is about 1/30 where 30 is the number of variables, so this is
what we would expect by chance if no variables are associated with oxygen update.
Diagnostics
The 2 core diagnostic tools are the residuals and influence or leverage. The residuals tell us
what the fitting error is for each point. Influence or leverage values tell us how much the
model depends on each point. They actually depend on the H hat matrix we mentioned before.
If a diagonal coefficient of H is large, then this data point has a large influence on the fitted
model. This is because the diagonal element of H is the coefficient of the value of y i in
estimating the predictor for yi:
The residuals can be used in different ways. They can be plotted in sequence to see if there is a
pattern in the sequence. They can be plotted against the fitted values to see if there is some
non-linearity in the overall responses or they can be plotted against each independent variable
to see if there is non-linearity for each variable. We can also look at the distribution of the
residuals to see if they look like they follow a Normal distribution (bell-shaped curve) (see
later).
Leverage plots show the influence of data points on the model, either overall, or in
terms of each independent variable.
Polynomial Regression
A special type of multiple regression is polynomial regression. In this case, the question is
what order of polynomial in X do we use to fit the model. Only in special situations would
we use high order polynomials, because the models usually give very extreme predictions
outside (or even inside) the data range. This is one reason why we would often prefer
looking at transformations instead, particularly if we can interpret the transformation
sensibly. The growth example below shows that a fifth order polynomial is statistically
significant but is very hard to interpret and yields an adjusted R2 value that is not that
much better than the simple log-log transformation (or as we will see, rank transformation).
Transformations
The most common transformation is the power transform
z=log(y)
log makes special sense when dealing with product or ratio variables, because it turns
them into linear relationships:
We already looked at this for the growth example, which explicitly involves ratios.
Note that transformations may simplify the model, e.g. if weight is proportional to some
power of height, then log(weight) is linear in log(height). It is important to consider models
that ‘make sense’, which often means they have a sound theoretical basis. Another nice
example is predicting the volume of trees given the diameter and height. We compare
regression models using the raw data and log transformed. The log model not only fits
better (compare R2), but also has a very simple interpretation: V is proportional to D2H,
which makes sense if we think of the formula for volume of a cone or cylinder.
The other very useful transformation is ranking, i.e., we replace the data by their ranks 1 to
n. Replacing the data by their ranks is usually equivalent to what we call a non-parametric
procedure. This is usually a less powerful procedure, but it is based on weaker
assumptions, so it is more robust. The tricky part is how to handle ties (values with the
same rank). “Proper” non-parametric procedures account for the fact that ties reduce the
variability and reflect less information.
We can visit the growth example using ranks. The simple linear model using ranks means
that the ratio increases (but perhaps at a non-constant rate) with age.
Normality Tests
The statistical inference we do for linear models assumes that the errors independently follow
a Normal distribution. Note: unlike the distribution of the sample mean when testing the
population mean, a large sample does NOT make this assumption correct!
There are several different formal tests of Normality, but one useful graphical method
plots residuals in order of increasing value against Normal quantiles (i.e. plot residuals
against their expected values if they come from a Normal distribution). If the
distribution of the residuals is Normal, the plot should be roughly linear – JMP provides
confidence bounds to allow an informal test of Normality. The growth example shows a
failure below.
12: Analysis of Variance for categories (Factors or Groups)
What if we have more than two populations, how can we test if they have the same
population mean? Analysis of Variance is also a way to simultaneously test for any
difference in the means of a set of populations, assuming that they have the same
population variance. There is no simple Confidence Interval idea any more, but we can
still do a significance test. This time, the thing we use to test for a difference (the test
statistic) is a bit more complex. The idea is as follows:
If the populations have the same mean, the sample means should be similar and hence the
sample means should be close to the overall mean. We look at the ratio of the squared
deviations of the sample means from the overall mean to the deviations of the individual data
points from their respective sample means. If the ratio is large, then we do not believe that the
sample means are similar and hence that the population means can be the same.
This should sound familiar. The idea is exactly the same ANOVA idea we used for multiple
regression. For a one-way classification (one-way ANOVA), the table looks like:
Source of Variation SS d.o.f. Mean Square F Ratio
Between Groups SSB k-1 MSB=SSB/(k-1) F=MSB/MSW
Within Groups SSw n-k MSw=SSW/(n-k)
Total SST n-1
where F=MSB/MSW and k is the number of groups. The Within Groups part is the same idea as
the Residual for the multiple regression model, i.e. the part that our model does not “explain”
The precise formulae are not important, but the relative simplicity is attractive and has lead
to the extension of this idea to much more complex situations, in particular where there are
multiple simultaneous groupings.
For example, if there are 2 classifications, the sources become Factor 1, Factor 2, Residual and
Total. This makes it relatively easy to identify the important sources (causes) of variability. If
the groups are very different, then SS B will be large. If the difference is large relative to the
number of groups, MSB will be large. If the variability left is small, then SS W will be small. If
SSW is small given the sample size, then MSW will be small.
Overall then, we can see why large values of F indicate that there is evidence of group
differences, i.e. we reject the null hypothesis of no group differences.
Statistical software tells us whether the F value is big enough to reject the null hypothesis.
Usually, it tells us the significance level at which we should reject the null - this is called the
observed significance level. If there are only 2 groups, we can use the T-test, which is
equivalent to F-test in this case (k=2, means k-1=1) with the F statistic equal to the square
of the t statistic.
Note that R2 is scale independent, which allows easy comparisons. Again, there is a
problem with R2 in that adding an extra group will always increase it. As in the multiple
regression case, we use adjusted R2, allowing comparison of models with different
numbers of groups.
General strategy is to find model that is simple (few terms), good explanation (high R 2) and
easy to interpret (theoretical basis and/or consistent inclusion pattern of variables and
interactions). The example below illustrates a simple randomized controlled trial where f is a
placebo and d,f are the active drugs.
Multiple Comparisons
What if we want to compare the individual groups? For example, we might want to compare
pairs of groups and see if they are significantly different. The simple approach is just to use a
t-test for those two groups, but the problem is like the one for multiple regression, where there
are many possible tests. One approach is to only look at pairs if the overall F test is
significant. This approach is quite simple, particularly if the numbers in each group are the
same (we call this a balanced design). The other better, but more complex, approach is to use a
modification of the t test that takes into account the multiple comparisons being done (e.g. the
Tukey-Kramer Honestly Significant Difference (HSD)). The drug example illustrates that the
result is less significant once we take the multiple comparisons into account:
Contrasts
Contrasts are linear combinations of group effects – the simplest example would be the
difference between any pair of groups. A more complex example is where we have 2
treatments and 1 control group and we want to test whether the treatment groups are different
from the control group. In other words, this is a situation where we are testing for zero
differences in specific known combinations of the groups. The drug example below tests
whether the average effect of a and d is better than placebo, even if we are unclear as to which
drug is better.
Nested vs. Crossed
What if there is more than one classification, e.g. Sex and Marital Status? In general, if there
is more than one classification, it can work two ways: nested (hierarchical) and crossed.
Nested means that the second classification works inside (nested) the first classification,
whereas crossed means that both classifications are generally applicable. Sex and Marital
Status will usually be crossed, whereas for School and Class Number, Class Number usually is
a within school classification, i.e. it is nested inside School.
For two crossed classifications (2 way ANOVA), we have 4 possible sources of variability:
Classification 1
Classification 2
Class 1 x Class 2
Residual
where Class 1 x Class 2 (sometimes denoted Class 1.Class 2). This effect means that
the Class 2 differences are not the same for different values of Class 1. In our example, it
might mean that the differences between single and married people may be different for
males and females. If this is not the case, we would say that Sex and Marital Status have
independent effects and merge this source back into the Residual.
