Professional Documents
Culture Documents
Proc Logistic
Proc Logistic
Contents
Overview: LOGISTIC Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5379
Getting Started: LOGISTIC Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5382
Syntax: LOGISTIC Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5389
PROC LOGISTIC Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5390
BY Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5403
CLASS Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5404
CODE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5407
CONTRAST Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5407
EFFECT Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5410
EFFECTPLOT Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5412
ESTIMATE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5413
EXACT Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5414
EXACTOPTIONS Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5416
FREQ Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5420
ID Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5420
LSMEANS Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5420
LSMESTIMATE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5422
MODEL Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5423
NLOPTIONS Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5441
ODDSRATIO Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5441
OUTPUT Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5443
ROC Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5448
ROCCONTRAST Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5449
SCORE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5450
SLICE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5452
STORE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5452
STRATA Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5453
TEST Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5454
UNITS Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5455
WEIGHT Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5456
Details: LOGISTIC Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5456
Missing Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5456
Response Level Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5457
Link Functions and the Corresponding Distributions . . . . . . . . . . . . . . . . . . 5458
Determining Observations for Likelihood Contributions . . . . . . . . . . . . . . . . 5459
Iterative Algorithms for Model Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . 5460
5378 F Chapter 73: The LOGISTIC Procedure
g./ D ˛ C ˇ 0 x
For ordinal response models, the response, Y, of an individual or an experimental unit might be restricted to
one of a (usually small) number of ordered values, denoted for convenience by 1; : : : ; k; k C 1. For example,
the severity of coronary disease can be classified into three response categories as 1 = no disease, 2 = angina
pectoris, and 3 = myocardial infarction. The LOGISTIC procedure fits a common slopes cumulative model,
which is a parallel lines regression model based on the cumulative probabilities of the response categories
rather than on their individual probabilities. The cumulative model defines k response functions of the form
g.Pr.Y i j x// D ˛i C ˇ 0 x; i D 1; : : : ; k
where ˛1 ; : : : ; ˛k are k intercept parameters and ˇ is a common vector of slope parameters. You can specify
cumulative logit, cumulative probit, and cumulative complementary log-log links. These models have been
5380 F Chapter 73: The LOGISTIC Procedure
considered by many researchers. Aitchison and Silvey (1957) and Ashford (1959) employ a probit scale
and provide a maximum likelihood analysis; Walker and Duncan (1967) and Cox and Snell (1989) discuss
the use of the log odds scale. For the log odds scale, the cumulative logit model is often referred to as the
proportional odds model.
Besides the three preceding cumulative response models, the LOGISTIC procedure provides a fourth ordinal
response model, the common slopes adjacent-category logit model. This model uses the adjacent-category
logit link function, which is based on the individual probabilities of the response categories and has the form
Pr.Y D i j x/
log D ˛i C ˇ 0 x; i D 1; : : : ; k
Pr.Y D i C 1 j x/
You can use this model when you want to interpret results in terms of the individual response categories
instead of cumulative categories.
For nominal response logistic models, where the k C 1 possible responses have no natural ordering, the logit
model can be extended to a multinomial model known as a generalized or baseline-category logit model,
which has the form
Pr.Y D i j x/
log D ˛i C ˇi0 x; i D 1; : : : ; k
Pr.Y D k C 1 j x/
where the ˛1 ; : : : ; ˛k are k intercept parameters and the ˇ1 ; : : : ; ˇk are k vectors of slope parameters. These
models are a special case of the discrete choice or conditional logit models introduced by McFadden (1974).
The LOGISTIC procedure enables you to relax the parallel lines assumption in ordinal response models, and
apply the parallel lines assumption to nominal response models, by specifying parallel line, constrained, and
unconstrained parameters as in Peterson and Harrell (1990) and Agresti (2010). The linear predictors for
these models have the form
˛i C ˇ10 x1 C ˇ2i
0
x2 C .i ˇ3 /0 x3 ; i D 1; : : : ; k
where the ˇ1 are the parallel line (equal slope) parameters, the ˇ21 ; : : : ; ˇ2k are k vectors of unequal slope
(unconstrained) parameters, and the ˇ3 are the constrained slope parameters whose constraints are provided
by the diagonal i matrix. To fit these models, you specify the EQUALSLOPES and UNEQUALSLOPES
options in the MODEL statement. Models that have cumulative logits and both equal and unequal slopes
parameters are called partial proportional odds models, and models that have only unequal slope parameters
are called general models.
The LOGISTIC procedure fits linear logistic regression models for discrete response data by the method of
maximum likelihood. It can also perform conditional logistic regression for binary response data and exact
logistic regression for binary and nominal response data. The maximum likelihood estimation is carried
out with either the Fisher scoring algorithm or the Newton-Raphson algorithm, and you can perform the
bias-reducing penalized likelihood optimization as discussed by Firth (1993) and Heinze and Schemper
(2002). You can specify starting values for the parameter estimates.
Any term specified in the model is referred to as an effect. The LOGISTIC procedure enables you to
specify categorical variables (also known as classification or CLASS variables) and continuous variables as
explanatory effects. You can also specify more complex model terms such as interactions and nested terms
in the same way as in the GLM procedure. You can create complex constructed effects with the EFFECT
statement. An effect in the model that is not an interaction or a nested term or a constructed effect is referred
to as a main effect.
Overview: LOGISTIC Procedure F 5381
The LOGISTIC procedure uses ODS Graphics to create graphs as part of its output. For general information
about ODS Graphics, see Chapter 21, “Statistical Graphics Using ODS.” For more information about the
plots implemented in PROC LOGISTIC, see the section “ODS Graphics” on page 5516.
The remaining sections of this chapter describe how to use PROC LOGISTIC and discuss the underlying
statistical methodology. The section “Getting Started: LOGISTIC Procedure” on page 5382 introduces
PROC LOGISTIC with an example for binary response data. The section “Syntax: LOGISTIC Procedure” on
page 5389 describes the syntax of the procedure. The section “Details: LOGISTIC Procedure” on page 5456
summarizes the statistical technique employed by PROC LOGISTIC. The section “Examples: LOGISTIC
Procedure” on page 5518 illustrates the use of the LOGISTIC procedure.
For more examples and discussion on the use of PROC LOGISTIC, see Stokes, Davis, and Koch (2012);
Allison (1999); SAS Institute Inc. (1995).
proc logistic;
model y=x1 x2;
run;
The response variable y can be either character or numeric. PROC LOGISTIC enumerates the total number
of response categories and orders the response levels according to the response variable option ORDER= in
the MODEL statement.
You can also input binary response data that are grouped. In the following statements, n represents the
number of trials and r represents the number of events:
proc logistic;
model r/n=x1 x2;
run;
Getting Started: LOGISTIC Procedure F 5383
The following example illustrates the use of PROC LOGISTIC. The data, taken from Cox and Snell (1989, pp.
10–11), consist of the number, r, of ingots not ready for rolling, out of n tested, for a number of combinations
of heating time and soaking time.
data ingots;
input Heat Soak r n @@;
datalines;
7 1.0 0 10 14 1.0 0 31 27 1.0 1 56 51 1.0 3 13
7 1.7 0 17 14 1.7 0 43 27 1.7 4 44 51 1.7 0 1
7 2.2 0 7 14 2.2 2 33 27 2.2 0 21 51 2.2 0 1
7 2.8 0 12 14 2.8 0 31 27 2.8 1 22 51 4.0 0 1
7 4.0 0 9 14 4.0 0 19 27 4.0 1 16
;
The following invocation of PROC LOGISTIC fits the binary logit model to the grouped data. The continuous
covariates Heat and Soak are specified as predictors, and the bar notation (“|”) includes their interaction,
Heat*Soak. The ODDSRATIO statement produces odds ratios in the presence of interactions, and a graphical
display of the requested odds ratios is produced when ODS Graphics is enabled.
Model Information
Data Set WORK.INGOTS
Response Variable (Events) r
Response Variable (Trials) n
Model binary logit
Optimization Technique Fisher's scoring
The “Response Profile” table (Figure 73.2) lists the response categories (which are Event and Nonevent when
grouped data are input), their ordered values, and their total frequencies for the given data.
5384 F Chapter 73: The LOGISTIC Procedure
Response Profile
Ordered Binary Total
Value Outcome Frequency
1 Event 12
2 Nonevent 375
The “Model Fit Statistics” table (Figure 73.3) contains Akaike’s information criterion (AIC), the Schwarz
criterion (SC), and the negative of twice the log likelihood (–2 Log L) for the intercept-only model and the
fitted model. AIC and SC can be used to compare different models, and the ones with smaller values are
preferred. Results of the likelihood ratio test and the efficient score test for testing the joint significance of the
explanatory variables (Soak, Heat, and their interaction) are included in the “Testing Global Null Hypothesis:
BETA=0” table (Figure 73.3); the small p-values reject the hypothesis that all slope parameters are equal to
zero.
Figure 73.3 Fit Statistics and Hypothesis Tests
The “Analysis of Maximum Likelihood Estimates” table in Figure 73.4 lists the parameter estimates, their
standard errors, and the results of the Wald test for individual parameters. Note that the Heat*Soak parameter
is not significantly different from zero (p=0.727), nor is the Soak variable (p=0.6916).
The “Association of Predicted Probabilities and Observed Responses” table (Figure 73.5) contains four
measures of association for assessing the predictive ability of a model. They are based on the number of pairs
of observations with different response values, the number of concordant pairs, and the number of discordant
pairs, which are also displayed. Formulas for these statistics are given in the section “Rank Correlation of
Observed Responses and Predicted Probabilities” on page 5473.
The ODDSRATIO statement produces the “Odds Ratio Estimates and Wald Confidence Intervals” table
(Figure 73.6), and a graphical display of these estimates is shown in Figure 73.7. The differences between
the odds ratios are small compared to the variability shown by their confidence intervals, which confirms the
previous conclusion that the Heat*Soak parameter is not significantly different from zero.
Because the Heat*Soak interaction is nonsignificant, the following statements fit a main-effects model:
The “Analysis of Maximum Likelihood Estimates” table in Figure 73.9 again shows that the Soak parameter
is not significantly different from zero (p=0.8639). The odds ratio for each effect parameter, estimated
by exponentiating the corresponding parameter estimate, is shown in the “Odds Ratios Estimates” table
(Figure 73.9), along with 95% Wald confidence intervals. The confidence interval for the Soak parameter
contains the value 1, which also indicates that this effect is not significant.
Using these parameter estimates, you can calculate the estimated logit of as
/ D
For example, if Heat=7 and Soak=1, then logit.b 4:9284. Using this logit estimate, you can calculate
as follows:
b
This gives the predicted probability of the event (ingot not ready for rolling) for Heat=7 and Soak=1. Note that
PROC LOGISTIC can calculate these statistics for you; use the OUTPUT statement with the PREDICTED=
option, or use the SCORE statement.
To illustrate the use of an alternative form of input data, the following program creates the ingots data set
with the new variables NotReady and Freq instead of n and r. The variable NotReady represents the response
of individual units; it has a value of 1 for units not ready for rolling (event) and a value of 0 for units ready
for rolling (nonevent). The variable Freq represents the frequency of occurrence of each combination of
Heat, Soak, and NotReady. Note that, compared to the previous data set, NotReady=1 implies Freq=r, and
NotReady=0 implies Freq=n–r.
data ingots;
input Heat Soak NotReady Freq @@;
datalines;
7 1.0 0 10 14 1.0 0 31 14 4.0 0 19 27 2.2 0 21 51 1.0 1 3
7 1.7 0 17 14 1.7 0 43 27 1.0 1 1 27 2.8 1 1 51 1.0 0 10
7 2.2 0 7 14 2.2 1 2 27 1.0 0 55 27 2.8 0 21 51 1.7 0 1
7 2.8 0 12 14 2.2 0 31 27 1.7 1 4 27 4.0 1 1 51 2.2 0 1
7 4.0 0 9 14 2.8 0 31 27 1.7 0 40 27 4.0 0 15 51 4.0 0 1
;
The following statements invoke PROC LOGISTIC to fit the main-effects model by using the alternative
form of the input data set:
Response Profile
Ordered Total
Value NotReady Frequency
1 0 375
2 1 12
Probability modeled is NotReady=1.
By default, Ordered Values are assigned to the sorted response values in ascending order, and PROC
LOGISTIC models the probability of the response level that corresponds to the Ordered Value 1. There
are several methods to change these defaults; the preceding statements specify the response variable option
EVENT= to model the probability of NotReady=1 as displayed in Figure 73.10. For more information, see
the section “Response Level Ordering” on page 5457.
Syntax: LOGISTIC Procedure F 5389
The PROC LOGISTIC statement invokes the LOGISTIC procedure. Optionally, it identifies input and output
data sets, suppresses the display of results, and controls the ordering of the response levels. Table 73.1
summarizes the options available in the PROC LOGISTIC statement.
Option Description
Input/Output Data Set Options
COVOUT Displays the estimated covariance matrix in the OUTEST= data set
DATA= Names the input SAS data set
INEST= Specifies the initial estimates SAS data set
INMODEL= Specifies the model information SAS data set
NOCOV Does not save covariance matrix in the OUTMODEL= data set
OUTDESIGN= Specifies the design matrix output SAS data set
OUTDESIGNONLY Outputs the design matrix only
OUTEST= Specifies the parameter estimates output SAS data set
OUTMODEL= Specifies the model output data set for scoring
Response and CLASS Variable Options
DESCENDING Reverses the sort order of the response variable
NAMELEN= Specifies the maximum length of effect names
ORDER= Specifies the sort order of the response variable
TRUNCATE Truncates class level names
Displayed Output Options
ALPHA= Specifies the significance level for confidence intervals
NOPRINT Suppresses all displayed output
PLOTS Specifies options for plots
SIMPLE Displays descriptive statistics
Large Data Set Option
MULTIPASS Does not copy the input SAS data set for internal computations
Control of Other Statement Options
EXACTONLY Performs exact analysis only
EXACTOPTIONS Specifies global options for EXACT statements
ROCOPTIONS Specifies global options for ROC statements
PROC LOGISTIC Statement F 5391
ALPHA=number
specifies the level of significance ˛ for 100.1 ˛/% confidence intervals. The value number must be
between 0 and 1; the default value is 0.05, which results in 95% intervals. This value is used as the
default confidence level for limits computed by the following options:
Statement Options
CONTRAST ESTIMATE=
EXACT ESTIMATE=
MODEL CLODDS= CLPARM=
ODDSRATIO CL=
OUTPUT LOWER= UPPER=
PROC LOGISTIC PLOTS=EFFECT(CLBAR CLBAND)
ROCCONTRAST ESTIMATE=
SCORE CLM
You can override the default in most of these cases by specifying the ALPHA= option in the separate
statements.
COVOUT
adds the estimated covariance matrix to the OUTEST= data set. For the COVOUT option to have
an effect, the OUTEST= option must be specified. See the section “OUTEST= Output Data Set” on
page 5499 for more information.
DATA=SAS-data-set
names the SAS data set containing the data to be analyzed. If you omit the DATA= option, the
procedure uses the most recently created SAS data set. The INMODEL= option cannot be specified
with this option.
DESCENDING
DESC
reverses the sort order for the levels of the response variable. If both the DESCENDING and ORDER=
options are specified, PROC LOGISTIC orders the levels according to the ORDER= option and then
reverses that order. This option has the same effect as the response variable option DESCENDING in
the MODEL statement. See the section “Response Level Ordering” on page 5457 for more detail.
EXACTONLY
requests only the exact analyses. The asymptotic analysis that PROC LOGISTIC usually performs is
suppressed.
5392 F Chapter 73: The LOGISTIC Procedure
EXACTOPTIONS (options)
specifies options that apply to every EXACT statement in the program. The available options are
summarized here, and full descriptions are available in the EXACTOPTIONS statement.
Option Description
ABSFCONV Specifies the absolute function convergence criterion
ADDTOBS Adds the observed sufficient statistic to the sampled exact distribution
BUILDSUBSETS Builds every distribution for sampling
EPSILON= Specifies the comparison fuzz for partial sums of sufficient statistics
FCONV Specifies the relative function convergence criterion
MAXTIME= Specifies the maximum time allowed in seconds
METHOD= Specifies the DIRECT, NETWORK, NETWORKMC, or MCMC algorithm
N= Specifies the number of Monte Carlo samples
ONDISK Uses disk space
SEED= Specifies the initial seed for sampling
STATUSN= Specifies the sampling interval for printing a status line
STATUSTIME= Specifies the time interval for printing a status line
XCONV Specifies the relative parameter convergence criterion
INEST=SAS-data-set
names the SAS data set that contains initial estimates for all the parameters in the model. If BY-group
processing is used, it must be accommodated in setting up the INEST= data set. See the section
“INEST= Input Data Set” on page 5501 for more information.
INMODEL=SAS-data-set
specifies the name of the SAS data set that contains the model information needed for scoring new
data. This INMODEL= data set is the OUTMODEL= data set saved in a previous PROC LOGISTIC
call. The OUTMODEL= data set should not be modified before its use as an INMODEL= data set.
The DATA= option cannot be specified with this option; instead, specify the data sets to be scored
in the SCORE statements. FORMAT statements are not allowed when the INMODEL= data set is
specified; variables in the DATA= and PRIOR= data sets in the SCORE statement should be formatted
within the data sets.
You can specify the BY statement provided that the INMODEL= data set is created under the same
BY-group processing.
The CLASS, EFFECT, EFFECTPLOT, ESTIMATE, EXACT, LSMEANS, LSMESTIMATE, MODEL,
OUTPUT, ROC, ROCCONTRAST, SLICE, STORE, TEST, and UNIT statements are not available
with the INMODEL= option.
MULTIPASS
forces the procedure to reread the DATA= data set as needed rather than require its storage in memory
or in a temporary file on disk. By default, the data set is cleaned up and stored in memory or in a
temporary file. This option can be useful for large data sets. All exact analyses are ignored in the
presence of the MULTIPASS option. If a STRATA statement is specified, then the data set must first
be grouped or sorted by the strata variables.
PROC LOGISTIC Statement F 5393
NAMELEN=number
specifies the maximum length of effect names in tables and output data sets to be number characters,
where number is a value between 20 and 200. The default length is 20 characters.
NOCOV
specifies that the covariance matrix not be saved in the OUTMODEL= data set. The covariance matrix
is needed for computing the confidence intervals for the posterior probabilities in the OUT= data set in
the SCORE statement. Specifying this option will reduce the size of the OUTMODEL= data set.
NOPRINT
suppresses all displayed output. Note that this option temporarily disables the Output Delivery System
(ODS); see Chapter 20, “Using the Output Delivery System,” for more information.
OUTDESIGN=SAS-data-set
specifies the name of the data set that contains the design matrix for the model. The data set contains
the same number of observations as the corresponding DATA= data set and includes the response
variable (with the same format as in the DATA= data set), the FREQ variable, the WEIGHT variable,
the OFFSET= variable, and the design variables for the covariates, including the Intercept variable of
constant value 1 unless the NOINT option in the MODEL statement is specified.
OUTDESIGNONLY
suppresses the model fitting and creates only the OUTDESIGN= data set. This option is ignored if the
OUTDESIGN= option is not specified.
OUTEST=SAS-data-set
creates an output SAS data set that contains the final parameter estimates and, optionally, their estimated
covariances (see the preceding COVOUT option). The output data set also includes a variable named
_LNLIKE_, which contains the log likelihood. See the section “OUTEST= Output Data Set” on
page 5499 for more information.
OUTMODEL=SAS-data-set
specifies the name of the SAS data set that contains the information about the fitted model. This data
set contains sufficient information to score new data without having to refit the model. It is solely used
as the input to the INMODEL= option in a subsequent PROC LOGISTIC call. The OUTMODEL=
option is not available with the STRATA statement. Information in this data set is stored in a very
compact form, so you should not modify it manually.
N OTE : The STORE statement can also be used to save your model. See the section “STORE Statement”
on page 5452 for more information.
5394 F Chapter 73: The LOGISTIC Procedure
PLOTS < (global-plot-options) > < =plot-request < (options) > >
PLOTS < (global-plot-options) > =(plot-request < (options) > < . . . plot-request < (options) > >)
controls the plots produced through ODS Graphics. When you specify only one plot-request , you can
omit the parentheses from around the plot-request . For example:
PLOTS = ALL
PLOTS = (ROC EFFECT INFLUENCE(UNPACK))
PLOTS(ONLY) = EFFECT(CLBAR SHOWOBS)
ODS Graphics must be enabled before plots can be requested. For example:
For more information about enabling and disabling ODS Graphics, see the section “Enabling and
Disabling ODS Graphics” on page 607 in Chapter 21, “Statistical Graphics Using ODS.”
If the PLOTS option is not specified or is specified with no plot-requests, then graphics are produced
by default in the following situations:
If the INFLUENCE or IPLOTS option is specified in the MODEL statement, then the INFLU-
ENCE plots are produced unless the MAXPOINTS= cutoff is exceeded.
If you specify the OUTROC= option in the MODEL statement, then ROC curves are produced.
If you also specify a SELECTION= method, then an overlaid plot of all the ROC curves for each
step of the selection process is displayed.
If the OUTROC= option is specified in a SCORE statement, then the ROC curve for the scored
data set is displayed.
If you specify ROC statements, then an overlaid plot of the ROC curves for the model (or the
selected model if a SELECTION= method is specified) and for all the ROC statement models is
displayed.
If you specify the CLODDS= option in the MODEL statement or if you specify the ODDSRATIO
statement, then a plot of the odds ratios and their confidence limits is displayed.
For general information about ODS Graphics, see Chapter 21, “Statistical Graphics Using ODS.”
The following global-plot-options are available:
LABEL
displays a label on diagnostic plots to aid in identifying the outlying observations. This option
enhances the plots produced by the DFBETAS, DPC, INFLUENCE, LEVERAGE, and PHAT
options. If an ID statement is specified, then the plots are labeled with the ID variables. Otherwise,
the observation number is displayed.