Lack of Fit & Replicates
For models of this sort, we usually check for lack of fit. This means we hope to not reject this
hypothesis. This analysis is possible if we have what we call replicates. This means that for
some combinations of the factors, we have multiple data points, giving us a way to estimate the
“pure error” (i.e. differences between respondents with identical factor combinations) which
excludes any possibility that our model does not include necessary interaction terms between
the factors. We divide up the residual error into “pure error” and “lack of fit” to check whether
our model looks OK (assuming we have not completely ignored important factors). Replicates
are a very good idea when designing an experiment as they allow a diagnostic check that does
not depend on whether the model is correct.
We look at an example of making popcorn and looking at which factors of popcorn type,
oil amount and batch size and their interactions predict high yield. The results show that
popcorn type, batch size and their interaction seem to matter.
However, it turns out that it is specifically the combination of small batches of gourmet
popcorn that performs better, so we have a good simple explanation:
Nested
For two nested classifications, we have 3 sources:
Classification 1
Classification 2 (inside Classification 1)
Residual
We would normally check whether the Class 2 differences (within Class 1) are important.
In our example, this would mean checking whether classes within schools are different.
The nested situation may be more complex, because Class 2 can be a sample from Class 1.
In this case we say that Class 2 is a random effect, which alters the error term in our F test
(too complex to explain fully here) and makes the correct statistical analysis more
complex, because we need to adjust the denominator in our F tests to account for the
random effects part.
Assumptions
The analysis makes 3 important assumptions:
1) Independent errors for different data points
2) The variability within groups is the same (across different values of X for
regression models)
3) errors follow a Normal (bell-shaped) distribution.
In practice, ANOVA is quite robust (not too sensitive to these assumptions), but sometimes we
can reduce divergence from these assumptions by transforming the data.
Transformations again
For one-way ANOVA, we can remove the assumption of equal variance using Welch’s ANOVA.
If the data look approximately linear, we can use statistical software to fit a range of models
and see how well they fit (again use R2, F test and ANOVA tables).
Random effects
Can integrate multiple regression with ANOVA for factors to get what is called General
Linear Model or Analysis of Covariance (ANACOVA). This allows investigation of which
factors (groups) and covariates (linear relations) are useful in ‘explaining’ the variability in
the data.
For individual covariates, can use T-test, but this is statistically equivalent to F-test. The T-
test is a test of a null hypothesis that the slope coefficient is zero. We can also allow for
interactions between Factors and Covariates (i.e. different slopes for different groups).
Interactions between Covariates are cross products of covariates.
We illustrate GLM using the drug data with a covariate, x which is the baseline measurement
of y (i.e. before taking the drug). This shows that there is no significant drug effect, once we
control for the baseline measurement. There is also no evidence of an interaction between the
drug and the baseline measurement.
14: Model selection criteria
For all linear models, the two common criteria are R2 and F-test.
R2 without adjustment is a very bad criterion because it will always choose the most
complex model. Adjusted R2 is better, but still tends to choose overly complex models.
The F-test is equivalent to the Log Likelihood ratio test for linear models, i.e. looking at
the 2 x the difference in the maximum Log Likelihood of the 2 models being compared.
However, this criterion has also been criticized for choosing models that are too complex.
The main competitors are AIC (Akaike Information Criterion) and BIC (Bayesian
Information Criterion)
The objective is to choose the model with lowest AIC or BIC, which both choose the highest
Log Likelihood after a penalty for the number of parameters (k). AIC is based on information
theory arguments, while BIC relies on Bayesian analysis arguments, although arguably BIC is
inconsistent with Bayesian inference as it takes no account of any prior information.
AIC = 2k -2 Loge(Lik), AIC has a variant called AICc which corrects for finite sample
For linear models with independent Gaussian errors, Loge(Lik) can be written as
–n/2 Loge(RSS) if we ignore constants that cancel for comparisons, so we can write
AIC=2k + n Loge(RSS)
BIC= k log(n) + n Loge(RSS)
BIC penalizes complexity more than AIC in all practical situations, as log(n) is usually much
more than 2.
These criteria are not restricted to linear models. The only real limitation is that we need to
be able to calculate the log likelihood and the number of parameters.
These criteria do not involve a statistical test, but generally differences of less than 2 in AIC or
BIC are considered to be meaningless (i.e. we cannot distinguish between the models). All of
these criteria except (adjusted) R2 assume that we have the same dependent variable values.
AIC and BIC do not require that the models are nested (i.e. all models are special cases of the
most complex model), unlike the F-test.
15: Experimental Design
We are often concerned with the analysis of data generated from an experiment. It is wise
to take time and effort to organize the experiment properly to ensure that the right type of
data, and enough of it, is available to answer the questions of interest as clearly and
efficiently as possible. This process is called experimental design.
The specific questions that the experiment is intended to answer must be clearly identified
before carrying out the experiment. We should also attempt to identify known or expected
sources of variability in the experimental units since one of the main aims of a designed
experiment is to reduce the effect of these sources of variability on the answers to
questions of interest. That is, we design the experiment in order to improve the precision of
our answers.
It is important to understand that optimizing the experiment in some ways may make it less
useful in others (e.g. minimizing the prediction error may not allow us to fit polynomial or
other non-linear models)
Main Effect
This is the simple effect of a factor on a dependent variable. It is the effect of the factor
alone averaged across the levels of other factors.
Interaction
An interaction is the variation among the differences between means for different levels of one
factor over different levels of the other factor.
Example
A cholesterol reduction clinic has two diets and one exercise regime. It was found that exercise
alone was effective, and diet alone was effective in reducing cholesterol levels (main effect of
exercise and main effect of diet). Also, for those patients who didn't exercise, the two diets
worked equally well (main effect of diet); those who followed diet A and exercised got the
benefits of both (main effect of diet A and main effect of exercise). However, it was found that
those patients who followed diet B and exercised got the benefits of both plus a bonus, an
interaction effect (main effect of diet B, main effect of exercise plus an interaction effect).
Blocking
This is the procedure by which experimental units are grouped into homogeneous
(similar) clusters in an attempt to improve the comparison of treatments by randomly
allocating the treatments within each cluster or 'block'.
Randomization
Randomization is the process by which experimental units (the basic objects upon which the
study or experiment is carried out) are allocated to treatments; that is, by a random process
and not by any subjective and hence possibly biased approach. The treatments should be
allocated to units in such a way that each treatment is equally likely to be applied to each unit.
Randomization is generally preferred because the alternatives may lead to biased results. The
logic is similar to why we need random samples in order to have samples that are
representative (unbiased) of the population.
The main point is that randomization tends to produce groups for study that are comparable
in unknown as well as known factors likely to influence the outcome, apart from the actual
treatment under study. This is critical, as in most situations there are unknown factors,
certainly when dealing with social or biological experiments.
The standard analysis of variance F tests assume that the treatments are fixed, but have
been allocated randomly to subjects.
Blinding
In a medical experiment, the comparison of treatments may be distorted if the patient, the
person administering the treatment and those evaluating it know which treatment is being
allocated. It is therefore necessary to ensure that the patient and/or the person
administering
the treatment and/or the trial evaluators are 'blind to' (don't know) which treatment is
allocated to whom.
Sometimes the experimental set-up of a clinical trial is referred to as double blind, that is,
neither the patient nor those treating and evaluating their condition are aware (they are
'blind' as to) which treatment a particular patient is allocated. A double-blind study is the
most scientifically acceptable option.
Placebo
A placebo is an inactive treatment or procedure. It literally means 'I do nothing'. The 'placebo
effect' (usually a positive or beneficial response) is attributable to the patient's expectation that
the treatment will have an effect. This is sometimes linked with the Hawthorne effect, which is
the effect due to attention being paid to people, i.e. people respond positively to being assessed
or measured. Clearly the Hawthorne effect should occur equally for all study participants as
long as the level of interaction is similar.
Ethical concerns
Clearly, it is unethical to randomize the allocation of treatments to humans, if we know that
one treatment is better than another (or than the placebo), so we sometimes have to use the
current best treatment for comparison instead. This can be complex when there are risks
(e.g. side effects) as well as benefits of treatments. When it is unethical to not provide the
one treatment being considered, one possibility is a wait control study where we compare
before (waiting), during and after treatment (to see if the effect persists) for each
individual.