PROC LOGISTIC Statement F 5395
MAXPOINTS=NONE | number
suppresses the plots produced by the DFBETAS, DPC, INFLUENCE, LEVERAGE, and PHAT
options if there are more than number observations. Also, observations are not displayed on the
EFFECT plots when the cutoff is exceeded. The default is MAXPOINTS=5000. The cutoff is
ignored if you specify MAXPOINTS=NONE.
ONLY
specifically requested plot-requests are displayed.
UNPACKPANELS | UNPACK
suppresses paneling. By default, multiple plots can appear in some output panels. Specify
UNPACKPANEL to display each plot separately.
ALL
produces all appropriate plots. You can specify other options with ALL. For example, to display
all plots and unpack the DFBETAS plots you can specify plots=(all dfbetas(unpack)).
MAXSIZE=Smax
specifies the maximum size when TYPE=BUBBLE or TYPE=LABEL. For
TYPE=BUBBLE, the size is the bubble radius and MAXSIZE=21 by default; for
TYPE=LABEL, the size is the font size and MAXSIZE=20 by default. This dpc-option is
ignored if TYPE=GRADIENT.
MAXVALUE=Cmax
displays all observations for which C Cmax at the value of the MAXSIZE= option when
TYPE=BUBBLE or TYPE=LABEL. By default, Cmax =maxi .Ci /. This dpc-option is
ignored if TYPE=GRADIENT.
MINSIZE=Smin
specifies the minimum size when TYPE=BUBBLE or TYPE=LABEL. Any observation
that maps to a smaller size is displayed at this size. For TYPE=BUBBLE, the size is the
bubble radius and MINSIZE=3.5 by default; for TYPE=LABEL, the size is the font size and
MINSIZE=2 by default. This dpc-option is ignored if TYPE=GRADIENT.
5396 F Chapter 73: The LOGISTIC Procedure
BUBBLE displays circular markers whose areas are proportional to C and whose colors
are determined by their response.
GRADIENT colors the markers according to the value of C.
LABEL displays the ID variables (if an ID statement is specified) or the observation
number. The colors of the ID variable or observation numbers are determined
by their response, and their font sizes are proportional to maxCi.C / .
i i
By default, TYPE=GRADIENT.
UNPACKPANELS | UNPACK
displays the plots separately.
NONE
suppresses all plots.
_ROC_ENTRY_ID sets the note for the ID= option on the ROC plot.
_ROC_ENTRYTITLE sets the first title line on the ROC plot.
_ROC_ENTRYTITLE2 sets the second title line on the ROC plot.
_ROC_XAXISOPTS_LABEL sets the X-axis label on the ROC and overlaid ROC plots.
_ROC_YAXISOPTS_LABEL sets the Y-axis label on the ROC and overlaid ROC plots.
_ROCOVERLAY_ENTRYTITLE sets the title on the overlaid ROC plot.
To revert to the default labels and titles, you can specify the macro variables in a %SYMDEL
statement. For example:
See Output 73.7.3 and Example 73.8 for examples of these ROC plots.
ROCOPTIONS (options)
specifies options that apply to every model specified in a ROC statement. Some of these options also
apply to the SCORE statement. The following options are available:
ALPHA=number
sets the significance level for creating confidence limits of the areas and the pairwise differences.
The ALPHA= value specified in the PROC LOGISTIC statement is the default. If neither
ALPHA= value is specified, then ALPHA=0.05 by default.
CROSSVALIDATE | X
uses cross validated predicted probabilities instead of the model-predicted probabilities for
all ROC and area under the ROC curve (AUC) computations; for more information, see the
section “Classification Table” on page 5475. The cross validated probabilities are also used in
computations for the “Association of Predicted Probabilities and Observed Responses” table.
If you use a SCORE statement, then the OUTROC= data set and the AUC statistic from the
FITSTAT option use the cross validated probabilities only when you score the original data set;
otherwise, the model-predicted probabilities are used.
5398 F Chapter 73: The LOGISTIC Procedure
EPS=value
is an alias for the ROCEPS= option in the MODEL statement. This value is used to determine
which predicted probabilities are equal. The default value is the square root of the machine
epsilon, which is about 1E–8.
The SENSIT, 1MSPEC, FALPOS, and FALNEG statistics are defined in the section “Receiver
Operating Characteristic Curves” on page 5480. The misclassification rate is the number of
events that are predicted as nonevents and the number of nonevents that are predicted as events
as calculated by using the given cutpoint (predicted probability) divided by the number of
observations. If the PEVENT= option is also specified, then FALPOS and FALNEG are computed
using the first PEVENT= value and Bayes’ theorem, as discussed in the section “False Positive,
False Negative, and Correct Classification Rates Using Bayes’ Theorem” on page 5476.
NODETAILS
suppresses the display of the model fitting information for the models specified in the ROC
statements.
OUT=SAS-data-set-name
is an alias for the OUTROC= option in the MODEL statement.
WEIGHTED
uses frequencyweight in the ROC computations (Izrael et al. 2002) instead of just frequency.
Typically, weights are considered in the fit of the model only, and hence are accounted for in the
parameter estimates. The “Association of Predicted Probabilities and Observed Responses” table
uses frequency (unless the BINWIDTH=0 option is also specified on the MODEL statement), and
is suppressed when ROC comparisons are performed. This option also affects SCORE statement
ROC and area under the ROC curve (AUC) computations.
PROC LOGISTIC Statement F 5399
SIMPLE
displays simple descriptive statistics (mean, standard deviation, minimum and maximum) for each
continuous explanatory variable. For each CLASS variable involved in the modeling, the frequency
counts of the classification levels are displayed. The SIMPLE option generates a breakdown of the
simple descriptive statistics or frequency counts for the entire data set and also for individual response
categories.
TRUNCATE
determines class levels by using no more than the first 16 characters of the formatted values of CLASS,
response, and strata variables. When formatted values are longer than 16 characters, you can use this
option to revert to the levels as determined in releases previous to SAS 9.0. This option invokes the
same option in the CLASS statement.
PLOTS=EFFECT Plots
Only one PLOTS=EFFECT plot is produced by default; you must specify other effect-options to produce
multiple plots. For binary response models, the following plots are produced when an EFFECT option is
specified with no effect-options:
If you only have continuous covariates in the model, then a plot of the predicted probability versus
the first continuous covariate fixing all other continuous covariates at their means is displayed. See
Output 73.7.4 for an example with one continuous covariate.
If you only have classification covariates in the model, then a plot of the predicted probability versus
the first CLASS covariate at each level of the second CLASS covariate, if any, holding all other CLASS
covariates at their reference levels is displayed.
If you have CLASS and continuous covariates, then a plot of the predicted probability versus the first
continuous covariate at up to 10 cross-classifications of the CLASS covariate levels, while fixing all
other continuous covariates at their means and all other CLASS covariates at their reference levels, is
displayed. For example, if your model has four binary covariates, there are 16 cross-classifications of
the CLASS covariate levels. The plot displays the 8 cross-classifications of the levels of the first three
covariates while the fourth covariate is fixed at its reference level.
For polytomous response models, similar plots are produced by default, except that the response levels are
used in place of the CLASS covariate levels. Plots for polytomous response models involving OFFSET=
variables with multiple values are not available.
The following effect-options specify the type of graphic to produce:
does not appear in the plot. When X does not define an axis it first produces plots setting X = 7 and
then produces plots setting X = 9. Note in this example that specifying AT( A=ALL ) is the same as
specifying the PLOTBY=A option.
FITOBSONLY
computes the predicted values only at the observed data. If the FITOBSONLY option is omitted and the
X-axis variable is continuous, the predicted values are computed at a grid of points extending slightly
beyond the range of the data (see the EXTEND= option for more information). If the FITOBSONLY
option is omitted and the X-axis effect is categorical, the predicted values are computed at all possible
categories.
INDIVIDUAL
displays the individual probabilities instead of the cumulative probabilities. This option is available
only with cumulative models, and it is not available with the LINK option.
LINK
displays the linear predictors instead of the probabilities on the Y axis. For example, for a binary
logistic regression, the Y axis will be displayed on the logit scale. The INDIVIDUAL and POLYBAR
options are not available with the LINK option.
PLOTBY=effect
displays an effect plot at each unique level of the PLOTBY= effect. You can specify effect as one
CLASS variable or as an interaction of classification covariates. For polytomous-response models, you
can also specify the response variable as the lone PLOTBY= effect. For nonsingular parameterizations,
the complete cross-classification of the CLASS variables specified in the effect define the different
PLOTBY= levels. When the GLM parameterization is used, the PLOTBY= levels can depend on the
model and the data.
SLICEBY=effect
displays predicted probabilities at each unique level of the SLICEBY= effect. You can specify effect
as one CLASS variable or as an interaction of classification covariates. For polytomous-response
models, you can also specify the response variable as the lone SLICEBY= effect. For nonsingular
parameterizations, the complete cross-classification of the CLASS variables specified in the effect
define the different SLICEBY= levels. When the GLM parameterization is used, the SLICEBY= levels
can depend on the model and the data.
X=effect
X=(effect. . . effect)
specifies effects to be used on the X axis of the effect plots. You can specify several different X axes:
continuous variables must be specified as main effects, while CLASS variables can be crossed. For
nonsingular parameterizations, the complete cross-classification of the CLASS variables specified in
the effect define the axes. When the GLM parameterization is used, the X= levels can depend on the
model and the data.
N OTE : Any variable not specified in a SLICEBY= or PLOTBY= option is available to be displayed on the X
axis. A variable can be specified in at most one of the SLICEBY=, PLOTBY=, and X= options.
The following effect-options enhance the graphical output:
PROC LOGISTIC Statement F 5401
ALPHA=number
specifies the size of the confidence limits. The ALPHA= value specified in the PROC LOGISTIC
statement is the default. If neither ALPHA= value is specified, then ALPHA=0.05 by default.
CLBAR
displays the error bars on the plots when you have CLASS covariates on the X axis; if the X axis is
continuous, then this invokes the CLBAND option. For polytomous-response models with CLASS
covariates only and with the POLYBAR option specified, the stacked bar charts are replaced by
side-by-side bar charts with error bars.
EXTEND=value
extends continuous X axes by a factor of value=2 in each direction. By default, EXTEND=0.2.
MAXATLEN=length
specifies the maximum number of characters used to display the levels of all the fixed variables. If
the text is too long, it is truncated and ellipses (“. . . ”) are appended. By default, length is equal to its
maximum allowed value, 256.
NOCLUSTER
prevents clustering of the levels of the SLICEBY= effect. This option is available when you have
CLASS covariates on the X axis.
NOCONNECT
removes the line that connects the predicted values. This option is available when you have CLASS
covariates on the X axis.
POLYBAR
replaces scatter plots of polytomous response models with bar charts. This option has no effect on
binary-response models, and it is overridden by the CONNECT option. By default, the X axis is chosen
to be a crossing of available classification variables so that there are no more than 16 levels; if no such
crossing is possible then the first available classification variable is used. You can override this default
by specifying the X= option.
5402 F Chapter 73: The LOGISTIC Procedure
DOTPLOT
displays dotted gridlines on the plot.
GROUP
displays the odds ratios in panels that are defined by the ODDSRATIO statements. The NPANELPOS=
option is ignored when this option is specified.
LOGBASE=2 | E | 10
displays the odds ratio axis on the specified log scale.
NPANELPOS=number
breaks the plot into multiple graphics that have at most |number | odds ratios per graphic. If number
is positive, then the number of odds ratios per graphic is balanced; if number is negative, then no
balancing of the number of odds ratios takes place. By default, number = 0 and all odds ratios are
displayed in a single plot. For example, suppose you want to display 21 odds ratios. Then specifying
NPANELPOS=20 displays two plots, the first with 11 odds ratios and the second with 10; but specifying
NPANELPOS=-20 displays 20 odds ratios in the first plot and only 1 odds ratio in the second plot.
BY Statement F 5403
ORDER=ASCENDING | DESCENDING
displays the odds ratios in sorted order. By default the odds ratios are displayed in the order in which
they appear in the corresponding table.
RANGE=(< min >< ,max >) | CLIP
specifies the range of the displayed odds ratio axis. Specifying the RANGE=CLIP option has the same
effect as specifying the minimum odds ratio as min and the maximum odds ratio as max . By default,
all odds ratio confidence intervals are displayed.
TYPE=HORIZONTAL | HORIZONTALSTAT | VERTICAL | VERTICALBLOCK
controls the look of the graphic. The default TYPE=HORIZONTAL option places the odds ratio values
on the X axis, while the TYPE=HORIZONTALSTAT option also displays the values of the odds ratios
and their confidence limits on the right side of the graphic. The TYPE=VERTICAL option places the
odds ratio values on the Y axis, while the TYPE=VERTICALBLOCK option (available only with the
CLODDS= option) places the odds ratio values on the Y axis and puts boxes around the labels.
BY Statement
BY variables ;
You can specify a BY statement with PROC LOGISTIC to obtain separate analyses of observations in groups
that are defined by the BY variables. When a BY statement appears, the procedure expects the input data
set to be sorted in order of the BY variables. If you specify more than one BY statement, only the last one
specified is used.
If your input data set is not sorted in ascending order, use one of the following alternatives:
Sort the data by using the SORT procedure with a similar BY statement.
Specify the NOTSORTED or DESCENDING option in the BY statement for the LOGISTIC procedure.
The NOTSORTED option does not mean that the data are unsorted but rather that the data are arranged
in groups (according to values of the BY variables) and that these groups are not necessarily in
alphabetical or increasing numeric order.
Create an index on the BY variables by using the DATASETS procedure (in Base SAS software).
If a SCORE statement is specified, then define the training data set to be the DATA= data set or the
INMODEL= data set in the PROC LOGISTIC statement, and define the scoring data set to be the DATA=
data set and PRIOR= data set in the SCORE statement. The training data set contains all of the BY variables,
and the scoring data set must contain either all of them or none of them. If the scoring data set contains all
the BY variables, matching is carried out between the training and scoring data sets. If the scoring data set
does not contain any of the BY variables, the entire scoring data set is used for every BY group in the training
data set and the BY variables are added to the output data sets that are specified in the SCORE statement.
C AUTION : The order of the levels in the response and classification variables is determined from all the data
regardless of BY groups. However, different sets of levels might appear in different BY groups. This might
affect the value of the reference level for these variables, and hence your interpretation of the model and the
parameters.
For more information about BY-group processing, see the discussion in SAS Language Reference: Concepts.
For more information about the DATASETS procedure, see the discussion in the Base SAS Procedures Guide.
5404 F Chapter 73: The LOGISTIC Procedure
CLASS Statement
CLASS variable < (options) > . . . < variable < (options) > > < / global-options > ;
The CLASS statement names the classification variables to be used as explanatory variables in the analysis.
Response variables do not need to be specified in the CLASS statement.
The CLASS statement must precede the MODEL statement. Most options can be specified either as individual
variable options or as global-options. You can specify options for each variable by enclosing the options in
parentheses after the variable name. You can also specify global-options for the CLASS statement by placing
them after a slash (/). Global-options are applied to all the variables specified in the CLASS statement. If you
specify more than one CLASS statement, the global-options specified in any one CLASS statement apply to
all CLASS statements. However, individual CLASS variable options override the global-options. You can
specify the following values for either an option or a global-option:
CPREFIX=n
specifies that, at most, the first n characters of a CLASS variable name be used in creating names for
the corresponding design variables. The default is 32 min.32; max.2; f //, where f is the formatted
length of the CLASS variable.
DESCENDING
DESC
reverses the sort order of the classification variable. If both the DESCENDING and ORDER= options
are specified, PROC LOGISTIC orders the categories according to the ORDER= option and then
reverses that order.
LPREFIX=n
specifies that, at most, the first n characters of a CLASS variable label be used in creating labels for the
corresponding design variables. The default is 256 min.256; max.2; f //, where f is the formatted
length of the CLASS variable.
MISSING
treats missing values (., ._, .A, . . . , .Z for numeric variables and blanks for character variables) as valid
values for the CLASS variable.
For more information about sort order, see the chapter on the SORT procedure in the Base SAS
Procedures Guide and the discussion of BY-group processing in SAS Language Reference: Concepts.
PARAM=keyword
specifies the parameterization method for the classification variable or variables. You can specify any
of the keywords shown in the following table;
the default is PARAM=EFFECT. Design matrix columns are created from CLASS variables according
to the corresponding coding schemes:
All parameterizations are full rank, except for the GLM parameterization. The REF= option in the
CLASS statement determines the reference level for EFFECT and REFERENCE coding and for their
orthogonal parameterizations. It also indirectly determines the reference level for a singular GLM
parameterization through the order of levels.
If PARAM=ORTHPOLY or PARAM=POLY and the classification variable is numeric, then the
ORDER= option in the CLASS statement is ignored, and the internal unformatted values are used. See
the section “Other Parameterizations” on page 389 in Chapter 19, “Shared Concepts and Topics,” for
further details.
5406 F Chapter 73: The LOGISTIC Procedure
REF=’level’ | keyword
specifies the reference level for PARAM=EFFECT, PARAM=REFERENCE, and their orthogonaliza-
tions. For PARAM=GLM, the REF= option specifies a level of the classification variable to be put at
the end of the list of levels. This level thus corresponds to the reference level in the usual interpretation
of the linear estimates with a singular parameterization.
For an individual variable REF= option (but not for a global REF= option), you can specify the level
of the variable to use as the reference level. Specify the formatted value of the variable if a format is
assigned. For a global or individual variable REF= option, you can use one of the following keywords.
The default is REF=LAST.
TRUNCATE< =n >
specifies the length n of CLASS variable values to use in determining CLASS variable levels. The
default is to use the full formatted length of the CLASS variable. If you specify TRUNCATE without
the length n, the first 16 characters of the formatted values are used. When formatted values are longer
than 16 characters, you can use this option to revert to the levels as determined in releases before SAS
9. The TRUNCATE option is available only as a global option.
Design Matrix
A(B=1) A(B=2)
B A A1 A2 A1 A2
1 1 1 0 0 0
1 2 0 1 0 0
1 3 –1 –1 0 0
2 1 0 0 1 0
2 2 0 0 0 1
CODE Statement F 5407
PROC LOGISTIC detects linear dependency among the last two design variables and sets the parameter for
A2(B=2) to zero, resulting in an interpretation of these parameters as if they were reference- or dummy-coded.
The REFERENCE or GLM parameterization might be more appropriate for such problems.
CODE Statement
CODE < options > ;
The CODE statement writes SAS DATA step code for computing predicted values of the fitted model either
to a file or to a catalog entry. This code can then be included in a DATA step to score new data.
Table 73.2 summarizes the options available in the CODE statement.
Option Description
CATALOG= Names the catalog entry where the generated code is saved
DUMMIES Retains the dummy variables in the data set
ERROR Computes the error function
FILE= Names the file where the generated code is saved
FORMAT= Specifies the numeric format for the regression coefficients
GROUP= Specifies the group identifier for array names and statement labels
IMPUTE Imputes predicted values for observations with missing or invalid
covariates
LINESIZE= Specifies the line size of the generated code
LOOKUP= Specifies the algorithm for looking up CLASS levels
RESIDUAL Computes residuals
For details about the syntax of the CODE statement, see the section “CODE Statement” on page 393 in
Chapter 19, “Shared Concepts and Topics.”
CONTRAST Statement
CONTRAST 'label' row-description< , . . . , row-description > < / options > ;
where a row-description is defined as follows:
effect values< , . . . , effect values >
The CONTRAST statement provides a mechanism for obtaining customized hypothesis tests. It is similar to
the CONTRAST and ESTIMATE statements in other modeling procedures.
The CONTRAST statement enables you to specify a matrix, L, for testing the hypothesis Lˇ D 0, where
ˇ is the vector of intercept and slope parameters. You must be familiar with the details of the model
parameterization that PROC LOGISTIC uses (for more information, see the PARAM= option in the section
“CLASS Statement” on page 5404). Optionally, the CONTRAST statement enables you to estimate each
row, li0 ˇ, of Lˇ and test the hypothesis li0 ˇ D 0. Computed statistics are based on the asymptotic chi-square
distribution of the Wald statistic.
5408 F Chapter 73: The LOGISTIC Procedure
There is no limit to the number of CONTRAST statements that you can specify, but they must appear after
the MODEL statement.
The following parameters are specified in the CONTRAST statement:
label identifies the contrast in the displayed output. A label is required for every contrast specified, and
it must be enclosed in quotation marks.
effect identifies an effect that appears in the MODEL statement. The name INTERCEPT can be used as
an effect when one or more intercepts are included in the model. You do not need to include all
effects that are included in the MODEL statement.
values are constants that are elements of the L matrix associated with the effect. To correctly specify
your contrast, it is crucial to know the ordering of parameters within each effect and the variable
levels associated with any parameter. The “Class Level Information” table shows the ordering
of levels within variables. The E option, described later in this section, enables you to verify the
proper correspondence of values to parameters. If too many values are specified for an effect, the
extra ones are ignored. If too few values are specified, the remaining ones are set to 0.
Multiple degree-of-freedom hypotheses can be tested by specifying multiple row-descriptions; the rows of
L are specified in order and are separated by commas. The degrees of freedom is the number of linearly
independent constraints implied by the CONTRAST statement—that is, the rank of L.
More details for specifying contrasts involving effects with full-rank parameterizations are given in the
section “Full-Rank Parameterized Effects” on page 5409, while details for less-than-full-rank parameterized
effects are given in the section “Less-Than-Full-Rank Parameterized Effects” on page 5410.
You can specify the following options after a slash (/):
ALPHA=number
specifies the level of significance ˛ for the 100.1 ˛/% confidence interval for each contrast when the
ESTIMATE option is specified. The value of number must be between 0 and 1. By default, number is
equal to the value of the ALPHA= option in the PROC LOGISTIC statement, or 0.05 if that option is
not specified.
E
displays the L matrix.
ESTIMATE=keyword
estimates and tests each individual contrast (that is, each row, li0 ˇ, of Lˇ), exponentiated contrast
0
(e li ˇ ), or predicted probability for the contrast (g 1 .li0 ˇ/). PROC LOGISTIC displays the point
estimate, its standard error, a Wald confidence interval, and a Wald chi-square test. The significance
level of the confidence interval is controlled by the ALPHA= option. You can estimate the individual
contrast, the exponentiated contrast, or the predicted probability for the contrast by specifying one of
the following keywords:
For more information about the computations of the standard errors and confidence limits, see the
section “Linear Predictor, Predicted Probability, and Confidence Limits” on page 5474.
SINGULAR=number
tunes the estimability check. This option is ignored when a full-rank parameterization is specified. If
v is a vector, define ABS.v/ to be the largest absolute value of the elements of v. For a row vector
l 0 of the contrast matrix L, define c D ABS.l / if ABS.l / is greater than 0; otherwise, c = 1. If
ABS.l 0 l 0 T/ is greater than cnumber , then l is declared nonestimable. The T matrix is the Hermite
form matrix I0 I0 , where I0 represents a generalized inverse of the (observed or expected) information
matrix I0 of the null model. The value for number must be between 0 and 1; the default value is 1E–4.
EFFECT Statement
EFFECT name=effect-type (variables < / options >) ;
The EFFECT statement enables you to construct special collections of columns for design matrices. These
collections are referred to as constructed effects to distinguish them from the usual model effects that are
formed from continuous or classification variables, as discussed in the section “GLM Parameterization of
Classification Variables and Effects” on page 385 in Chapter 19, “Shared Concepts and Topics.”
You can specify the following effect-types:
COLLECTION specifies a collection effect that defines one or more variables as a single
effect with multiple degrees of freedom. The variables in a collection are
considered as a unit for estimation and inference.
LAG specifies a classification effect in which the level that is used for a particular
period corresponds to the level in the preceding period.
MULTIMEMBER | MM specifies a multimember classification effect whose levels are determined by
one or more variables that appear in a CLASS statement.
POLYNOMIAL | POLY specifies a multivariate polynomial effect in the specified numeric variables.
SPLINE specifies a regression spline effect whose columns are univariate spline ex-
pansions of one or more variables. A spline expansion replaces the original
variable with an expanded or larger set of new variables.
EFFECT Statement F 5411
Option Description
Collection Effects Options
DETAILS Displays the constituents of the collection effect
Lag Effects Options
DESIGNROLE= Names a variable that controls to which lag design an observation
is assigned
DETAILS Displays the lag design of the lag effect
NLAG= Specifies the number of periods in the lag
PERIOD= Names the variable that defines the period. This option is required.
WITHIN= Names the variable or variables that define the group within which
each period is defined. This option is required.
Multimember Effects Options
NOEFFECT Specifies that observations with all missing levels for the
multimember variables should have zero values in the
corresponding design matrix columns
WEIGHT= Specifies the weight variable for the contributions of each of the
classification effects
Polynomial Effects Options
DEGREE= Specifies the degree of the polynomial
MDEGREE= Specifies the maximum degree of any variable in a term of the
polynomial
STANDARDIZE= Specifies centering and scaling suboptions for the variables that
define the polynomial
Spline Effects Options
BASIS= Specifies the type of basis (B-spline basis or truncated power
function basis) for the spline effect
DEGREE= Specifies the degree of the spline effect
KNOTMETHOD= Specifies how to construct the knots for the spline effect
For more information about the syntax of these effect-types and how columns of constructed effects are
computed, see the section “EFFECT Statement” on page 395 in Chapter 19, “Shared Concepts and Topics.”
5412 F Chapter 73: The LOGISTIC Procedure
EFFECTPLOT Statement
EFFECTPLOT < plot-type < (plot-definition-options) > > < / options > ;
The EFFECTPLOT statement produces a display of the fitted model and provides options for changing and
enhancing the displays. Table 73.4 describes the available plot-types and their plot-definition-options.
For full details about the syntax and options of the EFFECTPLOT statement, see the section “EFFECTPLOT
Statement” on page 414 in Chapter 19, “Shared Concepts and Topics.”
See Outputs 73.2.11, 73.2.12, 73.3.5, 73.4.8, 73.7.4, and 73.16.4 for examples of plots produced by this
statement.
ESTIMATE Statement F 5413
ESTIMATE Statement
ESTIMATE < 'label' > estimate-specification < (divisor =n) >
< , . . . < 'label' > estimate-specification < (divisor =n) > >
< / options > ;
The ESTIMATE statement provides a mechanism for obtaining custom hypothesis tests. Estimates are
formed as linear estimable functions of the form Lˇ. You can perform hypothesis tests for the estimable
functions, construct confidence limits, and obtain specific nonlinear transformations.
Table 73.5 summarizes the options available in the ESTIMATE statement.
Option Description
Construction and Computation of Estimable Functions
DIVISOR= Specifies a list of values to divide the coefficients
NOFILL Suppresses the automatic fill-in of coefficients for higher-order
effects
SINGULAR= Tunes the estimability checking difference
Statistical Output
CL Constructs confidence limits
CORR Displays the correlation matrix of estimates
COV Displays the covariance matrix of estimates
E Prints the L matrix
JOINT Produces a joint F or chi-square test for the estimable functions
PLOTS= Requests ODS statistical graphics if the analysis is sampling-based
SEED= Specifies the seed for computations that depend on random
numbers
For details about the syntax of the ESTIMATE statement, see the section “ESTIMATE Statement” on
page 442 in Chapter 19, “Shared Concepts and Topics.”
EXACT Statement
EXACT < 'label' > < INTERCEPT > < effects > < / options > ;
The EXACT statement performs exact tests of the parameters for the specified effects and optionally estimates
the parameters and outputs the exact conditional distributions. You can specify the keyword INTERCEPT
and any effects in the MODEL statement. Inference on the parameters of the specified effects is performed
by conditioning on the sufficient statistics of all the other model parameters (possibly including the intercept).
You can specify several EXACT statements, but they must follow the MODEL statement. Each statement can
optionally include an identifying label . If several EXACT statements are specified, any statement without
a label is assigned a label of the form “Exactn,” where n indicates the nth EXACT statement. The label is
included in the headers of the displayed exact analysis tables.
If a STRATA statement is also specified, then a stratified exact logistic regression is performed. The model
contains a different intercept for each stratum, and these intercepts are conditioned out of the model along
with any other nuisance parameters (parameters for effects specified in the MODEL statement that are not in
the EXACT statement).
If the LINK=GLOGIT option is specified in the MODEL statement, then the METHOD=DIRECT option is
invoked in the EXACTOPTIONS statement by default and a generalized logit model is fit. Since each effect
specified in the MODEL statement adds k parameters to the model (where k + 1 is the number of response
levels), exact analysis of the generalized logit model by using this method is limited to rather small problems.
The CONTRAST, ESTIMATE, LSMEANS, LSMESTIMATE, ODDSRATIO, OUTPUT, ROC, ROCCON-
TRAST, SCORE, SLICE, STORE, TEST, and UNITS statements are not available with an exact analysis;
results from these statements are based on the asymptotic results. Exact analyses are not performed when you
specify a WEIGHT statement, a link other than LINK=LOGIT or LINK=GLOGIT, an offset variable, the
NOFIT option, or a model selection method. Exact estimation is not available for ordinal response models.
For classification variables, use of the reference parameterization is recommended.
The following options can be specified in each EXACT statement after a slash (/):
ALPHA=number
specifies the level of significance ˛ for 100.1 ˛/% confidence limits for the parameters or odds
ratios. The value of number must be between 0 and 1. By default, number is equal to the value of the
ALPHA= option in the PROC LOGISTIC statement, or 0.05 if that option is not specified.
CLTYPE=EXACT | MIDP
requests either the exact or mid-p confidence intervals for the parameter estimates. By default, the exact
intervals are produced. The confidence coefficient can be specified with the ALPHA= option. The
mid-p interval can be modified with the MIDPFACTOR= option. See the section “Exact Conditional
Logistic Regression” on page 5495 for details.
EXACT Statement F 5415
JOINT
performs the joint test that all of the parameters are simultaneously equal to zero, performs individual
hypothesis tests for the parameter of each continuous variable, and performs joint tests for the parame-
ters of each classification variable. The joint test is indicated in the “Conditional Exact Tests” table by
the label “Joint.”
JOINTONLY
performs only the joint test of the parameters. The test is indicated in the “Conditional Exact Tests”
table by the label “Joint.” When this option is specified, individual tests for the parameters of each
continuous variable and joint tests for the parameters of the classification variables are not performed.
MIDPFACTOR=ı1 | (ı1 ; ı2 )
sets the tie factors used to produce the mid-p hypothesis statistics and the mid-p confidence intervals.
ı1 modifies both the hypothesis tests and confidence intervals, while ı2 affects only the hypothesis
tests. By default, ı1 D 0:5 and ı2 D 1:0. See the section “Exact Conditional Logistic Regression” on
page 5495 for details.
ONESIDED
requests one-sided confidence intervals and p-values for the individual parameter estimates and odds
ratios. The one-sided p-value is the smaller of the left- and right-tail probabilities for the observed
sufficient statistic of the parameter under the null hypothesis that the parameter is zero. The two-sided
p-values (default) are twice the one-sided p-values. See the section “Exact Conditional Logistic
Regression” on page 5495 for more details.
OUTDIST=SAS-data-set
names the SAS data set that contains the exact conditional distributions. This data set contains all of
the exact conditional distributions that are required to process the corresponding EXACT statement.
This data set contains the possible sufficient statistics for the parameters of the effects specified
in the EXACT statement, the counts, and, when hypothesis tests are performed on the parameters,
the probability of occurrence and the score value for each sufficient statistic. When you request an
OUTDIST= data set, the observed sufficient statistics are displayed in the “Sufficient Statistics” table.
See the section “OUTDIST= Output Data Set” on page 5503 for more information.
values of the sufficient statistics for the intercept, x2, and x3 parameters; likewise, the test for x2 is conditional
on the observed sufficient statistics for the intercept, x1, and x3.
proc logistic;
model y= x1 x2 x3;
exact x1 x2;
run;
PROC LOGISTIC determines, from all the specified EXACT statements, the distinct conditional distributions
that need to be evaluated. For example, there is only one exact conditional distribution for the following two
EXACT statements:
EXACTOPTIONS Statement
EXACTOPTIONS options ;
The EXACTOPTIONS statement specifies options that apply to every EXACT statement in the program.
The following options are available:
ABSFCONV=value
specifies the absolute function convergence criterion. Convergence requires a small change in the
log-likelihood function in subsequent iterations,
jli li 1j < value
where li is the value of the log-likelihood function at iteration i.
By default, ABSFCONV=1E–12. You can also specify the FCONV= and XCONV= criteria; optimiza-
tions are terminated as soon as one criterion is satisfied.
ADDTOBS
adds the observed sufficient statistic to the sampled exact distribution if the statistic was not sampled.
This option has no effect unless the METHOD=NETWORKMC or METHOD=MCMC option is
specified and the ESTIMATE option is specified in the EXACT statement. If the observed statistic
has not been sampled, then the parameter estimate does not exist; by specifying this option, you can
produce (biased) estimates.
EXACTOPTIONS Statement F 5417
BUILDSUBSETS
builds every distribution for sampling. By default, some exact distributions are created by taking
a subset of a previously generated exact distribution. When the METHOD=NETWORKMC or
METHOD=MCMC option is invoked, this subsetting behavior has the effect of using fewer than the
desired n samples; see the N= option for more details. Use the BUILDSUBSETS option to suppress
this subsetting.
EPSILON=value
Pj
controls how the partial sums i D1 yi xi are compared. value must be between 0 and 1; by default,
value=1E–8.
FCONV=value
specifies the relative function convergence criterion. Convergence requires a small relative change in
the log-likelihood function in subsequent iterations,
jli li 1 j
< value
jli 1 j C 1E–6
MAXSWAP=n
MAXR=n
specifies the maximum number of swaps for each sample that PROC LOGISTIC makes when you
specify the METHOD=MCMC option. If an intercept or a stratum is conditioned out, then n swaps are
performed: one event is changed to a nonevent, and one nonevent is changed to an event. Although
you might need large values of n in order to transition between any two points in the exact distribution,
such values quickly increase computation time. By default, MAXSWAP=2.
MAXTIME=seconds
specifies the maximum clock time (in seconds) that PROC LOGISTIC can use to calculate the exact
distributions. If the limit is exceeded, the procedure halts all computations and prints a note to the SAS
log. The default maximum clock time is seven days.
METHOD=keyword
specifies which exact conditional algorithm to use for every EXACT statement specified. You can
specify one of the following keywords:
DIRECT invokes the multivariate shift algorithm of Hirji, Mehta, and Patel (1987). This method
directly builds the exact distribution, but it can require an excessive amount of memory in its
intermediate stages. METHOD=DIRECT is invoked by default when you are conditioning
out at most the intercept, or when the LINK=GLOGIT option is specified in the MODEL
statement.
NETWORK invokes an algorithm described in Mehta, Patel, and Senchaudhuri (1992). This method
builds a network for each parameter that you are conditioning out, combines the networks,
then uses the multivariate shift algorithm to create the exact distribution. The NETWORK
method can be faster and require less memory than the DIRECT method. The NETWORK
method is invoked by default for most analyses.
5418 F Chapter 73: The LOGISTIC Procedure
NETWORKMC invokes the hybrid network and Monte Carlo algorithm of Mehta, Patel, and Sen-
chaudhuri (1992). This method creates a network, then samples from that network; this
method does not reject any of the samples at the cost of using a large amount of memory
to create the network. METHOD=NETWORKMC is most useful for producing parameter
estimates for problems that are too large for the DIRECT and NETWORK methods to
handle and for which asymptotic methods are invalid—for example, for sparse data on a
large grid.
MCMC invokes the Markov chain Monte Carlo (MCMC) algorithm of Forster, McDonald, and
Smith (2003). This method uses a Metropolis-Hastings algorithm to generate samples from
the exact distribution by repeatedly perturbing the response vector to obtain a new response
vector while maintaining the sufficient statistics for the nuisance parameters. You must
also condition out the intercept or the strata. (Using notation from Zamar et al. 2007, the
current implementation samples from VA , does not explicitly sample for v = 0, allows
candidates v and –v, allows d < 0, and limits r by the number of observations.) The sampling
is divided into parallel threads, where the number of threads is the value of the SAS system
option CPUCOUNT=. The MCMC method is most useful for problems for which the
NETWORKMC method has difficulty generating the network and for which asymptotic
results are suspect; however, to make sure you are sampling from the equilibrium distribution,
you should run your program multiple times and increase the N= and MAXSWAP= values
until you believe that your results are stable. The MCMC method can take a large amount
of time, depending on the number of observations in your data set, the number of samples,
and the number of swaps.
N=n
specifies the number of Monte Carlo samples to take when you specify the METHOD=NETWORKMC
or METHOD=MCMC option. By default, n = 10,000. If PROC LOGISTIC cannot obtain n samples
because of a lack of memory, then a note is printed in the SAS log (the number of valid samples is also
reported in the listing) and the analysis continues.
The number of samples used to produce any particular statistic might be smaller than n. For example,
let X1 and X2 be continuous variables, denote their joint distribution by f (X1,X2), and let f (X1 | X2 =
x2) denote the marginal distribution of X1 conditioned on the observed value of X2. If you request
the JOINT test of X1 and X2, then n samples are used to generate the estimate fO(X1,X2) of f (X1,X2),
from which the test is computed. However, the parameter estimate for X1 is computed from the subset
of fO(X1,X2) that has X2 = x2, and this subset need not contain n samples. Similarly, the distribution
for each level of a classification variable is created by extracting the appropriate subset from the joint
distribution for the CLASS variable.
In some cases, the marginal sample size can be too small to admit accurate estimation of a particular
statistic; a note is printed in the SAS log when a marginal sample size is less than 100. Increasing n
increases the number of samples used in a marginal distribution; however, if you want to control the
sample size exactly, you can either specify the BUILDSUBSETS option or do both of the following:
NBI=n
BURNIN=n
specifies the number of burn-in samples that are discarded when you specify the METHOD=MCMC
option. By default, NBI=0.
NOLOGSCALE
specifies that computations for the exact conditional models be computed by using normal scaling.
Log scaling can handle numerically larger problems than normal scaling; however, computations in the
log scale are slower than computations in normal scale.
NTHIN=n
controls the thinning rate of the sampling when you specify the METHOD=MCMC option. Every nth
sample is kept and the rest are discarded. By default, NTHIN=1.
ONDISK
uses disk space instead of random access memory to build the exact conditional distribution. Use this
option to handle larger problems at the cost of slower processing.
SEED=seed
specifies the initial seed for the random number generator used to take the Monte Carlo samples when
you specify the METHOD=NETWORKMC or METHOD=MCMC option. The value of the SEED=
option must be an integer. If you do not specify a seed, or if you specify a value less than or equal to 0,
then PROC LOGISTIC uses the time of day from the computer’s clock to generate an initial seed.
STATUSN=number
prints a status line in the SAS log after every number of Monte Carlo samples when you specify the
METHOD=NETWORKMC or METHOD=MCMC option. When you specify METHOD=MCMC, the
actual number that is used depends on the number of threads used in the computations. The number
of samples that are taken and the current exact p-value for testing the significance of the model are
displayed. You can use this status line to track the progress of the computation of the exact conditional
distributions.
STATUSTIME=seconds
specifies the time interval (in seconds) for printing a status line in the SAS log. You can use this
status line to track the progress of the computation of the exact conditional distributions. The time
interval that you specify is approximate; the actual time interval varies. By default, no status reports
are produced.
XCONV=value
specifies the relative parameter convergence criterion. Convergence requires a small relative parameter
change in subsequent iterations,
.i /
max jıj j < value
j
where
8 .i / .i 1/ .i 1/
< ˇj ˇj jˇj j < 0:01
.i /
ıj D .i/
ˇj
.i 1/
ˇj
: .i 1/ otherwise
ˇj
.i /
and ˇj is the estimate of the jth parameter at iteration i.
5420 F Chapter 73: The LOGISTIC Procedure
By default, XCONV=1E–4. You can also specify the ABSFCONV= and FCONV= criteria; if more
than one criterion is specified, then optimizations are terminated as soon as one criterion is satisfied.
FREQ Statement
FREQ variable ;
The FREQ statement identifies a variable that contains the frequency of occurrence of each observation.
PROC LOGISTIC treats each observation as if it appears n times, where n is the value of the FREQ variable
for the observation. If it is not an integer, the frequency value is truncated to an integer. If the frequency
value is less than 1 or missing, the observation is not used in the model fitting. When the FREQ statement is
not specified, each observation is assigned a frequency of 1. If you specify more than one FREQ statement,
then the first statement is used.
If a SCORE statement is specified, then the FREQ variable is used for computing fit statistics and the ROC
curve, but they are not required for scoring. If the DATA= data set in the SCORE statement does not contain
the FREQ variable, the frequency values are assumed to be 1 and a warning message is issued in the LOG.
If you fit a model and perform the scoring in the same run, the same FREQ variable is used for fitting and
scoring. If you fit a model in a previous run and input it with the INMODEL= option in the current run, then
the FREQ variable can be different from the one used in the previous run. However, if a FREQ variable was
not specified in the previous run, you can still specify a FREQ variable in the current run.
ID Statement
ID variable< variable,. . . > ;
The ID statement specifies variables in the DATA= data set that are used for labeling ROC curves and influence
diagnostic plots. If more than one ID variable is specified, then the plots are labeled by concatenating the ID
variable values together. For more information, see the PLOTS(LABEL) and ROCOPTIONS(ID) options in
the PROC LOGISTIC statement.
LSMEANS Statement
LSMEANS < model-effects > < / options > ;
The LSMEANS statement computes and compares least squares means (LS-means) of fixed effects. LS-means
are predicted population margins—that is, they estimate the marginal means over a balanced population. In a
sense, LS-means are to unbalanced designs as class and subclass arithmetic means are to balanced designs.
Table 73.6 summarizes the options available in the LSMEANS statement.
LSMEANS Statement F 5421
Option Description
Construction and Computation of LS-Means
AT Modifies the covariate value in computing LS-means
BYLEVEL Computes separate margins
DIFF Requests differences of LS-means
OM= Specifies the weighting scheme for LS-means computation as
determined by the input data set
SINGULAR= Tunes estimability checking
Statistical Output
CL Constructs confidence limits for means and mean differences
CORR Displays the correlation matrix of LS-means
COV Displays the covariance matrix of LS-means
E Prints the L matrix
LINES Produces a “Lines” display for pairwise LS-means differences
MEANS Prints the LS-means
PLOTS= Requests graphs of means and mean comparisons
SEED= Specifies the seed for computations that depend on random
numbers
For details about the syntax of the LSMEANS statement, see the section “LSMEANS Statement” on page 458
in Chapter 19, “Shared Concepts and Topics.”
N OTE : If you have classification variables in your model, then the LSMEANS statement is allowed only if
you also specify the PARAM=GLM option.
5422 F Chapter 73: The LOGISTIC Procedure
LSMESTIMATE Statement
LSMESTIMATE model-effect < 'label' > values < divisor =n >
< , . . . < 'label' > values < divisor =n > >
< / options > ;
The LSMESTIMATE statement provides a mechanism for obtaining custom hypothesis tests among least
squares means.
Table 73.7 summarizes the options available in the LSMESTIMATE statement.