It may also be important to look at cost-effectiveness, which requires collecting data about the
full costs of each treatment (and control).
Note that all studies that involve treatment of or personal data collection from human or
animal subjects require ethical approval. For clinical human studies, this involves the
Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong
West Cluster (IRB), for live animals it involves the Committee for the Use of Live Animals
in Teaching and Research (CULATR) while for other studies with personal data or human
subjects, it is the Human Research Ethics Committee for Non-Clinical Faculties
(HRECNCF).
Factorial Design
A factorial design is used to evaluate two or more factors simultaneously. The treatments are
combinations of levels of the factors. The advantages of factorial designs over one-factor-at-
a- time experiments are that they are more efficient and they allow interactions to be
detected.
For a), we usually want a full factorial design, which means generating all possible
combinations of the levels of the factors, which can be a large number as the number of factors
and/or levels increases.
For b) we usually assume a quadratic response and then choose data points to help us find
the maximum or minimum – this is called a response surface design
For c) we are interested in a screening design, which usually means assuming only 2 levels for
each factor and also using a fractional factorial design. In this case, we are usually assuming
that all the interactions are small (or at least that all higher order interactions are small) and
we focus on the main effects of each factor, e.g. 10 factors with 2 levels each = 1024
possibilities, one eighth fraction would be 128, a sixteenth would be 64, a thirty-second would
be 32, a sixty-fourth would be 16, while 5 factors with 3 levels each=243 possibilities, one
third fraction would 81, one ninth would be 27
For d) we need data spread out over different levels for the factors in order to identify any non-
linearity. This is only feasible for a small number of factors. We need at least 3 levels in order
to have any chance of identifying non-linearity at all.
When data are collected in a series over time, we call this a time series. Standard methods
require that the data are collected at equally spaced time points (e.g. hourly, daily, monthly
or yearly). The key difference from simple linear models is that the assumption that the
errors are independent is usually no longer valid – the simplest generalization is that the
errors are autocorrelated – in other words
Correlation of the errors for yt and yt+k = k, where yt is the observed value at time t and yt+k
is the observed value at time k units later.
You can see that this says that errors that are closer together in time have a stronger
correlation. If is small, then the usual linear models will be appropriate, so we can test
whether is zero in order to decide whether we need to use this generalization. This test is
called the Durbin-Watson test and is wise if our data has an obvious ordering. We examine this
test for the Longley dataset, where it shows marginal significance for a small autocorrelation:
This situation (of <>0) requires a relatively simple generalization of the linear models we
have been discussing. Autocorrelation means that we may have less information than we think
(i.e. the effective sample size is smaller).
Another type of time series model is a seasonal model. This means that we include an
independent variable that denotes which season (e.g. spring, summer, autumn and winter).
Clearly, this is just a particular factor, so it can be included in the linear models we already
know how to handle, if we know the length of the cycle. “Seasons” are not necessarily
seasons in the usual sense of the word, but can be hours of day or stages of any cycle. Note
that finding the length of the cycle may be an important part of the analysis for some data
sets. Special cases are where the seasonal pattern follows a sinusoidal curve (e.g. tide
heights and other lunar/solar patterns).
The next stage of sophistication in time series models is differencing, moving averages and
autoregressive models. The basic idea of these generalizations (usually considered together) is
that by using differences of the observed y (autoregressive models) or of the errors (moving
averages), possibly after initial differencing to remove trends, we can get a simple linear
model. These models are known as ARIMA models (autoregressive integrated moving
average)
In other words: Autoregressive (AR) models: forecasts are based on a linear combination of
previous values
Moving Average (MA) models: forecasts are based on a linear combination of previous errors
(i.e. a linear combination of residuals)
ARIMA models: combine differencing, AR and
In other words, we calculate the difference between an observation and the one before it in
time. This can be used to remove trends, as standard methods assume stationarity (the mean
and variance are stable over time)
General ARIMA models are hard to identify, but special cases are easier:
For pure MA with k terms, the autocorrelation function is zero after the kth term, where
autocorrelation is simply the sample correlation of yt with yt-k for different values of k.
For pure AR with k terms, the partial autocorrelation function (pacf) is zero after the kth
term, where the pacf is the sample correlation between yt with yt-k for different values of k,
after first removing the effects of the intervening y, i.e. y t-1 to yt-k+1 on both variables by
using linear regression, i.e. pacf is the autocorrelation of the residuals from the regressions.
The first example is looking at monthly Ozone concentrations, which show a strong
seasonal pattern and a downward trend.
The second example looks at the yearly numbers of lynx (a wild cat) caught in Canada and
shows a strong cycle of 9.5 years.
This example shows the pattern for passenger numbers – we can see that log
transformation is helpful here
Finally, this time series plot for CO2 shows a strong trend and seasonal pattern.
17: Generalized Linear Model
If Y is count data, then the distribution is Poisson with parameter λ, E(y)=λ and g(y) is
log(y) If Y is binary outcomes, i.e. success=1/failure=0 then the distribution is
Binomial(1,p), E(y) is p, the probability of success and g(y) is log(y/(1-y)) = logit (y)
Can also model ordinal Y, using ordinal logistic model.
We can apply similar ideas to identify a good model that explains the variation in y
Please see the book “Generalized Linear Models” by Nelder and McCullagh for more technical
details and examples.
i.e. expected log odds is linear in the factors and covariates (vs expected value for y is linear for
general linear models).
As the linear model moves from large negative to large positive, we shift from
These models need to be solved iteratively, but the computer shields us from this.
Can extend these models for ordinal data to what is called ordinal logistic regression, which
assumes a series of logistic regressions, which have the same slope but different intercepts.
As this model only adds one parameter per level, it is still feasible for large numbers of levels.
There is also an extension for nominal variables, but we usually use an alternative
formulation called log-linear models.
independence) or equivalently
Log(Pr(i,j))=Log(Pr(i))+Log(Pr(j))
Note that this is symmetric in the variables (does not distinguish between independent
and dependent variables).
18: JMP Basics
Learning objectives: understand basics of using JMP
Data Tables
We will assume that we can represent data as cases=rows and variables=columns, i.e. basic
spreadsheet concept. The main variation on this is that we may use one column as a count or
frequency column. This allows us to avoid entering every case, when some of the cases are
identical. (Illustrate simple 2x2 table in JMP)
Variable Type
One key issue when doing statistics is the type of variables that we have, in particular,
whether they are nominal, ordinal or continuous. In JMP, we have to select the type for
each column and this helps in that JMP can be smart in identifying appropriate analysis as
long as the type is correct. (Illustrate in JMP)
Entering Data
Data can come from multiple sources, including import from Excel or a text table,
keyboard entry or randomly generated. For importing, you need to know what the delimiter
is and whether the names of variables are in the first row. JMP tables can also be exported
similarly. (Illustrate import/export of animals from Sample Import folder and keyboard
entry of new data)
Row Selection
One of the beauties of an interactive package is that we can select rows in one place and
they will be highlighted in other places.
The rows can be selected manually (by selecting in the spreadsheet view),
programmatically (will show), or by lassoing points in a graph (will show). Selected rows
can then be coloured, hidden or excluded. This is invaluable if trying to understand
whether there is a problem with some data points. More detailed selections can be made
using repeated use of row selection to narrow down or expand the selected rows. (illustrate
different ways of selecting rows and then excluding or coloring using students)
Manipulating Tables
Sometimes you need to reorganize tables, using sorting, summarizing or stacking. You can
also join tables together. (summarise Companies, stack Cheese taste, subset(rows), sort(rows),
transpose(swap rows and columns), concatenate(rows), join (columns), split(reverse of stack))
Formulae for creating variables or generating random data
If we want to adjust data (e.g. standardize by subtracting the mean and dividing by the
S.D), recode or create random data, the formula editor is our friend (like a more
sophisticated formula in Excel applied to columns rather than cells). (illustrate using
students to standardize, hotdogs to recode, pendulum to transform and create randdist to
generate)
19: Data display with JMP
Learning objectives: understand how to display data in JMP
Univariate distributions
We start by looking at what JMP does when we ask for distributions for different types of
variable.
Key questions include whether the data is unimodal (if not, perhaps there is more than one
process generating the data) and whether it might follow a Normal distribution (possibly
after transformation). We can use alternative plots that may offer advantages in some
situations (e.g. box plots and stem-and-leaf) and Normal quantile plots (which plots the
data against what we would expect if the data came from a Normal distribution with
matching mean and variance and provides confidence limits to illustrate if the data is non-
Normal).
It is important to remember that statistical significance may be different to practical
significance
Graphics can help us identify when the data is truncated or there are atypical points.
Sometimes, the test of normality is rejected for very minor deviations from normality,
because there is a very large dataset. We also need to be careful to look at data within
each group, rather than the overall dataset.
Non-normal data
Sometimes a power transformation turns the data into Normal data (i.e. the residuals are
Normal). We can write this as:
Z=(y-1)/
So that the limiting case of goes to zero becomes Z=Ln(y) (which is the natural logarithm).
Log transforms often make sense because they turn multiplicative or ratio processes in
additive and subtractive ones (which underlie the Normal distribution). They can work as long
as y is always positive (examples later). If data includes zeros, may add a small amount before
doing log transform to account for rounding (add ¼ if rounded to integer) or truncation (add
½ if truncated to integer)
If power transform does not work, can use non-parametric methods – simplest idea is to
use ranks as a transformation – i.e. analyse ranks rather than raw data – t-test and F-test
become Mann-Whitney and Kruskal-Wallis tests. Only difference is that proper
nonparametric analysis takes accounts of ties (equal values) when calculating significance
levels, which is missed if just using ranks in standard analysis. Without accounting for
ties, we think we have more information than we really do. (Ranks can be added through
the save option in Distribution)
Multivariate data
Can do multiple bivariate plots or use principal component approach
Bivariate plots means plotting all the pairs of dependent variables, which does not
always reveal all the structure of the data.
Principal components is based on the idea that linear combinations of the dependent
variables may be able to summarise most of the variability in a relatively small number of
dimensions. The extreme case would be where all the variables are perfectly correlated (+1
or -1), when one variable can summarise all the variability. The disadvantage is that the
linear combinations are not always easy to interpret.
20: Model Building with JMP
Learning objectives: understand how to build models in JMP
Two groups
Testing for 2 groups is easy, as it is still based on the t-test.
We need to distinguish between paired and unpaired data - paired data turns back into 1
group data by looking at differences.
Multiple Regression
Model selection – dangers of “data mining” leading to models that are too complex (linnerand
example) (average number of variables added by chance will be significance level times the
number of variables tested, if we ignore correlations).
Diagnostics – residuals and leverage
Residuals help us identify model deficiencies (are the residuals Normal, is the variance
constant, is there are a missing variable, do we need a transformation)
Look for patterns in residuals against row, against covariates for mean or variance
Row (sequence) pattern for residual suggests missing variable (time?) or a change over time.
Pattern for residual against covariate suggests need for products or powers of covariates.
Pattern in variance suggests need for transformation to stabilize variance (assumed
constant in standard models)
Transformation – can try log, rank etc.
Leverage helps us identify high dependence of model fit on a few key data points (model
not necessarily “wrong”, but some risk and need to check).
Collinearity – problem with covariates that are collinear (i.e. can predict one covariate from
others), so parameter estimates not well defined, or may change considerably depending on
which other variables included.
The difference from the general linear model is that we have multiple dependent variables,
which are correlated (If uncorrelated, can look at them separately). Generally assume that
the same independent variables for all the dependent variables (i.e. they have similar causal
processes). In order to choose the appropriate tests, need to know whether the
measurements are all repeated measurements of the same construct (repeated measures) or
not. If not,
usually use identity matrix formulation, which means focus primarily on the separate
variables, rather than some specific combination of them.
Extend to data that does not follow the Normal distribution, but some other
We can apply similar ideas to identify a good model that explains the variation in y
Please see the book “Generalized Linear Models” by Nelder and McCullagh for more technical
details and examples.
Ordinal Data with JMP
As the linear model moves from large negative to large positive, we shift from
Pr(level 1) close to 1 and Pr(level 2) close to 0 to Pr(level 2) close to 1 and Pr(level 1)
close to 0. These models need to be solved iteratively, but the computer shields us from
this.
Can extend these models for ordinal data to what is called ordinal logistic regression, which
assumes a series of logistic regressions, which have the same slope but different intercepts.
As this model only adds one parameter per level, it is still feasible for large numbers of
levels.
There is also an extension for nominal variables, but we usually use an alternative
formulation called log-linear models.
These models assume a null hypothesis of:
Pr(Row i and Col j)=Pr(Row i) x Pr(Col j)
Or Log(Pr(i,j))=Log(Pr(i))+Log(Pr(j))
Which is why they are called log-linear.
Note that this is symmetric in the variables (does not distinguish between independent and
dependent variables) and corresponds to independence of row and column in Chisquared
goodness of fit.
Compositional data with JMP
This is data where non-negative variables sum to a known constant – usually 100% or 1,
such as when we have a set of proportions. In this case, correlations do not have the usual
meaning because of the constraint. If the variables are all within the range (0,1), i.e. there
are no zeroes or ones, we can use a log ratio transformation that is meaningful if we
believe that relative ratios are what matter. For example, if we have 3 components: (x1, x2,
x3) that sum to 1, then we can look at y1=log(x1/x3), and y2=log(x2/x3) using the standard
General Linear Model approach.
Note that we can map these log-ratios back to the original x space and that it does not matter
which of the x’s we use as a divisor for General Linear models. Please see John Aitchison’s
book: “The Statistical Analysis of Compositional Data” for more details of this approach.
How to install R depends on which operating system you use and even on which version of
that operating system. As R is open source, you can install it from scratch, i.e. compiling
the source code, but that is not recommended unless you are already familiar with
compiling code from scratch on your system, as it can be hard to get it to work properly.
Instead, visit the following repository where the latest version of R can always be
downloaded, together with installation instructions.
https://cran.r-project.org
Important Notes:
1) On OSX, you must install XQuartz first, which provides a UNIX based windowing system known as X11, and you must log out
and log in again before installing R.
2) You need to turn off anti-virus software, which otherwise may block R from moving files and hence completing the
installation of packages.
3) On Windows, you need access rights to the folder you have installed R in, or you can right click the R icon and
select "Run as administrator."
4) On Windows, please choose the "single-document "R interface (SDI). In the Startup options screen of the R
installer, select Yes (customized startup). Then select the SDI (single-document interface) in preference to the
default MDI (multiple-document interface).
5) It seems that packages installation may have problems if you have not set your location/region to an English
language location (e.g. Australia, Canada, UK, USA) so I strongly encourage making this change before installing R
Commander - you can change it back after installation.
6) If you get warnings like: Setting LC_CTYPE failed, using "C", then switch to US or British locale, or type into R:
On OSX: system("defaults write org.R-project.R force.LANG en_US.UTF-8").
On Windows/LINUX: Sys.setlocale("LC_ALL","English")
7) For OSX, if you have problems, you can try to run R from Terminal instead of double clicking on the R.app
8) If the files do not download, check if your firewall is blocking the download
9) Check that all dependencies are installed, if not install them manually until there are no errors when loading the
library.
10)R is case sensitive for variable names, so be careful.