Option Description
Construction and Computation of LS-Means
AT Modifies covariate values in computing LS-means
BYLEVEL Computes separate margins
DIVISOR= Specifies a list of values to divide the coefficients
OM= Specifies the weighting scheme for LS-means computation as
determined by a data set
SINGULAR= Tunes estimability checking
Statistical Output
CL Constructs confidence limits for means and mean differences
CORR Displays the correlation matrix of LS-means
COV Displays the covariance matrix of LS-means
E Prints the L matrix
ELSM Prints the K matrix
JOINT Produces a joint F or chi-square test for the LS-means and
LS-means differences
PLOTS= Requests graphs of means and mean comparisons
SEED= Specifies the seed for computations that depend on random
numbers
Option Description
ILINK Computes and displays estimates and standard errors of LS-means
(but not differences) on the inverse linked scale
For details about the syntax of the LSMESTIMATE statement, see the section “LSMESTIMATE Statement”
on page 477 in Chapter 19, “Shared Concepts and Topics.”
N OTE : If you have classification variables in your model, then the LSMESTIMATE statement is allowed
only if you also specify the PARAM=GLM option.
MODEL Statement
< label: > MODEL variable< (variable_options) > = < effects > < / options > ;
< label: > MODEL events/trials = < effects > < / options > ;
The MODEL statement names the response variable and the explanatory effects, including covariates, main
effects, interactions, and nested effects; see the section “Specification of Effects” on page 3670 in Chapter 47,
“The GLM Procedure,” for more information. If you omit the explanatory effects, PROC LOGISTIC fits an
intercept-only model. You must specify exactly one MODEL statement. The optional label must be a valid
SAS name; it is used to identify the resulting output when you specify the ROC statement or the ROCCI
option.
Two forms of the MODEL statement can be specified. The first form, referred to as single-trial syntax, is
applicable to binary, ordinal, and nominal response data. The second form, referred to as events/trials syntax,
is restricted to the case of binary response data. The single-trial syntax is used when each observation in
the DATA= data set contains information about only a single trial, such as a single subject in an experiment.
When each observation contains information about multiple binary-response trials, such as the counts of the
number of subjects observed and the number responding, then events/trials syntax can be used.
In the events/trials syntax, you specify two variables that contain count data for a binomial experiment. These
two variables are separated by a slash. The value of the first variable, events, is the number of positive
responses (or events). The value of the second variable, trials, is the number of trials. The values of both
events and (trials–events) must be nonnegative and the value of trials must be positive for the response to be
valid.
In the single-trial syntax, you specify one variable (on the left side of the equal sign) as the response variable.
This variable can be character or numeric. Variable_options specific to the response variable can be specified
immediately after the response variable with parentheses around them.
For both forms of the MODEL statement, explanatory effects follow the equal sign. Variables can be either
continuous or classification variables. Classification variables can be character or numeric, and they must be
declared in the CLASS statement. When an effect is a classification variable, the procedure inserts a set of
coded columns into the design matrix instead of directly entering a single column containing the values of
the variable.
5424 F Chapter 73: The LOGISTIC Procedure
EVENT=’category ’ | keyword
specifies the event category for the binary response model. PROC LOGISTIC models the probability
of the event category. The EVENT= option has no effect when there are more than two response
categories. You can specify the value (formatted if a format is applied) of the event category in
quotation marks, or you can specify one of the following keywords:
FIRST designates the first ordered category as the event. This is the default.
LAST designates the last ordered category as the event.
One of the most common sets of response levels is {0,1}, with 1 representing the event for which the
probability is to be modeled. Consider the example where Y takes the values 1 and 0 for event and
nonevent, respectively, and Exposure is the explanatory variable. To specify the value 1 as the event
category, use the following MODEL statement:
REFERENCE=’category ’ | keyword
REF=’category ’ | keyword
specifies the reference category for the generalized logit model and the binary response model. For the
generalized logit model, each logit contrasts a nonreference category with the reference category. For
the binary response model, specifying one response category as the reference is the same as specifying
the other response category as the event category. You can specify the value (formatted if a format
MODEL Statement F 5425
is applied) of the reference category in quotation marks, or you can specify one of the following
keywords:
Model Options
Table 73.8 summarizes the options available in the MODEL statement. These options can be specified after a
slash (/).
Option Description
Model Specification Options
EQUALSLOPES Specifies equal slope parameters
LINK= Specifies the link function
NOFIT Suppresses model fitting
NOINT Suppresses the intercept
OFFSET= Specifies the offset variable
SELECTION= Specifies the effect selection method
UNEQUALSLOPES Specifies unequal slope parameters
Effect Selection Options
BEST= Controls the number of models displayed for SCORE
selection
DETAILS Requests detailed results at each step
FAST Uses the fast elimination method
HIERARCHY= Specifies whether and how hierarchy is maintained and
whether a single effect or multiple effects are allowed to
enter or leave the model per step
INCLUDE= Specifies the number of effects included in every model
MAXSTEP= Specifies the maximum number of steps for STEPWISE
selection
SEQUENTIAL Adds or deletes effects in sequential order
SLENTRY= Specifies the significance level for entering effects
SLSTAY= Specifies the significance level for removing effects
START= Specifies the number of variables in the first model
STOP= Specifies the number of variables in the final model
STOPRES Adds or deletes variables by the residual chi-square
criterion
Model-Fitting Specification Options
ABSFCONV= Specifies the absolute function convergence criterion
FCONV= Specifies the relative function convergence criterion
FIRTH Specifies Firth’s penalized likelihood method
GCONV= Specifies the relative gradient convergence criterion
MAXFUNCTION= Specifies the maximum number of function calls for the
conditional analysis
5426 F Chapter 73: The LOGISTIC Procedure
Option Description
MAXITER= Specifies the maximum number of iterations
NOCHECK Suppresses checking for infinite parameters
RIDGING= Specifies the technique used to improve the log-likelihood
function when its value is worse than that of the previous
step
SINGULAR= Specifies the tolerance for testing singularity
TECHNIQUE= Specifies the iterative algorithm for maximization
XCONV= Specifies the relative parameter convergence criterion
Confidence Interval Options
ALPHA= Specifies ˛ for the 100.1 ˛/% confidence intervals
CLODDS= Computes confidence intervals for odds ratios
CLPARM= Computes confidence intervals for parameters
PLCONV= Specifies the profile-likelihood convergence criterion
Classification Options
CTABLE Displays the classification table
PEVENT= Specifies prior event probabilities
PPROB= Specifies probability cutpoints for classification
Overdispersion and Goodness-of-Fit Test Options
AGGREGATE= Determines subpopulations for Pearson chi-square and
deviance
LACKFIT Requests the Hosmer and Lemeshow goodness-of-fit test
SCALE= Specifies the method to correct overdispersion
ROC Curve Options
OUTROC= Names the output ROC data set
ROCCI Computes confidence intervals for the AUC
ROCEPS= Specifies the probability grouping criterion
Regression Diagnostics Options
INFLUENCE Displays influence statistics
IPLOTS Displays influence statistics
Display Options
CORRB Displays the correlation matrix
COVB Displays the covariance matrix
EXPB Displays exponentiated values of the estimates
ITPRINT Displays the iteration history
NODUMMYPRINT Suppresses the “Class Level Information” table
NOODDSRATIO Suppresses the default “Odds Ratio” table
ORPVALUE Displays p-values in tables produced by the CLODDS=
option and ODDSRATIO statement
PARMLABEL Displays parameter labels
PCORR Displays the partial correlation statistic
RSQUARE Displays the generalized R-square
STB Displays standardized estimates
MODEL Statement F 5427
Option Description
Computational Options
BINWIDTH= Specifies the bin size for estimating association statistics
NOLOGSCALE Performs calculations by using normal scaling
ABSFCONV=value
specifies the absolute function convergence criterion. Convergence requires a small change in the
log-likelihood function in subsequent iterations,
where li is the value of the log-likelihood function at iteration i. See the section “Convergence Criteria”
on page 5462 for more information.
ALPHA=number
sets the level of significance ˛ for 100.1 ˛/% confidence intervals for regression parameters or odds
ratios. The value of number must be between 0 and 1. By default, number is equal to the value of
the ALPHA= option in the PROC LOGISTIC statement, or 0.05 if the option is not specified. This
option has no effect unless confidence limits for the parameters (CLPARM= option) or odds ratios
(CLODDS= option or ODDSRATIO statement) are requested.
BEST=number
specifies that number models with the highest score chi-square statistics are to be displayed for each
model size. It is used exclusively with the SCORE model selection method. If the BEST= option
is omitted and there are no more than 10 explanatory variables, then all possible models are listed
for each model size. If the option is omitted and there are more than 10 explanatory variables, then
the number of models selected for each model size is, at most, equal to the number of explanatory
variables listed in the MODEL statement.
BINWIDTH=width
specifies the size of the bins used for estimating the association statistics. For more information, see
the section “Rank Correlation of Observed Responses and Predicted Probabilities” on page 5473.
Valid values are 0 width < 1 (for polytomous response models, 0 < width < 1). In releases before
5428 F Chapter 73: The LOGISTIC Procedure
SAS/STAT 14.1, the default width is 0.002. Beginning in SAS/STAT 14.1, if you have a binary response
and fewer than 5,000,000 observations, the default is BINWIDTH=0; this means that no binning is
performed and the exact values of the statistics are computed using the trapezoidal area described in
the section “ROC Computations” on page 5482. This method is a bit slower and might require more
memory than the binning approach.
If a nonzero width does not evenly divide the unit interval, it is reduced to a valid value and a message
is displayed in the SAS log. The number of bins is also constrained by the amount of memory available
on your machine; if a nonzero width is so small that you cannot allocate 8=width bytes of memory for
the bins, then width is adjusted to a value for which memory can be allocated and a note is displayed in
the SAS log.
The BINWIDTH= option is ignored and no binning is performed when a ROC statement is specified,
when ROC graphics are produced, or when the SCORE statement computes an ROC area.
CORRB
displays the correlation matrix of the parameter estimates.
COVB
displays the covariance matrix of the parameter estimates.
CTABLE
classifies the input binary response observations according to whether the predicted event probabilities
are above or below some cutpoint value z in the range .0; 1/. An observation is predicted as an event if
the predicted event probability exceeds or equals z. You can supply a list of cutpoints other than the
default list by specifying the PPROB= option (page 5435). Also, false positive and negative rates can
be computed as posterior probabilities by using Bayes’ theorem. You can use the PEVENT= option to
specify prior probabilities for computing these rates. The CTABLE option is ignored if the data have
more than two response levels. The CTABLE option is not available with the STRATA statement.
For more information, see the section “Classification Table” on page 5475.
MODEL Statement F 5429
DETAILS
produces a summary of computational details for each step of the effect selection process. It produces
the “Analysis of Effects Eligible for Entry” table before displaying the effect selected for entry for
forward or stepwise selection. For each model fitted, it produces the “Joint Tests” or “Type 3 Analysis
of Effects” table if the fitted model involves CLASS variables, the “Analysis of Maximum Likelihood
Estimates” table, and measures of association between predicted probabilities and observed responses.
For the statistics included in these tables, see the section “Displayed Output” on page 5508. The
DETAILS option has no effect when SELECTION=NONE.
EXPB
EXPEST
ˇ i in the “Analysis of Maximum
displays the exponentiated values (eb̌i ) of the parameter estimates b
Likelihood Estimates” table for the logit model. These exponentiated values are the estimated odds
ratios for parameters corresponding to the continuous explanatory variables, and for CLASS effects
that use reference or GLM parameterizations.
FAST
uses a computational algorithm of Lawless and Singhal (1978) to compute a first-order approximation
to the remaining slope estimates for each subsequent elimination of a variable from the model. Variables
are removed from the model based on these approximate estimates. The FAST option is extremely
efficient because the model is not refitted for every variable removed. The FAST option is used when
SELECTION=BACKWARD and in the backward elimination steps when SELECTION=STEPWISE.
The FAST option is ignored when SELECTION=FORWARD or SELECTION=NONE.
5430 F Chapter 73: The LOGISTIC Procedure
FCONV=value
specifies the relative function convergence criterion. Convergence requires a small relative change in
the log-likelihood function in subsequent iterations,
jli li 1 j
< value
jli 1 j C 1E 6
where li is the value of the log likelihood at iteration i. See the section “Convergence Criteria” on
page 5462 for more information.
FIRTH
performs Firth’s penalized maximum likelihood estimation to reduce bias in the parameter estimates
(Heinze and Schemper 2002; Firth 1993). This method is useful in cases of separability, as often occurs
when the event is rare, and is an alternative to performing an exact logistic regression. See the section
“Firth’s Bias-Reducing Penalized Likelihood” on page 5462 for more information.
N OTE : The intercept-only log likelihood is modified by using the full-model Hessian, computed with
the slope parameters equal to zero. When fitting a model and scoring a data set in the same PROC
LOGISTIC step, the model is fit using Firth’s penalty for parameter estimation purposes, but the
penalty is not applied to the scored log likelihood.
GCONV=value
specifies the relative gradient convergence criterion. Convergence requires that the normalized predic-
tion function reduction is small,
gi0 Ii 1 gi
< value
jli j C 1E 6
where li is the value of the log-likelihood function, gi is the gradient vector, and Ii is the negative
(expected) Hessian matrix, all at iteration i. This is the default convergence criterion, and the default
value is 1E–8. See the section “Convergence Criteria” on page 5462 for more information.
HIERARCHY=keyword
HIER=keyword
specifies whether and how the model hierarchy requirement is applied and whether a single effect or
multiple effects are allowed to enter or leave the model in one step. You can specify that only CLASS ef-
fects, or both CLASS and interval effects, be subject to the hierarchy requirement. The HIERARCHY=
option is ignored unless you also specify one of the following options: SELECTION=FORWARD,
SELECTION=BACKWARD, or SELECTION=STEPWISE.
Model hierarchy refers to the requirement that, for any term to be in the model, all effects contained
in the term must be present in the model. For example, in order for the interaction A*B to enter the
model, the main effects A and B must be in the model. Likewise, neither effect A nor B can leave the
model while the interaction A*B is in the model.
The keywords you can specify in the HIERARCHY= option are as follows:
NONE indicates that the model hierarchy is not maintained. Any single effect can enter or leave
the model at any given step of the selection process.
SINGLE indicates that only one effect can enter or leave the model at one time, subject to the model
hierarchy requirement. For example, suppose that you specify the main effects A and B
MODEL Statement F 5431
and the interaction A*B in the model. In the first step of the selection process, either A
or B can enter the model. In the second step, the other main effect can enter the model.
The interaction effect can enter the model only when both main effects have already been
entered. Also, before A or B can be removed from the model, the A*B interaction must first
be removed. All effects (CLASS and interval) are subject to the hierarchy requirement.
SINGLECLASS is the same as HIERARCHY=SINGLE except that only CLASS effects are subject
to the hierarchy requirement.
MULTIPLE indicates that more than one effect can enter or leave the model at one time, subject
to the model hierarchy requirement. In a forward selection step, a single main effect can
enter the model, or an interaction can enter the model together with all the effects that are
contained in the interaction. In a backward elimination step, an interaction itself, or the
interaction together with all the effects that the interaction contains, can be removed. All
effects (CLASS and continuous) are subject to the hierarchy requirement.
MULTIPLECLASS is the same as HIERARCHY=MULTIPLE except that only CLASS effects are
subject to the hierarchy requirement.
The default value is HIERARCHY=SINGLE, which means that model hierarchy is to be maintained
for all effects (that is, both CLASS and continuous effects) and that only a single effect can enter or
leave the model at each step.
INCLUDE=number | EQUALSLOPES
specifies effects in the MODEL statement to include in every model during model selection. You can
specify the following values:
number requests that the first number effects be included in every model.
EQUALSLOPES enables you to include all the equal slope effects in every model and perform the
selection process on the unequal slope effects.
By default, INCLUDE=0. The INCLUDE= option has no effect when SELECTION=NONE. You
cannot specify the INCLUDE=EQUALSLOPES option if you specify SELECTION=SCORE.
If you specify the same effect in both the EQUALSLOPES and UNEQUALSLOPES options, then
that effect is treated as two separate effects. For example, suppose you specify the following MODEL
statement:
The X1 and X2 variables both generate an equal slope effect and an unequal slope effect. Specifying IN-
CLUDE=1 includes the equal slope effect for X1 in every model; specifying INCLUDE=2 includes both
the equal and unequal slope effects for X1 in every model; specifying INCLUDE=EQUALSLOPES
includes the equal slope effects for X1 and for X2 in every model.
Note that the INCLUDE= and START= options perform different tasks: the INCLUDE= option
includes effects in every model, whereas the START= option requires only that the effects appear in
the first model.
5432 F Chapter 73: The LOGISTIC Procedure
IPLOTS
produces an index plot for the regression diagnostic statistics developed by Pregibon (1981). An index
plot is a scatter plot with the regression diagnostic statistic represented on the Y axis and the case
number on the X axis. See Example 73.6 for an illustration.
ITPRINT
displays the iteration history of the maximum-likelihood model fitting. The ITPRINT option also
displays the last evaluation of the gradient vector and the final change in the –2 Log Likelihood.
LINK=keyword
L=keyword
specifies the link function linking the response probabilities to the linear predictors. You can specify
one of the following keywords. The default is LINK=LOGIT.
ALOGIT is the adjacent-category logit function. PROC LOGISTIC fits the adjacent-category logit
model, in which each category is contrasted with the following category.
CLOGLOG is the complementary log-log function. PROC LOGISTIC fits the binary complementary
log-log model when there are two response categories and fits the cumulative complemen-
tary log-log model when there are more than two response categories. The aliases are
CCLOGLOG, CCLL, and CUMCLOGLOG.
GLOGIT is the generalized logit function. PROC LOGISTIC fits the generalized logit model, in
which each nonreference category is contrasted with the reference category. You can use
the response variable option REF= to specify the reference category.
LOGIT is the log odds function. PROC LOGISTIC fits the binary logit model when there are
two response categories and fits the cumulative logit model when there are more than two
response categories. The aliases are CLOGIT and CUMLOGIT.
MODEL Statement F 5433
PROBIT is the inverse standard normal distribution function. PROC LOGISTIC fits the binary probit
model when there are two response categories and fits the cumulative probit model when
there are more than two response categories. The aliases are NORMIT, CPROBIT, and
CUMPROBIT.
MAXFUNCTION=number
specifies the maximum number of function calls to perform when maximizing the conditional likelihood.
This option is valid only when you specify an EQUALSLOPES or UNEQUALSLOPES option, or you
specify a STRATA statement. The default values are as follows:
Because the optimization is terminated only after completing a full iteration, the number of function
calls that are actually performed can exceed number . If convergence is not attained, the displayed
output and all output data sets that PROC LOGISTIC creates contain results that are based on the last
maximum likelihood iteration.
MAXITER=number
specifies the maximum number of iterations to perform. By default, MAXITER=25. If convergence is
not attained in number iterations, the displayed output and all output data sets created by the procedure
contain results that are based on the last maximum likelihood iteration.
MAXSTEP=number
specifies the maximum number of times any explanatory variable is added to or removed from the
model when SELECTION=STEPWISE. The default number is twice the number of explanatory
variables in the MODEL statement. When the MAXSTEP= limit is reached, the stepwise selection
process is terminated. All statistics displayed by the procedure (and included in output data sets) are
based on the last model fitted. The MAXSTEP= option has no effect when SELECTION=NONE,
FORWARD, or BACKWARD.
NOCHECK
disables the checking process to determine whether maximum likelihood estimates of the regression
parameters exist. If you are sure that the estimates are finite, this option can reduce the execution time
if the estimation takes more than eight iterations. For more information, see the section “Existence of
Maximum Likelihood Estimates” on page 5463.
NODUMMYPRINT
NODESIGNPRINT
NODP
suppresses the “Class Level Information” table, which shows how the design matrix columns for the
CLASS variables are coded.
5434 F Chapter 73: The LOGISTIC Procedure
NOINT
suppresses the intercept for the binary response model, the first intercept for the ordinal response model
(which forces all intercepts to be nonnegative), or all intercepts for the generalized logit model. This
can be particularly useful in conditional logistic analysis; see Example 73.11.
NOFIT
performs the global score test without fitting the model. The global score test evaluates the joint
significance of the effects in the MODEL statement. No further analyses are performed. If the NOFIT
option is specified along with other MODEL statement options, NOFIT takes effect and all other
options except FIRTH, LINK=, NOINT, OFFSET=, and TECHNIQUE= are ignored. The NOFIT
option is not available with the STRATA statement.
NOLOGSCALE
specifies that computations for the conditional and exact logistic regression models should be computed
by using normal scaling. Log scaling can handle numerically larger problems than normal scaling;
however, computations in the log scale are slower than computations in normal scale.
NOODDSRATIO
NOOR
suppresses the default “Odds Ratio” table.
OFFSET=name
names the offset variable. The regression coefficient for this variable will be fixed at 1. For an example
that uses this option, see Example 73.14. You can also use the OFFSET= option to restrict parameters to
a fixed value. For example, if you want to restrict the parameter for variable X1 to 1 and the parameter
for X2 to 2, compute RestrictD X1 C 2 X 2 in a DATA step, specify the option offset=Restrict,
and leave X1 and X2 out of the model.
ORPVALUE
displays p-values in tables produced by the CLODDS= option and the ODDSRATIO statement. A p-
value for an odds ratio corresponds to the significance level such that the two-sided confidence interval
for the odds ratio has “1” as one of its endpoints. Computing profile-likelihood p-values requires
optimizing restricted log-likelihood equations; for more information, see the section “Confidence
Intervals for Parameters” on page 5469. This option is not available with the FIRTH option.
OUTROC=SAS-data-set
OUTR=SAS-data-set
creates, for binary response models, an output SAS data set that contains the data necessary to produce
the receiver operating characteristic (ROC) curve. The OUTROC= option is not available with the
STRATA statement. See the section “OUTROC= Output Data Set” on page 5505 for the list of variables
in this data set.
PARMLABEL
displays the labels of the parameters in the “Analysis of Maximum Likelihood Estimates” table.
PCORR r
2i 2
computes the partial correlation statistic sign.ˇi / 2 log L0 for each parameter i, where 2i is the
Wald chi-square statistic for the parameter and log L0 is the log-likelihood of the intercept-only model
(Hilbe 2009, p. 101). If 2i < 2 then the partial correlation is set to 0. The partial correlation for the
intercept terms is set to missing.