However, these installations provide command line interfaces rather than menu-based
systems. So, we will also install R Commander, which provides menus for many common
statistical procedures. Details can be found here:
https://socialsciences.mcmaster.ca/jfox/Misc/Rcmdr/index.html
R Commander is an add-on to R called a package, so we can install it from inside R, with this
instruction:
install.packages("Rcmdr")
(allow R to install any dependencies, meaning other software needed to make it work, if
it asks which mirror to use, select the first choice)
library(Rcmdr)
On the Mac, you can either start R by double-clicking the R GUI program installed, or by
typing R in the Terminal (which can be found in the Utilities folder). Using Terminal may
work better if you intend to use R Commander, which sometimes conflicts with the R GUI.
One good thing about R Commander is that it shows you which R commands it used in the R
Script window.
A more advanced system than R Commander, called R Studio, which includes a debugger
and visualisation tools and can operate as a browser interface to a server, can be found here
(the basic version is free):
https://www.rstudio.com
Data Tables
We will assume that we can represent data as cases=rows and variables=columns, i.e. basic
spreadsheet concept. Note that weights in R usually mean precision weights, not frequency
weights as in SPSS and will usually give different results as precision weights do not affect
degrees of freedom calculations.
Variable Type
One key issue when doing statistics is the type of variables that we have, in particular,
whether they are nominal, ordinal, interval or ratio scale. R only knows categorical (caused by
entering anything non-numeric as data, which it then assumes is nominal scale) and numeric
data. It is your responsibility to know whether your numeric data are nominal, ordinal,
interval or ratio scale and hence whether the analysis is appropriate. R Commander assumes
that all numeric data is interval scale, so if you have numeric coding of any categorical
variables, it is wise to use R Commander to create new factors to use instead and hence reduce
the risk of mistakes (illustrate creating a factor from a numeric variable).
Entering Data
Data can come from multiple sources, including import direct from Excel, SPSS, SAS,
Minitab, STAT, URL or a text table, keyboard entry or randomly generated. Some R packages
already include datasets, which can easily be loaded.
For importing from text, you need to know what the delimiter (field separator) is and
whether the names of variables are in the first row. R data tables (known as data frames in
R) can also be exported similarly. (Illustrate import and keyboard entry of new data). R
uses NA to mean missing and if you enter anything that is not a number or NA, it assumes
the variable is a categorical variable (factor).
Saving Data
Saving data means that the data is saved as an R data frame and can be easily reloaded.
Manipulating Tables
Sometimes you need to reorganize tables (Data/Active data set), using sorting, aggregating,
subsetting or stacking.
Univariate distributions
We start by looking at what R Commander does when we ask for (statistics) summaries for
different types of variable.
It also offers graphs such as stem and leaf plots, histograms, box plots for numeric data
Key questions for interval scale data include whether the data is unimodal (if not, perhaps
there is more than one process generating the data) and whether it might follow a Normal
distribution (possibly after transformation). We can also use Normal quantile plots (which
plots the data against what we would expect if the data came from a Normal distribution
with
matching mean and variance and provides confidence limits to illustrate if the data is
non- Normal).
Sometimes a power transformation can turn the data into Normal distributed data (i.e.
the residuals are Normal, not the raw data). We can write this as:
Z=(y-1)/
So that the limiting case of goes to zero becomes Z=log(y) (which is the natural logarithm in
R, use log10 for log base 10).
Log transforms often make sense because they turn multiplicative or ratio processes into
additive and subtractive ones (which underlie the Normal distribution). They can work as long
as y is always positive (examples later). If data includes zeros, may add a small amount before
doing log transform to account for rounding (add ¼ if rounded to integer) or truncation (add
½ if truncated to integer)
If power transform does not work, can use non-parametric methods – simplest idea is to
use ranks as a transformation – i.e. analyse ranks rather than raw data – t-test and F-test
become Mann-Whitney/Wilcoxon and Kruskal-Wallis tests. Only difference is that proper
nonparametric analysis takes accounts of ties (equal values) when calculating significance
levels, which is missed if just using ranks in standard analysis. Without accounting for ties,
we think we have more information than we really do. (Ranks can be added through the
rank function)
Multivariate data
Can do scatterplot matrix (under Graphs) or use principal component approach (under
Statistics: dimensional analysis in R commander).
Scatterplot matrix means plotting all the pairs of dependent variables, which does not
always reveal all the structure of the data. Principal components is based on the idea that
linear combinations of the dependent variables may be able to summarise most of the
variability in a relatively small number of dimensions. The extreme case would be where
all the variables are perfectly correlated (+1 or -1), when one variable can summarise all
the variability. The disadvantage is that the linear combinations are not always easy to
interpret.
23: Model Building with R
Learning objectives: understand how to build models in R
Two groups
Testing for 2 groups is easy, as it is based on the t-test.
We need to distinguish between paired and unpaired data - paired data turns back into 1
group data by looking at differences.
Multiple Regression
Model selection – dangers of “data mining” leading to models that are too complex (linnerand
example) (average number of variables added by chance will be significance level times the
number of variables tested, if we ignore correlations).
Leverage helps us identify high dependence of model fit on a few key data points (model
not necessarily “wrong”, but some risk and need to check).
Collinearity – problem with covariates that are collinear (i.e. can predict one covariate from
others), so parameter estimates not well defined, or may change considerably depending on
which other variables included.
R commander allows us to include polynomials and splines (smoothed polynomials) and offers
numerical diagnostics and graphs to help check model assumptions.
Can use best subset to help identify which is the best model, given different criteria (AIC,
BIC, Mallows Cp are good criteria that penalise complexity in different ways, adjusted R2
is not as good, R2 is terrible unless all models have same complexity). Do not use
stepwise, always prefer best subset, because stepwise may miss the best fitting models.
General Linear Models with R
GLM means Multiple Regression + Multi-way ANOVA combined, sometimes called Analysis of
Covariance
Model selection is now much more complex as we need to select from a wide range of models.
Aim at simple answers (small degrees of freedom for model) that fit well and that can be
explained in terms of a sensible theory.
For multiple factors, can have many interactions, but we often assume that higher level
interactions are likely to be (relatively) unimportant.
Generalized Linear Model with R
Extend GLM to data that does not follow the Normal distribution, but some other
distribution, like Poisson, Gamma, Binomial
e.g. Y is count data, then the distribution is Poisson with parameter λ, E(y)=λ and g(y) is log(y)
e.g. Y is binary outcomes, i.e. success=1/failure=0 then the distribution is Binomial(1,p), E(y)
is p, the probability of success and g(y) is log(y/(1-y)) = logit (y) or can use probit.
We can apply similar ideas to identify a good model that explains the variation in y
Please see the book “Generalized Linear Models” by Nelder and McCullagh for more technical
details and examples.
Binary, Ordinal and Multinomial Data with R
i.e. expected log odds is linear in the factors and covariates (vs expected value for y is linear for
general linear models).
As the linear model moves from large negative to large positive, we shift from
Pr(level 1) close to 1 and Pr(level 2) close to 0 to Pr(level 2) close to 1 and Pr(level 1)
close to 0. These models need to be solved iteratively, but the computer shields us from
this.
We can extend these models for ordinal data to what is called ordinal logistic regression, which
assumes a series of logistic regressions, which have the same slope but different intercepts.
As this model only adds one parameter per level, it is still feasible for large numbers of levels.
There is also an extension for nominal data called multinomial logit, which has many
more parameters, because the slope is no longer assumed to be common across levels.
Compositional data with R
This is data where non-negative variables sum to a known constant – usually 100% or 1,
such as when we have a set of proportions. In this case, correlations do not have the usual
meaning because of the constraint (in the trivial case, if p and q add to 1, they must have a
correlation of -1). If the variables are all within the range (0,1), i.e. there are no zeroes or
ones, we can use a log ratio transformation that is meaningful as relative ratios are what
matter. For example, if we have 3 components: (x1, x2, x3) that sum to 1, then we can
look at
Note that we can map these log-ratios back to the original x space and that it does not matter
which of the x’s we use as a divisor for General Linear models. Please see John Aitchison’s
book: “The Statistical Analysis of Compositional Data” for more details of this approach.
Beyond R Commander
R offers many packages beyond those included in R Commander. They can be installed just
like R Commander, inside R. However, some caution is advised as not all packages are well
maintained. Usually, if there is an up-to-date package at the R repository, that is a good sign
that this package is well maintained, but the documentation may not always be very clear.
24: Hierarchical Linear Models
Why HLM:
Hierarchical Linear Models (HLM), otherwise known as multilevel models, mixed models
or random-effects models are used to model data that vary at more than one level.
For example, data from cluster sampling, such as students in classes in schools, where
students, classes and schools would be the three levels. These models allow us to understand
variation at the different levels and model the impact of factors and covariates at the different
levels.
Simple linear models assume that all observations are independent and have fixed variance,
which is not true in our school example, as students in the same class or in the same school
share common effects. Using simple linear models on the students is flawed because it
does not account for the common variation and will be biased. Using simple linear models
on school averages does not allow us to account for characteristics of the students, such as
age and parental characteristics, which may vary greatly across students and schools,
distorting the school level relationships.
HLM means that the expected value may have factors and covariates at each level (which
is simple to account for) and the error term in our model may have elements from each
level of
our hierarchy, which is more complex to account for in the estimation and hypothesis testing
process.
HLM can also be used for longitudinal data, where we have repeated measurements on the
same individuals, so those repeated measurements are not independent because of common
individual variation, so we again distinguish between the individual level and the
measurement occasion for that individual level.
For example, a simple linear regression model assuming Gaussian errors and repeated
measurements on individuals (a two level model) where the regression coefficients may be
different for different individuals, would look like this:
Bayesian
In both cases, we need first to be able to state what the probability density function is (if our
dependent variable is categorical, this changes to needing the probability mass function).
We will start with a simple model with constant mean and independent Gaussian errors:
The probability density for each data point (Gaussian density) is proportional to
exp(-1/2((y-theta)/sigma)2)
sigma
where theta is the mean and sigma is the standard deviation
Maximum Likelihood (ML)
theta)/sigma)2)
sigman
We prefer to look at the log likelihood, which is quadratic in theta for Gaussian errors.
ML requires finding the maximum, often found using simple calculus, by setting the
first derivative to zero and solve:
We can also find the approximate variance of the estimate by looking at the inverse of minus
the second derivative of LogL:
d2 LogL = -n/sigma2
d2 theta
so the estimated variance = sigma 2/n
Unfortunately, for small samples, ML can be highly biased, especially if the number of
parameters is large. This can be improved using the bootstrap adjustment, but that is
beyond our scope here.
Also, for hierarchical models, ML does not work very well, as shown by the James-
Stein estimator:
In the simple case of Y following a Normal distribution with constant variance and n
measurements from each of m groups.
ML would suggest using the group sample means as the best estimate of the group population
means.
However, shrinkage of the group estimates towards the overall mean is superior, i.e.
group mean estimate = overall sample mean + k (group sample mean - overall sample mean)
Bayes
as p(data) is a constant and p(data|theta) is the likelihood L(theta|data) (notice that we have
switched from a function of the data conditional on theta to a function of theta conditional
on the data, this depends on the Likelihood principle), we get
Notice that this assumes we can make probability statements about theta, which cannot be
interpreted as long-run relative frequencies, but as subjective degree of belief, which has
been
criticized as unscientific, but this approach enables us to quantify all the elements of
uncertainty. In other words, we can apply probabilities not only to observable outcomes, but
also to parameters in a model, where there are no long-run relative frequencies.
We talk about this as Bayesian updating from the prior distribution (before seeing the
new evidence) using the likelihood to get the posterior distribution, which incorporates
the new evidence.
The conceptual challenge is how to construct the prior distribution before seeing any
evidence, but in practice, we can use a widely spread out distribution, such as Gaussian
with large variance.
if our prior distribution for theta is Gaussian with mean theta 0 and variance sigma02
If we use the term precision to mean the inverse of the variance, then the posterior
distribution for theta is also Gaussian with
posterior precision=P0+Pn,
where P0 is the prior precision (i.e. 1/sigma02) and Pn is the sample precision (i.e. n/sigma2)
so if P0 is very small, the answer is similar to ML, but the sample mean is shrunk
towards theta0
If we build a Bayesian HLM for the group example above and assume
that the group means are a Gaussian sample from the prior mean,
then the group means are shrunk towards the overall sample mean and these estimates
outperform ML estimates.
Bayesian methods are naturally hierarchical as regards the parameters, so we can have
second order priors for theta0 and sigma02
In short, there is a compelling empirical case that Bayes methods are superior for HLM in
terms of providing better estimates (in fact, all admissible procedures are either Bayes
methods or limits of Bayes methods, where admissible means I cannot find a procedure that is
always better)
As we shall see later, Bayesian methods also allow us to easily replace the Gaussian
assumption with other distributions, including discrete distributions, allowing many
different sensitivity analyses to be considered.
If we have a prior for theta of Beta(a,b), then the posterior for theta is Beta(a+r,b+n-r)
This is what we call a conjugate prior as the posterior is in the same class of distributions
as the prior. This is simple and does not need any sophisticated software, but if I sample
successes from k different individuals, who have different theta, then I have a hierarchical
model and the analysis gets more complex.
Comparison of classical, ML and Bayes in simple situation
If we just know that in n trials, there were r successes, what is our best estimator for theta,
the chance of success?
ML is clearly r/n, even though that may estimate theta as 0 or 1, which is impossible unless we
have a 2 headed or 2 tailed coin
Classical (unbiased minimum variance) cannot answer the question without knowing
what experiment we did!
Binomial (r is random, n is fixed), the sufficient statistic is r and the best estimator is r/n
Negative Binomial (r is fixed, n is random), the sufficient statistic is k (# of failures) and the
best estimator is (r-1)/(n-1)
Marginal Posterior Distribution
There is one practical problem with Bayesian methods that we have not yet discussed:
Bayes formula provides the joint posterior distribution for all parameters that we are
estimating. How do we obtain the marginal posterior distribution for specified
parameters?
In some simple cases, we can do exact symbolic integration. For example, if we use an
inverse gamma prior distribution for the variance (i.e. gamma prior for the precision) and
Gaussian prior for the mean, we can show that the marginal posterior for the mean is a
Student-T distribution and the marginal posterior for the variance is an inverse gamma (i.e.
gamma posterior for the precision).
Markov Chain Monte Carlo (MCMC)
However, in many cases, we cannot do exact symbolic integration and must use numeric
integration. Unfortunately, precise numerical integration takes a long time if the number of
parameters is large (i.e. much greater than 10) as the amount of calculation increases in
general exponentially with the number of parameters, so we need to rely on Monte Carlo
integration (i.e. sampling). Monte Carlo sampling works by allowing us to calculate
E(g(X)) = g(xi),
N
where xi is the ith sample out of N from g(x), and this is not very precise unless we generate a
very large number of points to ensure that the sampling error is small.
Fortunately, there was a major advance in the 1980s known as Markov Chain Monte Carlo,
which uses sampling in a much more efficient way to generate marginal posterior distributions
and any expectations we might be interested in. The key requirement is that we write the
statistical model as a full joint probability distribution for all parameters and observable data.
This is always possible for parametric models and distibutions, but often there is conditional
independence between many elements of the model, which greatly simplifies the calculations.
MCMC involves sampling in turn from each conditional distribution (this is called using the
Gibbs sampler).
Sensitivity analysis is often simple to implement by generalizing the model, for example, if we
replace a Gaussian by a T distribution, which includes Gaussian as a special case, where the
degrees of freedom are large.
This MCMC sampling takes advantage of the fact that if we know the conditional posterior
densities for each parameter (this is the Markov Chain part) and use them to generate
sequential samples, then we converge (under quite weak conditions) to the correct posterior
density, yielding all the marginal posterior distributions as an outcome. In many situations,
we can easily find those conditional densities if we have a parametric model.