MODEL Statement F 5435
PEVENT=value| (list)
specifies one prior probability or a list of prior probabilities for the event of interest. The false positive
and false negative rates are then computed as posterior probabilities by Bayes’ theorem. The prior
probability is also used in computing the rate of correct prediction. For each prior probability in the
given list, a classification table of all observations is computed. By default, the prior probability is
the total sample proportion of events. The PEVENT= option is useful for stratified samples. It has no
effect if the CTABLE option is not specified. For more information, see the section “False Positive,
False Negative, and Correct Classification Rates Using Bayes’ Theorem” on page 5476. Also see the
PPROB= option for information about how the list is specified.
PLCL
is the same as specifying CLPARM=PL.
PLCONV=value
controls the convergence criterion for confidence intervals based on the profile-likelihood function.
The quantity value must be a positive number, with a default value of 1E–4. The PLCONV= option
has no effect if profile-likelihood confidence intervals (CLPARM=PL) are not requested.
PLRL
is the same as specifying CLODDS=PL.
PPROB=value | (list)
specifies one critical probability value (or cutpoint) or a list of critical probability values for classifying
observations with the CTABLE option. Each value must be between 0 and 1. A response that has a
cross validated predicted probability greater than or equal to the current PPROB= value is classified as
an event response. The PPROB= option is ignored if the CTABLE option is not specified.
A classification table for each of several cutpoints can be requested by specifying a list. For example,
the following statement requests a classification of the observations for each of the cutpoints 0.3, 0.5,
0.6, 0.7, and 0.8:
If the PPROB= option is not specified, the default is to display the classification for a range of
probabilities from the smallest estimated probability (rounded down to the nearest 0.02) to the highest
estimated probability (rounded up to the nearest 0.02) with 0.02 increments.
RISKLIMITS
RL
WALDRL
is the same as specifying CLODDS=WALD.
ROCCI
displays standard errors and confidence limits for the area under the ROC curve (AUC) statistic when
you have a binary response variable. This option replaces the “Association of Predicted Probabilities
and Observed Responses” table with the “ROC Association Statistics” table.
ROCEPS=number
specifies a criterion for the ROC curve used for grouping estimated event probabilities that are close
to each other. In each group, the difference between the largest and the smallest estimated event
probabilities does not exceed the given value. The value for number must be between 0 and 1; the
default value is the square root of the machine epsilon, which is about 1E–8 (in releases prior to
9.2, the default was 1E–4). The smallest estimated probability in each group serves as a cutpoint for
predicting an event response. The ROCEPS= option has no effect unless the OUTROC= option, the
BINWIDTH=0 option, or a ROC statement is specified.
RSQUARE
RSQ
requests a generalized R-square measure for the fitted model. For more information, see the section
“Generalized Coefficient of Determination” on page 5466.
SCALE=scale
enables you to supply the value of the dispersion parameter or to specify the method for estimating
the dispersion parameter. It also enables you to display the “Deviance and Pearson Goodness-of-Fit
Statistics” table. To correct for overdispersion or underdispersion, the covariance matrix is multiplied
by the estimate of the dispersion parameter. Valid values for scale are as follows:
D | DEVIANCE specifies that the dispersion parameter be estimated by the deviance divided by its
degrees of freedom.
P | PEARSON specifies that the dispersion parameter be estimated by the Pearson chi-square
statistic divided by its degrees of freedom.
WILLIAMS < (constant ) > specifies that Williams’ method be used to model overdispersion. This
option can be used only with the events/trials syntax. An optional constant can be specified
as the scale parameter; otherwise, a scale parameter is estimated under the full model. A
set of weights is created based on this scale parameter estimate. These weights can then be
used in fitting subsequent models of fewer terms than the full model. When fitting these
submodels, specify the computed scale parameter as constant . See Example 73.10 for an
illustration.
N | NONE specifies that no correction is needed for the dispersion parameter; that is, the dispersion
parameter remains as 1. This specification is used for requesting the deviance and the
Pearson chi-square statistic without adjusting for overdispersion.
constant sets the estimate of the dispersion parameter to be the square of the given constant . For
example, SCALE=2 sets the dispersion parameter to 4. The value constant must be a
positive number.
MODEL Statement F 5437
You can use the AGGREGATE (or AGGREGATE=) option to define the subpopulations for calcu-
lating the Pearson chi-square statistic and the deviance. In the absence of the AGGREGATE (or
AGGREGATE=) option, each observation is regarded as coming from a different subpopulation. For
the events/trials syntax, each observation consists of n Bernoulli trials, where n is the value of the trials
variable. For single-trial syntax, each observation consists of a single response, and for this setting it is
not appropriate to carry out the Pearson or deviance goodness-of-fit analysis. Thus, PROC LOGISTIC
ignores specifications SCALE=P, SCALE=D, and SCALE=N when single-trial syntax is specified
without the AGGREGATE (or AGGREGATE=) option.
The “Deviance and Pearson Goodness-of-Fit Statistics” table includes the Pearson chi-square statistic,
the deviance, the degrees of freedom, the ratio of each statistic divided by its degrees of freedom, and
the corresponding p-value. The SCALE= option is not available with the STRATA statement. For
more information, see the section “Overdispersion” on page 5477.
SELECTION=BACKWARD | B
| FORWARD | F
| NONE | N
| STEPWISE | S
| SCORE
specifies the method used to select the variables in the model. BACKWARD requests backward
elimination, FORWARD requests forward selection, NONE fits the complete model specified in the
MODEL statement, and STEPWISE requests stepwise selection. SCORE requests best subset selection.
By default, SELECTION=NONE.
For more information, see the section “Effect-Selection Methods” on page 5464.
SEQUENTIAL
SEQ
forces effects to be added to the model in the order specified in the MODEL statement or eliminated
from the model in the reverse order of that specified in the MODEL statement. The model-building
process continues until the next effect to be added has an insignificant adjusted chi-square statistic or
until the next effect to be deleted has a significant Wald chi-square statistic. The SEQUENTIAL option
has no effect when SELECTION=NONE.
SINGULAR=value
specifies the tolerance for testing the singularity of the Hessian matrix (Newton-Raphson algorithm) or
the expected value of the Hessian matrix (Fisher scoring algorithm). The Hessian matrix is the matrix
of second partial derivatives of the log-likelihood function. The test requires that a pivot for sweeping
this matrix be at least this number times a norm of the matrix. Values of the SINGULAR= option must
be numeric. By default, value is the machine epsilon times 1E7, which is approximately 1E–9.
SLENTRY=value
SLE=value
specifies the significance level of the score chi-square for entering an effect into the model in the FOR-
WARD or STEPWISE method. Values of the SLENTRY= option should be between 0 and 1, inclusive.
By default, SLENTRY=0.05. The SLENTRY= option has no effect when SELECTION=NONE,
SELECTION=BACKWARD, or SELECTION=SCORE.
5438 F Chapter 73: The LOGISTIC Procedure
SLSTAY=value
SLS=value
specifies the significance level of the Wald chi-square for an effect to stay in the model in a back-
ward elimination step. Values of the SLSTAY= option should be between 0 and 1, inclusive. By
default, SLSTAY=0.05. The SLSTAY= option has no effect when SELECTION=NONE, SELEC-
TION=FORWARD, or SELECTION=SCORE.
START=number | EQUALSLOPES
specifies which effects in the MODEL statement are included in the initial model. You can specify the
following values:
number requests that the first number effects be included in the initial model. The
value of number ranges from 0 to s, where s is the total number of effects
that are specified in the MODEL statement. The default value of number is
s when SELECTION=BACKWARD and 0 when SELECTION=FORWARD
or SELECTION=STEPWISE. When SELECTION=SCORE, START=number
specifies that the smallest models contain number effects, where number ranges
from 1 to s; the default value is 1.
EQUALSLOPES enables you to begin the model selection process with all the equal slope effects
in the model.
The START= option has no effect when SELECTION=NONE. You cannot specify the
START=EQUALSLOPES option if you specify SELECTION=SCORE.
If you specify the same effect in both the EQUALSLOPES and UNEQUALSLOPES options, then
that effect is treated as two separate effects. For example, suppose you specify the following MODEL
statement:
The X1 and X2 variables both generate an equal slope effect and an unequal slope effect. Specifying
START=1 includes the equal slope effect for X1 in the initial model; specifying START=2 includes both
the equal and unequal slope effects for X1 in the initial model; specifying START=EQUALSLOPES
includes the equal slope effects for X1 and for X2 in the initial model.
Note that the INCLUDE= and START= options perform different tasks: the INCLUDE= option
includes effects in every model, whereas the START= option requires only that the effects appear in
the first model.
STB
displays the standardized estimates for the parameters in the “Analysis of Maximum Likelihood
Estimates” table. The standardized estimate of ˇi is given by b
ˇ i =.s=si /, where si is the total sample
standard deviation for the ith explanatory variable and
8 p
< = 3 Logistic
sD 1 p Normal
= 6 Extreme-value
:
The sample standard deviations for parameters associated with CLASS and EFFECT variables are
computed using their codings. For the intercept parameters, the standardized estimates are set to
missing.
MODEL Statement F 5439
STOP=number
specifies the maximum (SELECTION=FORWARD) or minimum (SELECTION=BACKWARD)
number of effects to be included in the final model. The effect selection process is stopped when
number effects are found. The value of number ranges from 0 to s, where s is the total number of
effects in the MODEL statement. The default value of number is s for the FORWARD method and
0 for the BACKWARD method. For the SCORE method, STOP=number specifies that the largest
models contain number effects, where number ranges from 1 to s; the default value of number is s.
The STOP= option has no effect when SELECTION=NONE or STEPWISE.
STOPRES
SR
specifies that the removal or entry of effects be based on the value of the residual chi-square. If
SELECTION=FORWARD, then the STOPRES option adds the effects into the model one at a time
until the residual chi-square becomes insignificant (until the p-value of the residual chi-square exceeds
the SLENTRY= value). If SELECTION=BACKWARD, then the STOPRES option removes effects
from the model one at a time until the residual chi-square becomes significant (until the p-value of the
residual chi-square becomes less than the SLSTAY= value). The STOPRES option has no effect when
SELECTION=NONE or SELECTION=STEPWISE.
TECHNIQUE=FISHER | NEWTON
TECH=FISHER | NEWTON
specifies the optimization technique for estimating the regression parameters. NEWTON (or NR) is
the Newton-Raphson algorithm and FISHER (or FS) is the Fisher scoring algorithm. Both techniques
yield the same estimates, but the estimated covariance matrices are slightly different except for the case
when the LOGIT link is specified for binary response data. By default, TECHNIQUE=FISHER. If
you specify the LINK=GLOGIT or LINK=ALOGIT option, then Newton-Raphson is the default and
only available method. The TECHNIQUE= option is not applied to conditional and exact conditional
analyses. This option is not available when the EQUALSLOPES or UNEQUALSLOPES option is
specified. For more information, see the section “Iterative Algorithms for Model Fitting” on page 5460.
WALDCL
CL
is the same as specifying CLPARM=WALD.
XCONV=value
specifies the relative parameter convergence criterion. Convergence requires a small relative parameter
change in subsequent iterations,
.i /
max jıj j < value
j
where
8 .i / .i 1/ .i 1/
< ˇj ˇj jˇj j < 0:01
.i /
ıj D .i/
ˇj ˇj
.i 1/
: .i 1/ otherwise
ˇj
.i /
and ˇj is the estimate of the jth parameter at iteration i. See the section “Convergence Criteria” on
page 5462 for more information.
NLOPTIONS Statement F 5441
NLOPTIONS Statement
NLOPTIONS < options > ;
The NLOPTIONS statement controls the optimization process for conditional analyses (which result from
specifying a STRATA statement) and for partial parallel slope models (which result from specifying the
EQUALSLOPES or UNEQUALSLOPES option in the MODEL statement). An option that is specified in the
NLOPTIONS statement takes precedence over the same option specified in the MODEL statement.
The default optimization techniques are chosen according to the number of parameters, p, as follows:
The available options are described in the section “NLOPTIONS Statement” on page 489 in Chapter 19,
“Shared Concepts and Topics.”
ODDSRATIO Statement
ODDSRATIO < 'label' > variable < / options > ;
The ODDSRATIO statement produces odds ratios for variable even when the variable is involved in inter-
actions with other covariates, and for classification variables that use any parameterization. You can also
specify variables on which constructed effects are based, in addition to the names of COLLECTION or
MULTIMEMBER effects. You can specify several ODDSRATIO statements.
If variable is continuous, then the odds ratios honor any values specified in the UNITS statement. If variable
is a classification variable, then odds ratios comparing each pairwise difference between the levels of variable
are produced. If variable interacts with a continuous variable, then the odds ratios are produced at the mean
of the interacting covariate by default. If variable interacts with a classification variable, then the odds ratios
are produced at each level of the interacting covariate by default. The computed odds ratios are independent
of the parameterization of any classification variable.
The odds ratios are uniquely labeled by concatenating the following terms to variable:
1. If this is a polytomous response model, then prefix the response variable and the level describing the
logit followed by a colon; for example, “Y 0:”.
2. If variable is continuous and the UNITS statement provides a value that is not equal to 1, then append
“Units=value”; otherwise, if variable is a classification variable, then append the levels being contrasted;
for example, “cat vs dog”.
3. Append all interacting covariates preceded by “At”; for example, “At X=1.2 A=cat”.
If you are also creating odds ratio plots, then this label is displayed on the plots (see the PLOTS option for
more information). If you specify a 'label' in the ODDSRATIO statement, then the odds ratios produced by
this statement are also labeled: 'label', 'label 2', 'label 3',. . . , and these are the labels used in the plots. If there
5442 F Chapter 73: The LOGISTIC Procedure
are any duplicated labels across all ODDSRATIO statements, then the corresponding odds ratios are not
displayed on the plots.
The following options are available:
CL=WALD | PL | BOTH
specifies whether to create Wald or profile-likelihood confidence limits, or both. By default, Wald
confidence limits are produced.
DIFF=REF | ALL
specifies whether the odds ratios for a classification variable are computed against the reference level,
or all pairs of variable are compared. By default, DIFF=ALL. The DIFF= option is ignored when
variable is continuous.
PLCONV=value
controls the convergence criterion for confidence intervals based on the profile-likelihood function.
The quantity value must be a positive number, with a default value of 1E–4. The PLCONV= option
has no effect if profile-likelihood confidence intervals (CL=PL) are not requested.
PLMAXITER=number
specifies the maximum number of iterations to perform. By default, PLMAXITER=25. If conver-
gence is not attained in number iterations, the odds ratio or the confidence limits are set to missing.
The PLMAXITER= option has no effect if profile-likelihood confidence intervals (CL=PL) are not
requested.
PLSINGULAR=value
specifies the tolerance for testing the singularity of the Hessian matrix (Newton-Raphson algorithm) or
the expected value of the Hessian matrix (Fisher scoring algorithm). The test requires that a pivot for
sweeping this matrix be at least this number times a norm of the matrix. Values of the PLSINGULAR=
option must be numeric. By default, value is the machine epsilon times 1E7, which is approximately
1E–9. The PLSINGULAR= option has no effect if profile-likelihood confidence intervals (CL=PL) are
not requested.
OUTPUT Statement F 5443
OUTPUT Statement
OUTPUT < OUT=SAS-data-set > < keyword-options > / < option > ;
The OUTPUT statement creates a new SAS data set that contains all the variables in the input data set and,
optionally, the estimated linear predictors and their standard error estimates, the estimates of the cumulative
or individual response probabilities, and the confidence limits for the cumulative probabilities. Regression
diagnostic statistics and estimates of cross validated response probabilities are also available for binary
response models. If you specify more than one OUTPUT statement, only the last one is used. Formulas for
the statistics are given in the sections “Linear Predictor, Predicted Probability, and Confidence Limits” on
page 5474 and “Regression Diagnostics” on page 5484, and, for conditional logistic regression, in the section
“Conditional Logistic Regression” on page 5492.
If you use the single-trial syntax, the data set also contains a variable named _LEVEL_, which indicates the
level of the response that the given row of output is referring to. For instance, the value of the cumulative
probability variable is the probability that the response variable is as large as the corresponding value of
_LEVEL_. For more information, see the section “OUT= Output Data Set in the OUTPUT Statement” on
page 5501.
The estimated linear predictor, its standard error estimate, all predicted probabilities, and the confidence
limits for the cumulative probabilities are computed for all observations in which the explanatory variables
have no missing values, even if the response is missing. By adding observations with missing response values
to the input data set, you can compute these statistics for new observations or for settings of the explanatory
variables not present in the data without affecting the model fit. Alternatively, the SCORE statement can be
used to compute predicted probabilities and confidence intervals for new observations.
Table 73.9 summarizes the options available in the OUTPUT statement. The statistic and diagnostic keyword-
options specify the statistics to be included in the output data set and name the new variables that contain
the statistics. If a STRATA statement is specified, only the PREDICTED=, RESCHI=, STDRESCHI=,
DFBETAS=, and H= options are available; for more information, see the section “Regression Diagnostic
Details” on page 5493.
Option Description
ALPHA= Specifies ˛ for the 100.1 ˛/% confidence intervals
OUT= Names the output data set
Statistic keyword-options
LOWER= Names the lower confidence limit
PREDICTED= Names the predicted probabilities
PREDPROBS= Requests the individual, cumulative, or cross validated predicted probabili-
ties
STDXBETA= Names the standard error estimate of the linear predictor
UPPER= Names the upper confidence limit
XBETA= Names the linear predictor
5444 F Chapter 73: The LOGISTIC Procedure
Option Description
Diagnostic keyword-options for binary response
C= Names the confidence interval displacement
CBAR= Names the confidence interval displacement
DFBETAS= Names the standardized deletion parameter differences
DIFCHISQ= Names the deletion chi-square goodness-of-fit change
DIFDEV= Names the deletion deviance change
H= Names the leverage
RESCHI= Names the Pearson chi-square residual
RESDEV= Names the deviance residual
RESLIK= Names the likelihood residual
STDRESCHI= Names the standardized Pearson chi-square residual
STDRESDEV= Names the standardized deviance residual
OUT=SAS-data-set
names the output data set. If you omit the OUT= option, the output data set is created and given a
default name by using the DATAn convention.
keyword-options
specifies the statistics to include in the output data set and provides names for the new variables that
contain the statistics. The keyword-options have the following form:
keyword=name | name-list | option
The keyword designates the desired statistic (see the following list of keywords), name provides the
name of the new variable that contains the statistic, name-list is a list of such names, and option
generates the statistics using default names.
The keywords allowed and the statistics they represent are as follows:
C=name
specifies the confidence interval displacement diagnostic that measures the influence of individual
observations on the regression estimates.
CBAR=name
specifies the confidence interval displacement diagnostic that measures the overall change in the
global regression estimates due to deleting an individual observation.
DFBETAS=_ALL_ | name-list
specifies the standardized differences in the regression estimates for assessing the effects of
individual observations on the estimated regression parameters in the fitted model. You can
specify a list of up to s C 1 variable names, where s is the number of explanatory variables in the
MODEL statement, or you can specify just the keyword _ALL_. In the former specification, the
first variable contains the standardized differences in the intercept estimate, the second variable
contains the standardized differences in the parameter estimate for the first explanatory variable in
the MODEL statement, and so on. In the latter specification, the DFBETAS statistics are named
OUTPUT Statement F 5445
DFBETA_xxx, where xxx is the name of the regression parameter. For example, if the model
contains two variables X1 and X2, the specification DFBETAS=_ALL_ produces three DFBETAS
statistics: DFBETA_Intercept, DFBETA_X1, and DFBETA_X2. If an explanatory variable is
not included in the final model, the corresponding output variable named in DFBETAS=var-list
contains missing values.
DIFCHISQ=name
specifies the change in the chi-square goodness-of-fit statistic attributable to deleting the individual
observation.
DIFDEV=name
specifies the change in the deviance attributable to deleting the individual observation.
H=name
specifies the diagonal element of the hat matrix for detecting extreme points in the design space.
LOWER=name
L=name
names the variable containing the lower confidence limits for , where is the probability of the
event response if events/trials syntax or single-trial syntax with binary response is specified; for a
cumulative model, is the cumulative probability (that is, the probability that the response is
less than or equal to the value of _LEVEL_); for the generalized logit model, it is the individual
probability (that is, the probability that the response category is represented by the value of
_LEVEL_). See the ALPHA= option to set the confidence level.
PREDICTED=name
PRED=name
PROB=name
P=name
names the variable containing the predicted probabilities. For the events/trials syntax or single-
trial syntax with binary response, it is the predicted event probability. For a cumulative model, it
is the predicted cumulative probability (that is, the probability that the response variable is less
than or equal to the value of _LEVEL_); and for the generalized logit model, it is the predicted
individual probability (that is, the probability of the response category represented by the value
of _LEVEL_).
PREDPROBS=(options)
requests individual, cumulative, or cross validated predicted probabilities. Descriptions of the
options are as follows.
INDIVIDUAL | I requests the predicted probability of each response level. For a response
variable Y with three levels, 1, 2, and 3, the individual probabilities are Pr(Y=1),
Pr(Y=2), and Pr(Y=3).
CUMULATIVE | C requests the cumulative predicted probability of each response level. For
a response variable Y with three levels, 1, 2, and 3, the cumulative probabilities are
Pr(Y1), Pr(Y2), and Pr(Y3). The cumulative probability for the last response
level always has the constant value of 1. For generalized logit models, the cumulative
predicted probabilities are not computed and are set to missing.
5446 F Chapter 73: The LOGISTIC Procedure
For more information, see the section “Details of the PREDPROBS= Option” on page 5447 at
the end of this section.
RESCHI=name
specifies the Pearson (chi-square) residual for identifying observations that are poorly accounted
for by the model.
RESDEV=name
specifies the deviance residual for identifying poorly fitted observations.
RESLIK=name
specifies the likelihood residual for identifying poorly fitted observations.
STDRESCHI=name
specifies the standardized Pearson (chi-square) residual for identifying observations that are
poorly accounted for by the model.
STDRESDEV=name
specifies the standardized deviance residual for identifying poorly fitted observations.
STDXBETA=name
names the variable containing the standard error estimates of XBETA. For more information, see
the section “Linear Predictor, Predicted Probability, and Confidence Limits” on page 5474.
UPPER=name
U=name
names the variable containing the upper confidence limits for , where is the probability of
the event response if events/trials syntax or single-trial syntax with binary response is specified;
for a cumulative model, is cumulative probability (that is, the probability that the response is
less than or equal to the value of _LEVEL_); for the generalized logit model, it is the individual
probability (that is, the probability that the response category is represented by the value of
_LEVEL_). See the ALPHA= option to set the confidence level.
XBETA=name
names the variable containing the estimates of the linear predictor ˛i C ˇ 0 x, where i is the
corresponding ordered value of _LEVEL_.
The following option is available in the OUTPUT statement and is specified after a slash (/):
ALPHA=number
sets the level of significance ˛ for 100.1 ˛/% confidence limits for the appropriate response
probabilities. The value of number must be between 0 and 1. By default, number is equal to the value
of the ALPHA= option in the PROC LOGISTIC statement, or 0.05 if that option is not specified. You
can specify this option after a slash(/).
OUTPUT Statement F 5447
If you specify events/trials syntax, xxx is either ‘Event’ or ‘Nonevent’. Thus, the variable containing
the event probabilities is named IP_Event and the variable containing the nonevent probabilities is
named IP_Nonevent.
If you specify the single-trial syntax with more than one BY group, xxx is 1 for the first ordered level
of the response, 2 for the second ordered level of the response, and so forth, as given in the “Response
Profile” table. The variable containing the predicted probabilities Pr(Y=1) is named IP_1, where Y is
the response variable. Similarly, IP_2 is the name of the variable containing the predicted probabilities
Pr(Y=2), and so on.
If you specify the single-trial syntax with no BY-group processing, xxx is the left-justified formatted
value of the response level (the value might be truncated so that IP_xxx does not exceed 32 characters).
For example, if Y is the response variable with response levels ‘None’, ‘Mild’, and ‘Severe’, the
variables representing individual probabilities Pr(Y=’None’), P(Y=’Mild’), and P(Y=’Severe’) are
named IP_None, IP_Mild, and IP_Severe, respectively.
If you specify PREDPROBS=CUMULATIVE, the OUT= data set contains k additional variables representing
the cumulative probabilities, one for each response level, where k is the maximum number of response
levels across all BY groups. The names of these variables have the form CP_xxx, where xxx represents the
particular response level. The naming convention is similar to that given by PREDPROBS=INDIVIDUAL.
The PREDPROBS=CUMULATIVE values are the same as those output by the PREDICT= option, but are
arranged in variables on each output observation rather than in multiple output observations.
If you specify PREDPROBS=CROSSVALIDATE, the OUT= data set contains k additional variables repre-
senting the cross validated predicted probabilities of the k response levels, where k is the maximum number
of response levels across all BY groups. The names of these variables have the form XP_xxx, where xxx
represents the particular level. The representation is the same as that given by PREDPROBS=INDIVIDUAL
except that for the events/trials syntax there are four variables for the cross validated predicted probabilities
instead of two:
XP_EVENT_R1E is the cross validated predicted probability of an event when a single event is removed
from the current observation.
XP_NONEVENT_R1E is the cross validated predicted probability of a nonevent when a single event is
removed from the current observation.
5448 F Chapter 73: The LOGISTIC Procedure
XP_EVENT_R1N is the cross validated predicted probability of an event when a single nonevent is removed
from the current observation.
XP_NONEVENT_R1N is the cross validated predicted probability of a nonevent when a single nonevent is
removed from the current observation.
The cross validated predicted probabilities are precisely those used in the CTABLE option. For more
information about the computation, see the section “Predicted Probability of an Event for Classification” on
page 5476.
ROC Statement
ROC < 'label' > < specification > < / options > ;
The ROC statements specify models to be used in the ROC comparisons. You can specify more than one
ROC statement. ROC statements are identified by their label —if you do not specify a label , the ith ROC
statement is labeled “ROCi”. Additionally, the specified or selected model is labeled with the MODEL
statement label or “Model” if the MODEL label is not present. The specification can be either a list of effects
that have previously been specified in the MODEL statement, or PRED=variable, where the variable does not
have to be specified in the MODEL statement. The PRED= option enables you to input a criterion produced
outside PROC LOGISTIC; for example, you can fit a random-intercept model by using PROC GLIMMIX
or use survey weights in PROC SURVEYLOGISTIC, then use the predicted values from those models to
produce an ROC curve for the comparisons. If you do not make a specification, then an intercept-only model
is fit to the data, resulting in a noninformative ROC curve that can be used for comparing the area under
another ROC curve to 0.5.
You can specify a ROCCONTRAST statement and a ROCOPTIONS option in the PROC LOGISTIC
statement to control how the models are compared, while the PLOTS=ROC option controls the ODS Graphics
displays. See Example 73.8 for an example that uses the ROC statement.
If you specify any options, then a “ROC Model Information” table summarizing the new ROC model is
displayed. The options are ignored for the PRED= specification. The following options are available:
NOINT
suppresses the intercept for the model. Note that specifying the NOINT option in the MODEL
statement does not affect the ROC model.
NOOFFSET
does not include an offset variable if the OFFSET= option is specified in the MODEL statement.
A constant offset has no effect on the ROC curve, although the cutpoints might be different, but a
nonconstant offset can affect the parameter estimates and hence the ROC curve.
LINK=keyword
specifies the link function to be used in the model. The available keywords are LOGIT, NORMIT, and
CLOGLOG. The logit link is the default. Note that the LINK= option in the MODEL statement is
ignored.
ROCCONTRAST Statement F 5449
ROCCONTRAST Statement
ROCCONTRAST < 'label' > < contrast > < / options > ;
The ROCCONTRAST statement compares the different ROC models. You can specify only one ROCCON-
TRAST statement. The ROCOPTIONS options in the PROC LOGISTIC statement control how the models
are compared. You can specify one of the following contrast specifications:
ADJACENTPAIRS
produces a contrast matrix of each ROC curve minus the succeeding curve.
matrix
specifies the contrast in the form row1,row2,..., where each row contains the coefficients used to
compare the ROC curves. Each row must contain the same number of entries as there are ROC curves
being compared. The elements of each row refer to the ROC statements in the order in which they are
specified. However, the first element of each row refers either to the fitted model, the selected model
if a SELECTION= method is specified, or the first specified ROC statement if the NOFIT option is
specified.
If no contrast is specified, then the REFERENCE contrast with the default reference curve is used. See the
section “Comparing ROC Curves” on page 5481 for more information about comparing ROC curves, and see
Example 73.8 for an example.
The following options are available:
E
displays the contrast.
COV
displays covariance matrices used in the computations.
5450 F Chapter 73: The LOGISTIC Procedure
SCORE Statement
SCORE < options > ;
The SCORE statement creates a data set that contains all the data in the DATA= data set together with
posterior probabilities and, optionally, prediction confidence intervals. Fit statistics are displayed on request.
If you have binary response data, the SCORE statement can be used to create a data set containing data for
the ROC curve. You can specify several SCORE statements. FREQ, WEIGHT, and BY statements can be
used with the SCORE statements. Weights do not affect the computation of predicted probabilities, their
confidence limits, or the predicted response level. Weights affect some fit statistics as described in “Fit
Statistics for Scored Data Sets” on page 5488. The SCORE statement is not available with the STRATA
statement.
If a SCORE statement is specified in the same run as fitting the model, FORMAT statements should be
specified after the SCORE statement in order for the formats to apply to all the DATA= and PRIOR= data
sets in the SCORE statement.
See the section “Scoring Data Sets” on page 5487 for more information, and see Example 73.16 for an
illustration of how to use this statement.
Table 73.10 summarizes the options available in the SCORE statement.
Option Description
ALPHA= Specifies the significance level
CLM Outputs the Wald-test-based confidence limits
CUMULATIVE Outputs the cumulative predicted probabilities
DATA= Names the SAS data that you want to score
FITSTAT Displays fit statistics
OUT= Names the SAS data set that contains the predicted information
OUTROC= Names the SAS data set that contains the ROC curve
PRIOR= Names the SAS data set that contains the priors of the response categories
PRIOREVENT= Specifies the prior event probability
ROCEPS= Specifies the criterion for grouping estimated event probabilities
ALPHA=number
specifies the significance level ˛ for 100.1 ˛/% confidence intervals. By default, the value of
number is equal to the ALPHA= option in the PROC LOGISTIC statement, or 0.05 if that option is
not specified. This option has no effect unless the CLM option in the SCORE statement is requested.
CLM
outputs the Wald-test-based confidence limits for the predicted probabilities. This option is not available
when the INMODEL= data set is created with the NOCOV option.
SCORE Statement F 5451
CUMULATIVE
outputs the cumulative predicted probabilities Pr.Y i /; i D 1; : : : ; k C 1, to the OUT= data set.
This option is valid only when you have more than two response levels; otherwise, the option is ignored
and a note is printed in the SAS log. These probabilities are named CP_level_i, where level_i is the ith
response level.
If the CLM option is also specified in the SCORE statement, then the Wald-based confidence limits for
the cumulative predicted probabilities are also output. The confidence limits are named CLCL_level_i
and CUCL_level_i. In particular, for the lowest response level, the cumulative values (CP, CLCL,
CUCL) should be identical to the individual values (P, LCL, UCL), and for the highest response level
CP=CLCL=CUCL=1.
DATA=SAS-data-set
names the SAS data set that you want to score. If you omit the DATA= option in the SCORE statement,
then scoring is performed on the DATA= input data set in the PROC LOGISTIC statement, if specified;
otherwise, the DATA=_LAST_ data set is used.
It is not necessary for the DATA= data set in the SCORE statement to contain the response variable
unless you are specifying the FITSTAT or OUTROC= option.
Only those variables involved in the fitted model effects are required in the DATA= data set in the
SCORE statement. For example, the following statements use forward selection to select effects:
Suppose Treatment and Age are the effects selected for the final model. You can score a data set that
does not contain the variable Sex because the effect Sex is not in the model that the scoring is based
on. For example, the following statements score the Neuralgia data set after dropping the Sex variable:
FITSTAT
displays fit statistics for the data set you are scoring. The data set must contain the response variable.
For more information, see the section “Fit Statistics for Scored Data Sets” on page 5488.
OUT=SAS-data-set
names the SAS data set that contains the predicted information. If you omit the OUT= option, the
output data set is created and given a default name by using the DATAn convention.
OUTROC=SAS-data-set
names the SAS data set that contains the ROC curve for the DATA= data set. The ROC curve is
computed only for binary response data. See the section “OUTROC= Output Data Set” on page 5505
for the list of variables in this data set.
5452 F Chapter 73: The LOGISTIC Procedure
PRIOR=SAS-data-set
names the SAS data set that contains the priors of the response categories. The priors can be values
proportional to the prior probabilities; thus, they do not necessarily sum to one. This data set should
include a variable named _PRIOR_ that contains the prior probabilities. For events/trials MODEL
statement syntax, this data set should also include an _OUTCOME_ variable that contains the values
EVENT and NONEVENT; for single-trial syntax, this data set should include the response variable
that contains the unformatted response categories. See Example 73.16 for an example.
PRIOREVENT=value
specifies the prior event probability for a binary response model. If both PRIOR= and PRIOREVENT=
options are specified, the PRIOR= option takes precedence.
ROCEPS=value
specifies the criterion for grouping estimated event probabilities that are close to each other for the ROC
curve. In each group, the difference between the largest and the smallest estimated event probability
does not exceed the given value. The value must be between 0 and 1; the default value is the square
root of the machine epsilon, which is about 1E–8 (in releases prior to 9.2, the default was 1E–4). The
smallest estimated probability in each group serves as a cutpoint for predicting an event response. The
ROCEPS= option has no effect if the OUTROC= option is not specified in the SCORE statement.
SLICE Statement
SLICE model-effect < / options > ;
The SLICE statement provides a general mechanism for performing a partitioned analysis of the LS-means
for an interaction. This analysis is also known as an analysis of simple effects.
The SLICE statement uses the same options as the LSMEANS statement, which are summarized in Ta-
ble 19.21. For details about the syntax of the SLICE statement, see the section “SLICE Statement” on
page 506 in Chapter 19, “Shared Concepts and Topics.”
N OTE : If you have classification variables in your model, then the SLICE statement is allowed only if you
also specify the PARAM=GLM option.
STORE Statement
STORE < OUT= >item-store-name < / LABEL='label' > ;
The STORE statement requests that the procedure save the context and results of the statistical analysis. The
resulting item store has a binary file format that cannot be modified. The contents of the item store can be
processed with the PLM procedure. For details about the syntax of the STORE statement, see the section
“STORE Statement” on page 509 in Chapter 19, “Shared Concepts and Topics.”
The STORE statement is not available for models created with the LINK=ALOGIT option, the EQUAL-
SLOPES or UNEQUALSLOPES options, a STRATA statement, or an EXACT statement.
STRATA Statement F 5453
STRATA Statement
STRATA variable < (option) > . . . < variable < (option) > > < / options > ;
The STRATA statement names the variables that define strata or matched sets to use in stratified logistic
regression of binary response data.
Observations that have the same variable values are in the same matched set. For a stratified logistic model,
you can analyze 1:1, 1:n, m:n, and general mi :ni matched sets where the number of cases and controls
varies across strata. At least one variable must be specified to invoke the stratified analysis, and the usual
unconditional asymptotic analysis is not performed. The stratified logistic model has the form
logit.hi / D ˛h C x0hi ˇ
where hi is the event probability for the ith observation in stratum h with covariates xhi and where the
stratum-specific intercepts ˛h are the nuisance parameters that are to be conditioned out.
STRATA variables can also be specified in the MODEL statement as classification or continuous covariates;
however, the effects are nondegenerate only when crossed with a nonstratification variable. Specifying several
STRATA statements is the same as specifying one STRATA statement that contains all the strata variables.
The STRATA variables can be either character or numeric, and the formatted values of the STRATA variables
determine the levels. Thus, you can also use formats to group values into levels; see the discussion of the
FORMAT procedure in the Base SAS Procedures Guide.
The “Strata Summary” table is displayed by default. For an exact logistic regression, it displays the number
of strata that have a specific number of events and non-events. For example, if you are analyzing a 1:5
matched study, this table enables you to verify that every stratum in the analysis has exactly one event and
five non-events. Strata that contain only events or only non-events are reported in this table, but such strata
are uninformative and are not used in the analysis.
If an EXACT statement is also specified, then a stratified exact logistic regression is performed.
The EFFECTPLOT, SCORE, and WEIGHT statements are not available with a STRATA statement. The
following MODEL options are also not supported with a STRATA statement: CLPARM=PL, CLODDS=PL,
CTABLE, FIRTH, LACKFIT, LINK=, NOFIT, OUTMODEL=, OUTROC=, ROC, and SCALE=.
The following option can be specified for a stratification variable by enclosing the option in parentheses after
the variable name, or it can be specified globally for all STRATA variables after a slash (/).
MISSING
treats missing values (‘.’, ._, .A, . . . , .Z for numeric variables and blanks for character variables) as
valid STRATA variable values.
The following strata options are also available after the slash:
CHECKDEPENDENCY | CHECK=keyword
specifies which variables are to be tested for dependency before the analysis is performed. The available
keywords are as follows:
NONE performs no dependence checking. Typically, a message about a singular information matrix
is displayed if you have dependent variables. Dependent variables can be identified after the
analysis by noting any missing parameter estimates.
5454 F Chapter 73: The LOGISTIC Procedure
COVARIATES checks dependence between covariates and an added intercept. Dependent covariates
are removed from the analysis. However, covariates that are linear functions of the strata
variable might not be removed, which results in a singular information matrix message
being displayed in the SAS log. This is the default.
ALL checks dependence between all the strata and covariates. This option can adversely affect
performance if you have a large number of strata.
NOSUMMARY
suppresses the display of the “Strata Summary” table.
INFO
displays the “Strata Information” table, which includes the stratum number, levels of the STRATA
variables that define the stratum, the number of events, the number of non-events, and the total
frequency for each stratum. Since the number of strata can be very large, this table is displayed only
by request.
TEST Statement
< label: > TEST equation1 < , equation2, . . . > < / option > ;
The TEST statement tests linear hypotheses about the regression coefficients. The Wald test is used to
perform a joint test of the null hypotheses H0 W Lˇ D c specified in a single TEST statement, where ˇ is the
vector of intercept and slope parameters. When c D 0 you should specify a CONTRAST statement instead.
Each equation specifies a linear hypothesis (a row of the L matrix and the corresponding element of the c
vector). Multiple equations are separated by commas. The label , which must be a valid SAS name, is used
to identify the resulting output and should always be included. You can submit multiple TEST statements.
The form of an equation is as follows:
term< ˙ term : : : > < D ˙term < ˙term: : : > >
where term is a parameter of the model, or a constant, or a constant times a parameter. Intercept and CLASS
variable parameter names should be specified as described in the section “Parameter Names in the OUTEST=
Data Set” on page 5500. For multinomial response models, this form enables you to construct tests of
parameters from specific logits. When no equal sign appears, the expression is set to 0. The following
statements illustrate possible uses of the TEST statement:
proc logistic;
model y= a1 a2 a3 a4;
test1: test intercept + .5 * a2 = 0;
test2: test intercept + .5 * a2;
test3: test a1=a2=a3;
test4: test a1=a2, a2=a3;
run;
Note that the first and second TEST statements are equivalent, as are the third and fourth TEST statements.
You can specify the following option in the TEST statement after a slash(/):
UNITS Statement F 5455
PRINT
displays intermediate calculations in the testing of the null hypothesis H0 W Lˇ D c. These calculations
include Lb V.b̌/L0 bordered by .Lb̌ c/ and ŒLb V.b̌/L0 1 bordered by ŒLb V.b̌/L0 1 .Lb̌ c/, where
b̌ is the maximum likelihood estimator of ˇ and b V.b̌/ is the estimated covariance matrix of b̌.
For more information, see the section “Testing Linear Hypotheses about the Regression Coefficients”
on page 5483.
UNITS Statement
UNITS < independent1=list1 < independent2=list2: : : > > < / option > ;
The UNITS statement enables you to specify units of change for the continuous explanatory variables so
that customized odds ratios can be estimated. If you specify more than one UNITS statement, only the last
one is used. An estimate of the corresponding odds ratio is produced for each unit of change specified for
an explanatory variable. The UNITS statement is ignored for CLASS variables. Odds ratios are computed
only for main effects that are not involved in interactions or nestings, unless an ODDSRATIO statement is
also specified. If the CLODDS= option is specified in the MODEL statement, the corresponding confidence
limits for the odds ratios are also displayed, as are odds ratios and confidence limits for any CLASS main
effects that are not involved in interactions or nestings. The CLASS effects must use the GLM, reference, or
effect coding.
The UNITS statement also enables you to customize the odds ratios for effects specified in ODDSRATIO
statements, in which case interactions and nestings are allowed, and CLASS variables can be specified with
any parameterization.
The term independent is the name of an explanatory variable and list represents a list of units of change,
separated by spaces, that are of interest for that variable. Each unit of change in a list has one of the following
forms:
number
SD or –SD
number * SD
where number is any nonzero number, and SD is the sample standard deviation of the corresponding
independent variable. For example, X D 2 requests an odds ratio that represents the change in the odds
when the variable X is decreased by two units. X D 2SD requests an estimate of the change in the odds
when X is increased by two sample standard deviations.
You can specify the following option in the UNITS statement after a slash(/):
DEFAULT=list
gives a list of units of change for all explanatory variables that are not specified in the UNITS statement.
Each unit of change can be in any of the forms described previously. If the DEFAULT= option is not
specified, PROC LOGISTIC does not produce customized odds ratio estimates for any continuous
explanatory variable that is not listed in the UNITS statement. For more information, see the section
“Odds Ratio Estimation” on page 5470.
5456 F Chapter 73: The LOGISTIC Procedure
WEIGHT Statement
WEIGHT variable < / option > ;
When a WEIGHT statement appears, each observation in the input data set is weighted by the value of the
WEIGHT variable. Unlike a FREQ variable, the values of the WEIGHT variable can be nonintegral and are
not truncated. Observations with negative, zero, or missing values for the WEIGHT variable are not used
in the model fitting. When the WEIGHT statement is not specified, each observation is assigned a weight
of 1. The WEIGHT statement is not available with the STRATA statement. If you specify more than one
WEIGHT statement, then the first WEIGHT variable is used.
If a SCORE statement is specified, then the WEIGHT variable is used for computing fit statistics (see the
section “Fit Statistics for Scored Data Sets” on page 5488) and the ROC curve (see the WEIGHTED option
of the ROCOPTIONS option), but it is not required for scoring. Weights do not affect the computation of
predicted probabilities, their confidence limits, or the predicted response level. If the DATA= data set in the
SCORE statement does not contain the WEIGHT variable, the weights are assumed to be 1 and a warning
message is issued in the SAS log. If you fit a model and perform the scoring in the same run, the same
WEIGHT variable is used for fitting and scoring. If you fit a model in a previous run and input it with the
INMODEL= option in the current run, then the WEIGHT variable can be different from the one used in the
previous run; however, if a WEIGHT variable was not specified in the previous run, you can still specify a
WEIGHT variable in the current run.
C AUTION : PROC LOGISTIC does not compute the proper variance estimators if you are analyzing survey
data and specifying the sampling weights through the WEIGHT statement. The SURVEYLOGISTIC
procedure is designed to perform the necessary, and correct, computations.