The biggest challenge is arguably to confirm convergence – have we converged yet, or are
we still moving very slowly across the parameter space? The technical term is that our
Markov chain should be stationary. One basic approach is to run parallel simulations
(chains) starting from different initial values and see if they all converge to the same region
and by comparing the variability within chains to variability across chains. This fits well
with using a cluster of high performance computers. We can often improve convergence by
parameterizing our model such that the parameters have low correlation.
This was implemented in software called WinBUGS, which only runs under Windows and is
no longer developed so we will use open source software called Just Another Gibbs Sampler
(JAGS), that was completely rewritten but uses the same model description language (known
as BUGS), so it can solve all the many examples developed for WinBUGS, including many
hierarchical models.
The sampler needs to be run long enough to ensure that we have converged on the
posterior distribution (so we have completed what is called "burn-in") first. Each new
iteration then produces a set of estimates for all the parameters, taken from the posterior
distribution, so we will run for sufficient iterations to generate sufficient sampling
precision in the posterior summaries of interest, such as posterior means, posterior standard
deviations or marginal posterior distribution densities.
JAGS has R interfaces called rjags, runjags, R2jags,jagsUI but to save time ensuring everyone
has R correctly installed, we will instead work through some examples in JAGS directly. The
R interfaces include a package called Coda to check convergence, which is probably the most
challenging element of MCMC.
References
JAGS manuals:
https://sourceforge.net/projects/mcmc-jags/files/Manuals/4.x/
JAGS examples:
https://sourceforge.net/projects/mcmc-jags/files/Examples/4.x/
JAGS software:
https://sourceforge.net/projects/mcmc-jags/files/JAGS/4.x/
BUGS book:
https://www.mrc-bsu.cam.ac.uk/software/bugs/the-bugs-project-the-bugs-book/
Windows installer hint: The path to the installer must contain only latin letters, no non-
English symbols, such as Chinese.
BUGS language
BUGS language understands scalars and arrays, including subsets of arrays and simple for
loop structures
model {
for (i in 1:N) {
Y[i] ~ dnorm(mu[i], tau)
mu[i] <- alpha + beta * (x[i] - x.bar)
}
x.bar <- mean(x)
alpha ~ dnorm(0.0, 1.0E-
4) beta ~ dnorm(0.0, 1.0E-
4) sigma <- 1.0/sqrt(tau)
tau ~ dgamma(1.0E-3, 1.0E-3)
}
Notes:
x is centered to reduce the correlation in the estimates for alpha and beta
sigma is there purely for understanding the results
BUGS data
The data part can be very simple and data is stored in R dump
as:
matrices are more tricky as we need to tell BUGS the structure, such as for a 3 x 2 matrix A
Our example of a simple linear model would need a data section like:
N<- 8
x<c(1,2,3,4,5,6,7,8)
Y<-c(2,5,7,9,11,13,15,16)
It is also possible to simulate data inside BUGS like this:
data {
for (i in 1:N) {
y[i] ~ dnorm(0,1)}
}
or transform data
data {
for (i in 1:N) {
z[i] <- sqrt(y[i])}
}
Before running the examples, we first change our directory of the folder containing the
example files (cd changes directory (use double quotes for the pathname, ls lists files (dir for
Windows), pwd to show current directory)
In my case:
cd "/Volumes/Mac HD/Applications/JAGS/examples/classic-bugs/"
is where JAGS installed my examples, so I can easily access the example files.
HLM Gaussian regression example
This is the rats example originally used for WinBUGS in the Volume 1 set but now also
included in JAGS, from Illustration of Bayesian Inference in Normal Data Models Using Gibbs
Sampling by Alan E. Gelfand, Susan E. Hills, Amy Racine-Poon and Adrian F. M. Smith, in
Journal of the American Statistical Association, Vol. 85, No. 412 (Dec., 1990), pp. 972-985.
This example is taken from section 6 of the paper, and concerns 30 young rats whose weights
were measured weekly for five weeks.
The model is a random effects linear growth curve where the intercept and slope depend on
the rat and xi is the day (8,15,22,29,36) of the measurement.
where xbar = 22, and τ represents the precision (1/variance) of a normal distribution.
In BUGS language, this is expressed as:
model {
for (i in 1:N) {
for (j in 1:T) {
mu[i,j] <- alpha[i] + beta[i]*(x[j] - x.bar);
Y[i,j] ~ dnorm(mu[i,j],tau.c)
}
alpha[i] ~ dnorm(alpha.c,tau.alpha);
beta[i] ~ dnorm(beta.c,tau.beta);
}
alpha.c ~ dnorm(0,1.0E-
4); beta.c ~
dnorm(0,1.0E-4);
tau.c ~ dgamma(1.0E-3,1.0E-3);
tau.alpha ~ dgamma(1.0E-3,1.0E-
3); tau.beta ~ dgamma(1.0E-3,1.0E-
3); sigma <- 1.0/sqrt(tau.c);
x.bar <- mean(x[]);
alpha0 <- alpha.c - beta.c*x.bar;
}
JAGS output in CODA format
The standard CODA format for monitors that do not pool values over iterations is to create an
index file and one or more output files.
The index file is has three columns with, one each line:
1. A string giving the name of the (scalar) value being recorded
2. The first line in the output file(s)
3. The last line in the output file(s)
If we do not have R (and the package coda) we can easily look at these values in Excel to
check for convergence. If we run at least 2 chains, we can check that they converge to the
same values
JAGS script example
We will run this for 1,000 updates to converge and then 10,000 more to yield estimates using
this JAGS script:
model in "rats.bug"
data in "rats-
data.R" compile,
nchains(2) inits in
"rats-init.R"
initialize
update 1000
monitor alpha0
monitor beta.c
update 10000
coda *
This is the seeds example from Volume 1, using data from Table 3 of Crowder, M. J.
1978. Beta-binomial Anova for proportions, Journal of the Royal Statistical Society (C,
Applied Statistics) 27(1), 34–37.
This example concerns the proportion of seeds that germinated on each of 21 plates arranged
according to a 2 by 2 factorial layout by seed and type of root extract.
ri ~ Binomial(pi , ni )
where x1i , x2i are the seed type and root extract of the ith plate, and an interaction term
model {
alpha0 ~ dnorm(0.0,1.0E-6); # intercept
alpha1 ~ dnorm(0.0,1.0E-6); # seed coeff
alpha2 ~ dnorm(0.0,1.0E-6); # extract coeff
alpha12 ~ dnorm(0.0,1.0E-6); # interaction coeff
tau ~ dgamma(1.0E-3,1.0E-3); # 1/sigma^2
sigma <- 1.0/sqrt(tau);
for (i in 1:N) {
b[i] ~ dnorm(0.0,tau);
logit(p[i]) <- alpha0 + alpha1*x1[i] + alpha2*x2[i] + alpha12*x1[i]*x2[i] + b[i];
r[i] ~ dbin(p[i],n[i]);
}
}
Again we will run 1000 times to "burn-in", then 10,000 times to get estimates
School hierarchical model with pupil and school covariates
This is the schools example from Volume 2, using data from “A Multilevel Analysis of School
Examination Results”, by Harvey Goldstein; Jon Rasbash; Min Yang; Geoffrey Woodhouse;
Huiqi Pan; Desmond Nuttall; Sally Thomas, Oxford Review of Education, Vol. 19, No. 4.
(1993), pp. 425-433 (pdf in folder).
They present an analysis of examination results from inner London schools. They use
hierarchical or multilevel models to study the between-school variation and calculate
school- level residuals in an attempt to differentiate between 'good' and 'bad' schools.
Data
Standardized mean examination scores (Y) were available for 1978 pupils from 38 different
schools. The median number of pupils per school was 48, with a range of 1--198. Pupil-level
covariates included gender plus a standardized London Reading Test (LRT) score and a verbal
reasoning (VR) test category (1, 2 or 3, where 1 represents the highest ability group) measured
when each child was aged 11. Each school was classified by gender intake (all girls, all boys or
mixed) and denomination (Church of England, Roman Catholic, State school or other); these
were used as categorical school-level covariates.