The following option can be added to the WEIGHT statement after a slash (/):
NORMALIZE
NORM
causes the weights specified by the WEIGHT variable to be normalized so that they add up to the
actual sample size. Weights wi are normalized by multiplying them by Pn n w , where n is the sample
iD1 i
size. With this option, the estimated covariance matrix of the parameter estimators is invariant to the
scale of the WEIGHT variable.
Missing Values
Any observation with missing values for the response, offset, strata, or explanatory variables is excluded
from the analysis; however, missing values are valid for variables specified with the MISSING option in the
CLASS or STRATA statement. Observations with a nonpositive or missing weight or with a frequency less
than 1 are also excluded. The estimated linear predictor and its standard error estimate, the fitted probabilities
and confidence limits, and the regression diagnostic statistics are not computed for any observation with
missing offset or explanatory variable values. However, if only the response value is missing, the linear
Response Level Ordering F 5457
predictor, its standard error, the fitted individual and cumulative probabilities, and confidence limits for the
cumulative probabilities can be computed and output to a data set by using the OUTPUT statement.
logit.Pr.Y i jx//; i D 1; : : : ; k
and for the generalized logit model the functions modeled are
Pr.Y D i jx/
log ; i D 1; : : : ; k
Pr.Y D k C 1jx/
where the highest Ordered Value Y D k C 1 is the reference level. You can change which probabilities are
modeled by specifying the EVENT=, REF=, DESCENDING, or ORDER= response variable options in the
MODEL statement.
For binary response data with event and nonevent categories, if your event category has a higher Ordered
Value, then by default the nonevent is modeled. Because the default response function modeled is
logit./ D log
1
where is the probability of the response level assigned Ordered Value 1, and because
logit./ D logit.1 /
the effect of modeling the nonevent is to change the signs of ˛ and ˇ in the model for the event, logit./ D
˛ C ˇ 0 x.
For example, suppose the binary response variable Y takes the values 1 and 0 for event and nonevent,
respectively, and Exposure is the explanatory variable. By default, PROC LOGISTIC assigns Ordered Value
1 to response level Y=0, and Ordered Value 2 to response level Y=1. As a result, PROC LOGISTIC models
the probability of the nonevent (Ordered Value=1) category, and your parameter estimates have the opposite
sign from those in the model for the event. To model the event without using a DATA step to change the
values of the variable Y, you can control the ordering of the response levels or select the event or reference
level, as shown in the following list:
Explicitly state which response level is to be modeled by using the response variable option EVENT=
in the MODEL statement:
Specify the nonevent category for the response variable in the response variable option REF= in the
MODEL statement. This option is most useful for generalized logit models where the EVENT= option
cannot be used.
5458 F Chapter 73: The LOGISTIC Procedure
Specify the response variable option DESCENDING in the MODEL statement to assign the lowest
Ordered Value to Y=1:
model Y(descending)=Exposure;
Assign a format to Y such that the first formatted value (when the formatted values are put in sorted
order) corresponds to the event. In the following example, Y=1 is assigned the formatted value ‘event’
and Y=0 is assigned the formatted value ‘nonevent’. Because ORDER=FORMATTED by default,
Ordered Value 1 is assigned to response level Y=1, so the procedure models the event.
proc format;
value Disease 1='event' 0='nonevent';
run;
proc logistic;
format Y Disease.;
model Y=Exposure;
run;
Traditionally, the probit function contains the additive constant 5, but throughout PROC LOGISTIC,
the terms probit and normit are used interchangeably.
Determining Observations for Likelihood Contributions F 5459
is the inverse of the cumulative extreme-value function (also called the Gompertz distribution), which
is
The generalized logit function extends the binary logit link to a vector of levels .p1 ; : : : ; pkC1 / by
contrasting each level with a fixed level:
The adjacent-category logit function extends the binary logit link to a vector of levels .p1 ; : : : ; pkC1 /
by contrasting each level with the following level:
The variances of the normal, logistic, and extreme-value distributions are not the same. Their respective
means and variances are shown in the following table:
Here
is the Euler constant. In comparing parameter estimates from different link functions, you need to
take into account the different scalings of the corresponding distributions and, for the complementary log-log
function, a possible shift in location. For example, if the fitted probabilities are inpthe neighborhood of 0.1 to
0.9, then the parameter estimates from the logit link function should be about = 3 larger than the estimates
from the probit link function.
Suppose the response variable in a cumulative response model can take on the ordered values 1; : : : ; k; k C 1,
where k is an integer 1. The likelihood for the jth observation with ordered response value yj and
explanatory variables vector xj is given by
< F .˛1 C ˇ 0 xj /
8
yj D 1
Lj D F .˛i C ˇ 0 xj / F .˛i 1 C ˇ 0x / 1 < y D i k
j j
1 F .˛k C ˇ 0 xj / yj D k C 1
:
where F ./ is the logistic, normal, or extreme-value distribution function, ˛1 ; : : : ; ˛k are ordered intercept
parameters, and ˇ is the common slope parameter vector.
For the adjacent-category logit model, the likelihood for the jth observation with ordered response value yj
and explanatory variables vector xj is given by
8 Pk 0
i Dyj ˛i Cxj ˇ
ˆ
ˆ e
ˆ
ˆ
Pk Pk 0
1 yj k
iDm ˛i Cxj ˇ
<
Lj D Pr.Y D yj jxj / D 1 C mD1 e
ˆ 1
yj D k C 1
ˆ
ˆ
ˆ Pk Pk 0
1 C mD1 e iDm ˛i Cxj ˇ
:
With ij denoting the probability that the jth observation has response value i, the expected value of Zj
Pk
is j D .1j ; : : : ; kC1;j /0 where kC1;j D 1 i D1 ij . The covariance matrix of Zj is Vj , which is
the covariance matrix of a multinomial random variable for one trial with parameter vector j . Let ˇ be
the vector of regression parameters; in other words, ˇ D .˛1 ; : : : ; ˛k ; ˇ1 ; : : : ; ˇs /0 . Let Dj be the matrix of
partial derivatives of j with respect to ˇ. The estimating equation for the regression parameters is
X
D0j Wj .Zj j / D 0
j
where Wj D wj fj Vj , wj and fj are the weight and frequency of the jth observation, and Vj is a
generalized inverse of Vj . PROC LOGISTIC chooses Vj as the inverse of the diagonal matrix with j as
the diagonal.
With a starting value of ˇ .0/ , the maximum likelihood estimate of ˇ is obtained iteratively as
X X
ˇ .mC1/ D ˇ .m/ C . D0j Wj Dj / 1 D0j Wj .Zj j /
j j
where Dj , Wj , and j are evaluated at ˇ .m/ . The expression after the plus sign is the step size. If the
likelihood evaluated at ˇ .mC1/ is less than that evaluated at ˇ .m/ , then ˇ .mC1/ is recomputed by step-halving
or ridging as determined by the value of the RIDGING= option. The iterative scheme continues until
convergence is obtained—that is, until ˇ .mC1/ is sufficiently close to ˇ .m/ . Then the maximum likelihood
estimate of ˇ is b̌ D ˇ .mC1/ .
The covariance matrix of b̌ is estimated by
b
Cov.b̌/ D .
X
D0j W
b bjb Dj / 1 D b I 1
j
Dj and W
where b b j are, respectively, Dj and Wj evaluated at b̌. b
I is the information matrix, or the negative
expected Hessian matrix, evaluated at b̌.
By default, starting values are zero for the slope parameters, and for the intercept parameters, starting
values are the observed cumulative logits (that is, logits of the observed cumulative proportions of response).
Alternatively, the starting values can be specified with the INEST= option.
Newton-Raphson Algorithm
For cumulative models, let the parameter vector be ˇ D .˛1 ; : : : ; ˛k ; ˇ1 ; : : : ; ˇs /0 , and for the generalized
logit model let ˇ D .˛1 ; : : : ; ˛k ; ˇ10 ; : : : ; ˇk0 /0 . The gradient vector and the Hessian matrix are given,
5462 F Chapter 73: The LOGISTIC Procedure
respectively, by
X @lj
g D wj fj
@ˇ
j
X @2 lj
H D wj fj
@ˇ 2
j
where lj D log Lj is the log likelihood for the jth observation. With a starting value of ˇ .0/ , the maximum
likelihood estimate b̌ of ˇ is obtained iteratively until convergence is obtained:
ˇ .mC1/ D ˇ .m/ H 1
g
where H and g are evaluated at ˇ .m/ . If the likelihood evaluated at ˇ .mC1/ is less than that evaluated at ˇ .m/ ,
then ˇ .mC1/ is recomputed by step-halving or ridging.
The covariance matrix of b̌ is estimated by
b
Cov.b̌/ D b
I 1
ID
where the observed information matrix b H
b is computed by evaluating H at b̌.
where p is the number of parameters in the model, with the modified score equation
n
X
g.ˇj / D fyi i C hi .0:5 i /gxij D 0 .j D 1; : : : ; p/
i D1
where the hi s are the ith diagonal elements of the hat matrix W1=2 X.X0 WX/ 1 X0 W1=2 and W D
diagfi .1 i /g. The Hessian matrix is not modified by this penalty, and the optimization method is
performed in the usual manner.
Convergence Criteria
Four convergence criteria are available: ABSFCONV=, FCONV=, GCONV=, and XCONV=. If you specify
more than one convergence criterion, the optimization is terminated as soon as one of the criteria is satisfied.
If none of the criteria is specified, the default is GCONV=1E–8.
If you specify a STRATA statement or the EQUALSLOPES or UNEQUALSLOPES options in the MODEL
statement, all unspecified (or nondefault) criteria are also compared to zero. For example, specifying only the
criterion XCONV=1E–8 but attaining FCONV=0 terminates the optimization even if the XCONV= criterion
is not satisfied, because the log likelihood has reached its maximum. More convergence criteria are also
available; for more information, see the section “NLOPTIONS Statement” on page 5441.
Existence of Maximum Likelihood Estimates F 5463
Complete Separation There is a complete separation of data points if there exists a vector b that correctly
allocates all observations to their response groups; that is,
0
b xj > 0 Yj D 1
b0 xj < 0 Yj D 2
This configuration gives nonunique infinite estimates. If the iterative process of maximizing the
likelihood function is allowed to continue, the log likelihood diminishes to zero, and the dispersion
matrix becomes unbounded.
Quasi-complete Separation The data are not completely separable, but there is a vector b such that
0
b xj 0 Yj D 1
b0 xj 0 Yj D 2
and equality holds for at least one subject in each response group. This configuration also yields
nonunique infinite estimates. If the iterative process of maximizing the likelihood function is
allowed to continue, the dispersion matrix becomes unbounded and the log likelihood diminishes
to a nonzero constant.
Overlap If neither complete nor quasi-complete separation exists in the sample points, there is an overlap
of sample points. In this configuration, the maximum likelihood estimates exist and are unique.
Complete separation and quasi-complete separation are problems typically encountered with small data sets.
Although complete separation can occur with any type of data, quasi-complete separation is not likely with
truly continuous explanatory variables.
The LOGISTIC procedure uses a simple empirical approach to recognize the data configurations that lead
to infinite parameter estimates. The basis of this approach is that any convergence method of maximizing
the log likelihood must yield a solution giving complete separation, if such a solution exists. In maximizing
the log likelihood, there is no checking for complete or quasi-complete separation if convergence is attained
in eight or fewer iterations. Subsequent to the eighth iteration, the probability of the observed response is
computed for each observation. If the predicted response equals the observed response for every observation,
there is a complete separation of data points and the iteration process is stopped. If the complete separation of
data has not been determined and an observation is identified to have an extremely large probability (0.95)
of predicting the observed response, there are two possible situations. First, there is overlap in the data set,
and the observation is an atypical observation of its own group. The iterative process, if allowed to continue,
will stop when a maximum is reached. Second, there is quasi-complete separation in the data set, and the
5464 F Chapter 73: The LOGISTIC Procedure
asymptotic dispersion matrix is unbounded. If any of the diagonal elements of the dispersion matrix for the
standardized observations vectors (all explanatory variables standardized to zero mean and unit variance)
exceeds 5000, quasi-complete separation is declared and the iterative process is stopped. If either complete
separation or quasi-complete separation is detected, a warning message is displayed in the procedure output.
Checking for quasi-complete separation is less foolproof than checking for complete separation. The
NOCHECK option in the MODEL statement turns off the process of checking for infinite parameter
estimates. In cases of complete or quasi-complete separation, turning off the checking process typically
results in the procedure failing to converge.
To address the separation issue, you can change your model, specify the FIRTH option to use Firth’s penalized
likelihood method, or for small data sets specify an EXACT statement to perform an exact logistic regression.
Effect-Selection Methods
Five effect-selection methods are available by specifying the SELECTION= option in the MODEL statement.
The simplest method (and the default) is SELECTION=NONE, for which PROC LOGISTIC fits the complete
model as specified in the MODEL statement. The other four methods are FORWARD for forward selection,
BACKWARD for backward elimination, STEPWISE for stepwise selection, and SCORE for best subsets
selection. Intercept parameters are forced to stay in the model unless the NOINT option is specified.
When SELECTION=FORWARD, PROC LOGISTIC first estimates parameters for effects forced into the
model. These effects are the intercepts and the first n explanatory effects in the MODEL statement, where n
is the number specified by the START= or INCLUDE= option in the MODEL statement (n is zero by default).
Next, the procedure computes the score chi-square statistic for each effect not in the model and examines the
largest of these statistics. If it is significant at the SLENTRY= level, the corresponding effect is added to the
model. Once an effect is entered in the model, it is never removed from the model. The process is repeated
until none of the remaining effects meet the specified level for entry or until the STOP= value is reached. For
more information, see the section “Score Statistics and Tests” on page 5467.
When SELECTION=BACKWARD, parameters for the complete model as specified in the MODEL statement
are estimated unless the START= option is specified. In that case, only the parameters for the intercepts and
the first n explanatory effects in the MODEL statement are estimated, where n is the number specified by
the START= option. Results of the Wald test for individual parameters are examined. The least significant
effect that does not meet the SLSTAY= level for staying in the model is removed. Once an effect is removed
from the model, it remains excluded. The process is repeated until no other effect in the model meets the
specified level for removal or until the STOP= value is reached. Backward selection is often less successful
than forward or stepwise selection because the full model fit in the first step is the model most likely to
result in a complete or quasi-complete separation of response values as described in the section “Existence
of Maximum Likelihood Estimates” on page 5463. For more information, see the section “Testing Linear
Hypotheses about the Regression Coefficients” on page 5483.
The SELECTION=STEPWISE option is similar to the SELECTION=FORWARD option except that effects
already in the model do not necessarily remain. Effects are entered into and removed from the model in such
a way that each forward selection step can be followed by one or more backward elimination steps. The
stepwise selection process terminates if no further effect can be added to the model or if the current model is
identical to a previously visited model.
For SELECTION=SCORE, PROC LOGISTIC uses the branch-and-bound algorithm of Furnival and Wilson
(1974) adapted to find a specified number of models with the highest score chi-square statistic for all possible
Model Fitting Information F 5465
model sizes, from 1, 2, 3 effect models, and so on, up to the single model containing all of the explanatory
effects. The number of models displayed for each model size is controlled by the BEST= option. You can
use the START= option to impose a minimum model size, and you can use the STOP= option to impose a
maximum model size. For instance, with BEST=3, START=2, and STOP=5, the SCORE selection method
displays the best three models (that is, the three models with the highest score chi-squares) containing 2, 3, 4,
and 5 effects. The SELECTION=SCORE option is not available for models with CLASS variables. For more
information, see the section “Score Statistics and Tests” on page 5467.
The options FAST, SEQUENTIAL, and STOPRES can alter the default criteria for entering or removing
effects from the model when they are used with the FORWARD, BACKWARD, or STEPWISE selection
method.
–2 log likelihood:
X wj
2 Log L D 2 fj log.b
j /
2
j
where wj and fj are the weight and frequency values of the jth observation, and 2 is the dispersion
parameter, which equals 1 unless the SCALE= option is specified. For binary response models that use
events/trials MODEL statement syntax, this is
!
X wj nj
2 Log L D 2 fj Œlog C rj log.b
j / C .nj rj / log.1 b j /
2 rj
j
where rj is the number of events, nj is the number of trials, b j is the estimated event probability, and
the statistic is reported both with and without the constant term.
AIC D 2 Log L C 2p
where p is the number of parameters in the model. For cumulative response models, p D k C s, where
k is the total number of response levels minus one and s is the number of explanatory effects. For the
generalized logit model, p D k.s C 1/.
where p is the number of parameters in the model, nj is the number of trials when events/trials syntax
is specified, and nj D 1 with single-trial syntax.
5466 F Chapter 73: The LOGISTIC Procedure
The AIC and SC statistics give two different ways of adjusting the –2 Log L statistic for the number of terms
in the model and the number of observations used. These statistics can be used when comparing different
models for the same data (for example, when you use the SELECTION=STEPWISE option in the MODEL
statement). The models being compared do not have to be nested; lower values of the statistics indicate a
more desirable model.
The difference in the –2 Log L statistics between the intercepts-only model and the specified model has a
p k degree-of-freedom chi-square distribution under the null hypothesis that all the explanatory effects
in the model are zero, where p is the number of parameters in the specified model and k is the number of
intercepts. The likelihood ratio test in the “Testing Global Null Hypothesis: BETA=0” table displays this
difference and the associated p-value for this statistic. The score and Wald tests in that table test the same
hypothesis and are asymptotically equivalent; for more information, see the sections “Residual Chi-Square”
on page 5467 and “Testing Linear Hypotheses about the Regression Coefficients” on page 5483.
PL.0/ is the likelihood of the intercept-only model, L. / is the likelihood of the specified model,
where b̌
n D j fj nj is the sample size, fj is the frequency of the jth observation, and nj is the number of trials
when events/trials syntax is specified or nj D 1 with single-trial syntax.
The quantity R2 achieves a maximum of less than one for discrete models, where the maximum is given by
2
2
Rmax D1 fL.0/g n
Nagelkerke (1991) proposes the following adjusted coefficient, which can achieve a maximum value of one:
2 R2
RN D 2
Rmax
Specifying the NORMALIZE option in the WEIGHT statement makes these coefficients invariant to the
scale of the weights.
Like the AIC and SC statistics described in the section “Model Fitting Information” on page 5465, R2 and
2
RN are most useful for comparing competing models that are not necessarily nested—larger values indicate
2
better models. More properties and interpretation of R2 and RN are provided in Nagelkerke (1991). In
2 2
the “Testing Global Null Hypothesis: BETA=0” table, R is labeled as “RSquare” and RN is labeled as
2 2
“Max-rescaled RSquare.” Use the RSQUARE option to request R and RN .
Score Statistics and Tests F 5467
g0 .b̌H0 /I 1
.b̌H0 /g.b̌H0 /
and it has an asymptotic 2 distribution with r degrees of freedom under H0 , where r is the number of
restrictions imposed on ˇ by H0 .
Score statistics are used when performing forward, stepwise and score selection; for more information see
the section “Effect-Selection Methods” on page 5464.
Residual Chi-Square
When you use SELECTION=FORWARD, BACKWARD, or STEPWISE, the procedure calculates a residual
chi-square score statistic and reports the statistic, its degrees of freedom, and the p-value. This section
describes how the statistic is calculated.
Suppose there are s explanatory effects of interest. The full cumulative response model has a parameter
vector
ˇ D .˛1 ; : : : ; ˛k ; ˇ1 ; : : : ; ˇs /0
where ˛1 ; : : : ; ˛k are intercept parameters, and ˇ1 ; : : : ; ˇs are the common slope parameters for the s
explanatory effects. The full generalized logit model has a parameter vector
where ˇij is the slope parameter for the jth effect in the ith logit.
5468 F Chapter 73: The LOGISTIC Procedure
Consider the null hypothesis H0 W ˇt C1 D D ˇs D 0, where t < s for the cumulative response model, and
H0 W ˇi;t C1 D D ˇi s D 0; t < s; i D 1; : : : ; k, for the generalized logit model. For the reduced model
with t explanatory effects, let b˛1; : : : ; b
˛ k be the MLEs of the unknown intercept parameters, let b ˇ1; : : : ; b
ˇt
0
be the MLEs of the unknown slope parameters, and let i.t / D .ˇ i1 ; : : : ; ˇ i t /; i D 1; : : : ; k, be those for the
b̌ b b
generalized logit model. The residual chi-square is the chi-square score statistic testing the null hypothesis
H0 ; that is, the residual chi-square is
g0 .b̌H0 /I 1
.b̌H0 /g.b̌H0 /
where for the cumulative response model b̌H0 D .b ˛1; : : : ; b˛k ; b ˇ t ; 0; : : : ; 0/0 , and for the generalized
ˇ1; : : : ; b
0 0 0 0
logit model b̌H0 D .b ˛ k ; b̌1.t / ; 0.s t / ; : : : b̌k.t / ; 0.s t / /0 , where 0.s t / denotes a vector of s t
˛1; : : : ; b
zeros.
The residual chi-square has an asymptotic chi-square distribution with s t degrees of freedom (k.s t / for
the generalized logit model). A special case is the global score chi-square, where the reduced model consists
of the k intercepts and no explanatory effects. The global score statistic is displayed in the “Testing Global
Null Hypothesis: BETA=0” table. The table is not produced when the NOFIT option is used, but the global
score statistic is displayed.
where g./ is the link function, and ˇi D .˛i ; ˇi1 ; : : : ; ˇi s /0 is a vector of unknown parameters consisting of
Confidence Intervals for Parameters F 5469
an intercept ˛i and s slope parameters ˇi1 ; : : : ; ˇi s . The parameter vector for this general cumulative model
is
ˇ D .ˇ10 ; : : : ; ˇk0 /0
Under the null hypothesis of parallelism H0 W ˇ1m D ˇ2m D D ˇkm ; 1 m s, there is a single
common slope parameter for each of the s explanatory variables. Let ˇ1 ; : : : ; ˇs be the common slope
˛1; : : : ; b
parameters. Let b ˇ1; : : : ; b
˛ k and b ˇ s be the MLEs of the intercept parameters and the common slope
parameters. Then, under H0 , the MLE of ˇ is
and the chi-square score statistic g0 .b̌H0 /I 1 .b̌H0 /g.b̌H0 / has an asymptotic chi-square distribution with
s.k 1/ degrees of freedom. This tests the parallel lines assumption by testing the equality of separate slope
parameters simultaneously for all explanatory variables.
where Bj .