Model
We consider the following model, which essentially corresponds to Goldstein et al.'s model 1.
logτij = θ + φ LRT ij
We wish to specify a regression model for the variance components, and here we model the
logarithm of τ ij (the inverse of the between-pupil variance) as a linear function of each pupil's
LRT score. This differs from Goldstein et al.'s model which allows the variance σ 2 ij to depend
linearly on LRT. However, such a parameterization may lead to negative estimates of σ2 ij .
This yields this model:
model {
for (p in 1 : N) {
Y[p] ~ dnorm(mu[p], tau[p])
mu[p] <- alpha[school[p], 1] + alpha[school[p], 2] * LRT[p] + alpha[school[p], 3] * VR[p, 1] +
beta[1] * LRT2[p] + beta[2] * VR[p, 2] + beta[3] * Gender[p] + beta[4] * School.gender[p, 1]
+ beta[5] * School.gender[p, 2] + beta[6] * School.denom[p, 1] +
beta[7] * School.denom[p, 2] + beta[8] * School.denom[p, 3]
log(tau[p]) <- theta + phi * LRT[p]
sigma2[p] <- 1 / tau[p]
LRT2[p] <- LRT[p] * LRT[p]
}
min.var <- exp(-(theta + phi * (-34.6193))) # lowest LRT score = -34.6193
max.var <- exp(-(theta + phi * (37.3807))) # highest LRT score = 37.3807
# Priors for fixed effects:
for (k in 1 : 8) {
beta[k] ~ dnorm(0.0, 0.0001) }
theta ~ dnorm(0.0, 0.0001)
phi ~ dnorm(0.0, 0.0001)
# Hyper-priors:
gamma[1 : 3] ~ dmnorm(mn[1:3 ], prec[1:3 ,1:3 ]);
T[1 : 3, 1 : 3 ] ~ dwish(R[1:3 ,1:3 ], 3)}
Note that alpha1[] measures the 'residual effect' for school j after adjusting for pupil- and
school-level covariates. This might represent an appropriate quantity by which to rank schools'
performance in terms of how good the results are after accounting for the language ability of
the students. We can calculate the ranks for each complete iteration and hence find the mean
and s.d. of the rank for each school.
Inhalers (ordered categorical repeated measures)
This example is from Ezzet and Whitehead (1993), who analyse data from a two-treatment,
two-period crossover trial to compare 2 inhalation devices for delivering the drug
salbutamol in 286 asthma patients. Patients were asked to rate the clarity of leaflet
instructions accompanying each device, using a 4-point ordinal scale.
The response Rit from the ith subject (i = 1 to 286) in the tth period ( t = 1,2) thus assumes
integer values between 1 and 4. It may be expressed in terms of a continuous latent variable Y
is taking values on (-inf, inf) as follows:
where a0 = -inf and a4 = inf. Assuming a logistic distribution with mean mit for Y it , then the
cumulative probability Qitj of subject i rating the treatment in period t as worse than category j
(i.e. Prob( Y it >= a j ) is given by
where bi represents the random effect for subject i . Here, m s i t depends only on the period
t and the sequence s i = 1,2 to which patient i belongs. It is defined as
m11 = b / 2 + p / 2
m12 = - b / 2 - p / 2 - k
m21 = - b / 2 + p / 2
m22 = b / 2 - p / 2 + k
where b represents the treatment effect, p represents the period effect and k represents the
carryover effect. The probability of subject i giving response j in period t is thus given by:
The BUGS language for this model is shown below. We assume the b i 's to be normally
distributed with zero mean and common precision t. The fixed effects b, p and k are given
vague normal priors, as are the unknown cut points a 1, a2 and a3. We also impose order
constraints on the latter using the I(,) notation in BUGS (T() in JAGS as this is truncation,
not censoring), to ensure that a1 < a2 < a3.
model
{
#
# Construct individual response data from contingency table
#
for (i in 1 : Ncum[1, 1])
{ group[i] <- 1
for (t in 1 : T) { response[i, t] <- pattern[1, t] }
}
for (i in (Ncum[1,1] + 1) : Ncum[1, 2]) {
group[i] <- 2 for (t in 1 : T) { response[i, t] <- pattern[1, t] }
}
for (k in 2 : Npattern) {
for(i in (Ncum[k - 1, 2] + 1) : Ncum[k, 1]) {
group[i] <- 1 for (t in 1 : T) { response[i, t] <- pattern[k, t] }
}
for(i in (Ncum[k, 1] + 1) : Ncum[k, 2]) {
group[i] <- 2 for (t in 1 : T) { response[i, t] <- pattern[k, t] }
}
}
#
# Model
#
for (i in 1 : N) {
for (t in 1 : T) {
for (j in 1 : Ncut) {
#
# Cumulative probability of worse response than j
#
logit(Q[i, t, j]) <- -(a[j] + mu[group[i], t] + b[i])
}
#
# Probability of response =
j#
p[i, t, 1] <- 1 - Q[i, t, 1]
for (j in 2 : Ncut) { p[i, t, j] <- Q[i, t, j - 1] - Q[i, t, j] }
p[i, t, (Ncut+1)] <- Q[i, t, Ncut]
response[i, t] ~ dcat(p[i, t, ])
culmative.response[i, t] <- culmative(response[i, t], response[i, t])
}
#
# Subject (random) effects
#
b[i] ~ dnorm(0.0, tau)
}
#
# Fixed effects
#
for (g in 1 : G) {
for(t in 1 : T) {
# logistic mean for group i in period t
mu[g, t] <- beta * treat[g, t] / 2 + pi * period[g, t] / 2 + kappa * carry[g, t]
}
}
beta ~ dnorm(0, 1.0E-06)
pi ~ dnorm(0, 1.0E-06)
kappa ~ dnorm(0, 1.0E-
06)
Note that the data is read into BUGS in the original contigency table format to economize
on space and effort. The indivdual responses for each of the 286 patients are then
constructed within BUGS.
Other examples
There are many other examples in the 3 volumes of WinBUGS example pdfs and full JAGS
sample files for Volumes 1 and 2.
Related tools
Instead of using rjags, runjags or R2jags, you can also consider NIMBLE, which is like JAGS
operating completely inside R
Stan
If you are trying to model something beyond what JAGS can manage (e.g. JAGS is too
slow or you need a distribution not included in JAGS), the best tool is probably the open
source software called Stan, which is more flexible, but requires stronger programming
skills. Stan is more efficient as it uses automatic differentiation of the log density to use a
more efficient sampler (Hamiltonian sampler instead of the Gibbs sampler), but cannot
handle categorical or missing data easily.
URL: http://mc-stan.org
For example, my PhD student was modelling compositional data, i.e. p1+p2+p3=1 and
p1,p2,p3>0. This is modelled using log ratios, i.e. log(p1/p3) and log(p2/p3) follow a bivariate
Gaussian distribution. She wanted to examine censoring of p1,p2,p3 and this cannot be done in
JAGS as the censoring option in JAGS can only be applied to the sampled distributions (i.e. the
log ratios), not on p1,p2,p3 directly. In Stan we could find the censored log-density by
integrating numerically over the censoring regions. Convergence was slow, but this was the
correctly formulated problem (versus censoring of the log-ratios).
Statistical Advice Centre for Students (STACS)
All HKU research students are entitled to a maximum of 4 hours free statistical consulting
(over their time as a graduate student), provided by me.
https://www.gradsch.hku.hk/gradsch/current-students/courses-workshops-dialogues-career-
preparation/supporting-courses-services/statistical-advice-centre-for-students-stacs#
where you can find the link to download the form for your supervisor to sign and return to
the Graduate School for checking. The form is called a support services form.
View publication stats