/ is the set of all ˇ with the jth element fixed at
, and l.ˇ/ is the log-likelihood function for ˇ.
(The penalized log-likelihood function is used when you specify the FIRTH option in the MODEL statement.)
If lmax D l.b̌/ is the log likelihood evaluated at the maximum likelihood estimate b̌, then 2.lmax lj .ˇj //
has a limiting chi-square distribution with one degree of freedom if ˇj is the true parameter value. Let
l0 D lmax 0:521 .1 ˛/, where 21 .1 ˛/ is the 100.1 ˛/th percentile of the chi-square distribution with
one degree of freedom. A 100.1 ˛/% confidence interval for ˇj is
f
W lj .
/ l0 g
The endpoints of the confidence interval are found by solving numerically for values of ˇj that satisfy
equality in the preceding relation. To obtain an iterative algorithm for computing the confidence limits, the
log-likelihood function in a neighborhood of ˇ is approximated by the quadratic function
Q C ı/ D l.ˇ/ C ı 0 g C 1 ı 0 Vı
l.ˇ
2
5470 F Chapter 73: The LOGISTIC Procedure
where g D g.ˇ/ is the gradient vector and V D V.ˇ/ is the Hessian matrix. The increment ı for the next
iteration is obtained by solving the likelihood equations
d Q
fl.ˇ C ı/ C .e0j ı
/g D 0
dı
where is the Lagrange multiplier, ej is the jth unit vector, and
is an unknown constant. The solution is
1
ıD V .g C ej /
Q C ı/ D l0 , you can estimate as
By substituting this ı into the equation l.ˇ
1 21
2.l0 l.ˇ/ C 12 g0 V g/
D˙ 1
e0j V ej
The upper confidence limit for ˇj is computed by starting at the maximum likelihood estimate of ˇ and
iterating with positive values of until convergence is attained. The process is repeated for the lower
confidence limit by using negative values of .
Convergence is controlled by the value specified with the PLCONV= option in the MODEL statement (the
default value of is 1E–4). Convergence is declared on the current iteration if the following two conditions
are satisfied:
jl.ˇ/ l0 j
and
.g C ej /0 V 1
.g C ej /
ˇ j ˙ z1
b j
˛=2b
The odds ratio is defined as the ratio of the odds for those with the risk factor (X = 1) to the odds for those
without the risk factor (X = 0). The log of the odds ratio is given by
log. / log. .X D 1; X D 0// D logit.X D 1/ logit.X D 0/ D .˛ C 1 ˇ/ .˛ C 0 ˇ/ D ˇ
In general, the odds ratio can be computed by exponentiating the difference of the logits between any two
population profiles. This is the approach taken by the ODDSRATIO statement, so the computations are
available regardless of parameterization, interactions, and nestings. However, as shown in the preceding
equation for log. /, odds ratios of main effects can be computed as functions of the parameter estimates,
and the remainder of this section is concerned with this methodology.
The parameter, ˇ, associated with X represents the change in the log odds from X D 0 to X D 1. So the
odds ratio is obtained by simply exponentiating the value of the parameter associated with the risk factor.
The odds ratio indicates how the odds of the event change as you change X from 0 to 1. For instance, D 2
means that the odds of an event when X = 1 are twice the odds of an event when X = 0. You can also express
this as follows: the percent change in the odds of an event from X = 0 to X = 1 is . 1/100% D 100%.
Suppose the values of the dichotomous risk factor are coded as constants a and b instead of 0 and 1. The
odds when X D a become exp.˛ C aˇ/, and the odds when X D b become exp.˛ C bˇ/. The odds ratio
corresponding to an increase in X from a to b is
D expŒ.b a/ˇ D Œexp.ˇ/b a
Œexp.ˇ/c
Note that for any a and b such that c D b a D 1; D exp.ˇ/. So the odds ratio can be interpreted as the
change in the odds for any increase of one unit in the corresponding risk factor. However, the change in
odds for some amount other than one unit is often of greater interest. For example, a change of one pound in
body weight might be too small to be considered important, while a change of 10 pounds might be more
meaningful. The odds ratio for a change in X from a to b is estimated by raising the odds ratio estimate for a
unit change in X to the power of c D b a as shown previously.
For a polytomous risk factor, the computation of odds ratios depends on how the risk factor is parameterized.
For illustration, suppose that Race is a risk factor with four categories: White, Black, Hispanic, and Other.
For the effect parameterization scheme (PARAM=EFFECT) with White as the reference group
(REF=’White’), the design variables for Race are as follows:
Design Variables
Race X1 X2 X3
Black 1 0 0
Hispanic 0 1 0
Other 0 0 1
White –1 –1 –1
The log odds for Black is
For the reference cell parameterization scheme (PARAM=REF) with White as the reference cell, the design
variables for race are as follows:
Design Variables
Race X1 X2 X3
Black 1 0 0
Hispanic 0 1 0
Other 0 0 1
White 0 0 0
The log odds ratio of Black versus White is given by
For the GLM parameterization scheme (PARAM=GLM), the design variables are as follows:
Design Variables
Race X1 X2 X3 X4
Black 1 0 0 0
Hispanic 0 1 0 0
Other 0 0 1 0
White 0 0 0 1
The log odds ratio of Black versus White is
Consider the hypothetical example of heart disease among race in Hosmer and Lemeshow (2000, p. 56). The
entries in the following contingency table represent counts:
Race
Disease Status White Black Hispanic Other
Present 5 20 15 10
Absent 20 10 10 10
Rank Correlation of Observed Responses and Predicted Probabilities F 5473
The computation of odds ratio of Black versus White for various parameterization schemes is tabulated in
Table 73.11.
Parameter Estimates
PARAM= ˇ1
b bˇ2 ˇ3
b ˇ4
b Odds Ratio Estimates
EFFECT 0.7651 0.4774 0.0719 exp.2 0:7651 C 0:4774 C 0:0719/ D 8
REF 2.0794 1.7917 1.3863 exp.2:0794/ D 8
GLM 2.0794 1.7917 1.3863 0.0000 exp.2:0794/ D 8
Because the log odds ratio (log. /) is a linear function of the parameters, the Wald confidence interval
for log. / can be derived from the parameter estimates and the estimated covariance matrix. Confidence
intervals for the odds ratios are obtained by exponentiating the corresponding confidence limits for the log
odd ratios. In the displayed output of PROC LOGISTIC, the “Odds Ratio Estimates” table contains the odds
ratio estimates and the corresponding 95% Wald confidence intervals. For continuous explanatory variables,
these odds ratios correspond to a unit increase in the risk factors.
To customize odds ratios for specific units of change for a continuous risk factor, you can use the UNITS
statement to specify a list of relevant units for each explanatory variable in the model. Estimates of these
customized odds ratios are given in a separate table. Let .Lj ; Uj / be a confidence interval for log. /. The
corresponding lower and upper confidence limits for the customized odds ratio exp.cˇj / are exp.cLj /
and exp.cUj /, respectively (for c > 0), or exp.cUj / and exp.cLj /, respectively (for c < 0). You use the
CLODDS= option or ODDSRATIO statement to request the confidence intervals for the odds ratios.
For a generalized logit model, odds ratios are computed similarly, except k odds ratios are computed for each
effect, corresponding to the k logits in the model.
The CONTRAST statement also enables you to estimate the exponentiated contrast, e O i . The corresponding
standard error is e Oi O .O i /, and the confidence limits are computed by exponentiating those for the linear
predictor: expfO i ˙ z˛=2 O .O i /g.
Classification Table F 5475
@i 0
2 @i
.i / D V.ˇ/
@ˇ @ˇ
i ˙ z˛=2 O .b
b i /
Note that the contrast O i and exponentiated contrast e O i , their standard errors, and their confidence intervals
are computed in the same fashion as for the cumulative response models, replacing ˇ with ˇi .
Classification Table
For binary response data, the response is either an event or a nonevent. In PROC LOGISTIC, the response
with Ordered Value 1 is regarded as the event, and the response with Ordered Value 2 is the nonevent.
PROC LOGISTIC models the probability of the event. From the fitted model, a predicted event probability
can be computed for each observation. A method to compute a reduced-bias estimate of the predicted
probability is given in the section “Predicted Probability of an Event for Classification” on page 5476. If the
predicted event probability exceeds or equals some cutpoint value z 2 Œ0; 1, the observation is predicted to
be an event observation; otherwise, it is predicted as a nonevent. A 2 2 frequency table can be obtained
by cross-classifying the observed and predicted responses. The CTABLE option produces this table, and
the PPROB= option selects one or more cutpoints. Each cutpoint generates a classification table. If the
PEVENT= option is also specified, a classification table is produced for each combination of PEVENT= and
PPROB= values.
The accuracy of the classification is measured by its sensitivity (the ability to predict an event correctly) and
specificity (the ability to predict a nonevent correctly). Sensitivity is the proportion of event responses that
were predicted to be events. Specificity is the proportion of nonevent responses that were predicted to be
nonevents. PROC LOGISTIC also computes three other conditional probabilities: false positive rate, false
negative rate, and rate of correct classification. The false positive rate is the proportion of predicted event
responses that were observed as nonevents. The false negative rate is the proportion of predicted nonevent
responses that were observed as events. Given prior probabilities specified with the PEVENT= option, these
conditional probabilities can be computed as posterior probabilities by using Bayes’ theorem.
5476 F Chapter 73: The LOGISTIC Procedure
where
False Positive, False Negative, and Correct Classification Rates Using Bayes’ Theorem
Suppose n1 of n individuals experience an event, such as a disease. Let this group be denoted by C1 , and let
the group of the remaining n2 D n n1 individuals who do not have the disease be denoted by C2 . The jth
individual is classified as giving a positive response if the predicted probability of disease (b .j / ) is large.
.j / is the reduced-bias estimate based on the one-step approximation given in the preceding
The probability b
.j / z.
section. For a given cutpoint z, the jth individual is predicted to give a positive response if b
Let B denote the event that a subject has the disease, and let BN denote the event of not having the disease. Let
A denote the event that the subject responds positively, and let AN denote the event of responding negatively.
Results of the classification are represented by two conditional probabilities, Pr.AjB/ and Pr.AjB/, N where
N
Pr.AjB/ is the sensitivity and Pr.AjB/ is one minus the specificity.
These probabilities are given by
.j / z/
P
j 2C1 I.b
Pr.AjB/ D
n1
.j / z/
P
N D j 2C2 I.b
Pr.AjB/
n2
Bayes’ theorem is used to compute several rates of the classification. For a given prior probability Pr.B/ of
the disease, the false positive rate PF C , the false negative rate PF , and the correct classification rate PC
are given by Fleiss (1981, pp. 4–5) as follows:
N
Pr.AjB/Œ1 Pr.B/
N
PF C D Pr.BjA/ D
N C Pr.B/ŒPr.AjB/ Pr.AjB/
Pr.AjB/ N
N D Œ1 Pr.AjB/Pr.B/
PF D Pr.BjA/
N
1 Pr.AjB/ Pr.B/ŒPr.AjB/ Pr.AjB/ N
N A/
PC D Pr.BjA/ C Pr.Bj N D Pr.AjB/Pr.B/ C Pr.Aj
N B/Œ1
N Pr.B/
The prior probability Pr.B/ can be specified by the PEVENT= option. If the PEVENT= option is not
specified, the sample proportion of diseased individuals is used; that is, Pr.B/ D n1 =n. In such a case, the
false positive rate and the false negative rate reduce to
.j / z/
P
j 2C2 I.b
PF C D
.j / .j / z/
P P
j 2C1 I.b z/ C j 2C2 I.b
.j / < z/
P
j 2C1 I.b
PF D
.j / .j / < z/
P P
j 2C1 I.b < z/ C j 2C2 I.b
.j / .j / < z/
P P
j 2C1 I.b z/ C j 2C2 I.b
PC D
n
Note that for a stratified sampling situation in which n1 and n2 are chosen a priori, n1 =n is not a desirable
estimate of Pr.B/. For such situations, the PEVENT= option should be specified.
Overdispersion
For a correctly specified model, the Pearson chi-square statistic and the deviance, divided by their degrees of
freedom, should be approximately equal to one. When their values are much larger than one, the assumption
of binomial variability might not be valid and the data are said to exhibit overdispersion. Underdispersion,
which results in the ratios being less than one, occurs less often in practice.
When fitting a model, there are several problems that can cause the goodness-of-fit statistics to exceed their
degrees of freedom. Among these are such problems as outliers in the data, using the wrong link function,
omitting important terms from the model, and needing to transform some predictors. These problems should
be eliminated before proceeding to use the following methods to correct for overdispersion.
The Pearson chi-square statistic 2P and the deviance 2D are given by
m kC1
X X .rij ij /2
ni b
2P D
ni b
ij
i D1 j D1
m kC1
X X rij
2D D 2 rij log
ni b
ij
i D1 j D1
where m is the number of subpopulation profiles, k C 1 is the number of response levels, rij is the total
weight (sum of the product of the frequencies and the weights) associated with jth level responses in the
ith profile, ni D kC1
P
j D1 rij , and b
ij is the fitted probability for the jth level at the ith profile. Each of these
chi-square statistics has mk p degrees of freedom, where p is the number of parameters estimated. The
dispersion parameter is estimated by
8 2
< P =.mk p/ SCALE=PEARSON
c2 D
2 =.mk p/ SCALE=DEVIANCE
: D
.constant/2 SCALE=constant
In order for the Pearson statistic and the deviance to be distributed as chi-square, there must be sufficient
replication within the subpopulations. When this is not true, the data are sparse, and the p-values for these
statistics are not valid and should be ignored. Similarly, these statistics, divided by their degrees of freedom,
cannot serve as indicators of overdispersion. A large difference between the Pearson statistic and the deviance
provides some evidence that the data are too sparse to use either statistic.
You can use the AGGREGATE (or AGGREGATE=) option to define the subpopulation profiles. If you do not
specify this option, each observation is regarded as coming from a separate subpopulation. For events/trials
syntax, each observation represents n Bernoulli trials, where n is the value of the trials variable; for single-trial
syntax, each observation represents a single trial. Without the AGGREGATE (or AGGREGATE=) option,
the Pearson chi-square statistic and the deviance are calculated only for events/trials syntax.
Note that the parameter estimates are not changed by this method. However, their standard errors are adjusted
for overdispersion, affecting their significance tests.
Williams’ Method
Suppose that the data consist of n binomial observations. For the ith observation, let ri =ni be the observed
proportion and let xi be the associated vector of explanatory variables. Suppose that the response probability
for the ith observation is a random variable Pi with mean and variance
where pi is the probability of the event, and is a nonnegative but otherwise unknown scale parameter. Then
the mean and variance of ri are
Williams (1982) estimates the unknown parameter by equating the value of Pearson’s chi-square statistic
for the full model to its approximate expected value. Suppose wi is the weight associated with the ith
The Hosmer-Lemeshow Goodness-of-Fit Test F 5479
Let g 0 ./ be the first derivative of the link function g./. The approximate expected value of 2 is
n
X
E2 D wi .1 wi vi di /Œ1 C .ni 1/
i D1
2 .n p/
O 0 D P
i .ni 1/.1 vi di /
where p is the total number of parameters. The initial estimates of the weights become wO i0 D Œ1 C .ni
1/O 0 1 . After a weighted fit of the model, the b
˛ i and b̌ are recalculated, and so is 2 . Then a revised
estimate of is given by
P
2 i wi .1 wi vi di /
O 1 D
wi .ni 1/.1 wi vi di /
The iterative procedure is repeated until 2 is very close to its degrees of freedom.
Once has been estimated by O under the full model, weights of .1 C .ni 1//O 1 can be used to fit models
that have fewer terms than the full model. See Example 73.10 for an illustration.
N OTE : If the WEIGHT statement is specified with the NORMALIZE option, then the initial wi values are
set to the normalized weights, and the weights resulting from Williams’ method will not add up to the actual
sample size. However, the estimated covariance matrix of the parameter estimates remains invariant to the
scale of the WEIGHT variable.
groups according to the following scheme. Let N be the total number of subjects. Let M be the target number
of subjects for each group given by
M D Œ0:1 N C 0:5
where Œx represents the integral value of x. If the single-trial syntax is used, blocks of subjects are formed
of observations with identical values of the explanatory variables. Blocks of subjects are not divided when
being placed into groups.
Suppose there are n1 subjects in the first block and n2 subjects in the second block. The first block of subjects
is placed in the first group. Subjects in the second block are added to the first group if
Otherwise, they are placed in the second group. In general, suppose subjects of the (j – 1) block have been
placed in the kth group. Let c be the total number of subjects currently in the kth group. Subjects for the jth
block (containing nj subjects) are also placed in the kth group if
Otherwise, the nj subjects are put into the next group. In addition, if the number of subjects in the last group
does not exceed Œ0:05 N (half the target group size), the last two groups are collapsed to form only one
group.
Note that the number of groups, g, can be smaller than 10 if there are fewer than 10 patterns of explanatory
variables. There must be at least three groups in order for the Hosmer-Lemeshow statistic to be computed.
The Hosmer-Lemeshow goodness-of-fit statistic is obtained by calculating the Pearson chi-square statistic
from the 2 g table of observed and expected frequencies, where g is the number of groups. The statistic is
written
g
X .Oi Ni N i /2
2HL D
Ni N i .1 N i /
i D1
where Ni is the total frequency of subjects in the ith group, Oi is the total frequency of event outcomes in the
ith group, and N i is the average estimated predicted probability of an event outcome for the ith group. (Note
that the predicted probabilities are computed as shown in the section “Linear Predictor, Predicted Probability,
and Confidence Limits” on page 5474 and are not the cross validated estimates discussed in the section
“Classification Table” on page 5475.) The Hosmer-Lemeshow statistic is then compared to a chi-square
distribution with .g n/ degrees of freedom, where the value of n can be specified in the LACKFIT option
in the MODEL statement. The default is n = 2. Large values of 2HL (and small p-values) indicate a lack of
fit of the model.
In a sample of n individuals, suppose n1 individuals are observed to have a certain condition or event. Let
this group be denoted by C1 , and let the group of the remaining n2 D n n1 individuals who do not have
the condition be denoted by C2 . Risk factors are identified for the sample, and a logistic regression model is
j of the event of interest is calculated. By
fitted to the data. For the jth individual, an estimated probability b
j are computed as shown in the section “Linear Predictor, Predicted Probability, and Confidence
default, the b
Limits” on page 5474; you can use the cross validated probabilities (discussed in the section “Classification
Table” on page 5475) by specifying the CROSSVALIDATE suboption of the ROCOPTIONS option.
Suppose the n individuals undergo a test for predicting the event and the test is based on the estimated
probability of the event. Higher values of this estimated probability are assumed to be associated with the
event. You can construct an ROC curve by varying the cutpoint that determines which estimated event
probabilities are considered to predict the event. For each cutpoint z, the following measures can be output to
a data set by specifying the OUTROC= option in the MODEL statement or the OUTROC= option in the
SCORE statement:
X
_POS_.z/ D i z/
I.b
i 2C1
X
_NEG_.z/ D I.b
i < z/
i 2C2
X
_FALPOS_.z/ D i z/
I.b
i 2C2
X
_FALNEG_.z/ D I.b
i < z/
i 2C1
_POS_.z/
_SENSIT_.z/ D
n1
_FALPOS_.z/
_1MSPEC_.z/ D
n2
For example, if you have three curves and the second curve is the reference, the contrast used for the overall
test is
0
l1 1 1 0
L1 D D
l20 0 1 1
and you can optionally estimate and test each row of this contrast, in order to test the difference between
the reference curve and each of the other curves. If you do not want to use a reference curve, the global test
optionally uses the following contrast:
0
l1 1 1 0
L2 D D
l20 0 1 1
You can also specify your own contrast matrix. Instead of estimating the rows of these contrasts, you can
request that the difference between every pair of ROC curves be estimated and tested. Demler, Pencina, and
D’Agostino (2012) caution that testing the difference in the AUC between two nested models is not a valid
approach if the added predictor is not significantly associated with the response; in any case, if you use this
approach, you are more likely to fail to reject the null.
By default for the reference contrast, the specified or selected model is used as the reference unless the
NOFIT option is specified in the MODEL statement, in which case the first ROC model is the reference.
In order to label the contrasts, a name is attached to every model. The name for the specified or selected
model is the MODEL statement label, or “Model” if the MODEL label is not present. The ROC statement
models are named with their labels, or as “ROCi” for the ith ROC statement if a label is not specified.
The contrast L1 is labeled as “Reference = ModelName”, where ModelName is the reference model name,
while L2 is labeled “Adjacent Pairwise Differences”. The estimated rows of the contrast matrix are labeled
“ModelName1 – ModelName2”. In particular, for the rows of L1 , ModelName2 is the reference model name.
If you specify your own contrast matrix, then the contrast is labeled “Specified” and the ith contrast row
estimates are labeled “Rowi”.
If ODS Graphics is enabled, then all ROC curves are displayed individually and are also overlaid in a final
display. If a selection method is specified, then the curves produced in each step of the model selection
process are overlaid onto a single plot and are labeled “Stepi”, and the selected model is displayed on a
separate plot and on a plot with curves from specified ROC statements. See Example 73.8 for an example.
ROC Computations
The trapezoidal area under an empirical ROC curve is equal to the Mann-Whitney two-sample rank measure
of association statistic (a generalized U-statistic) applied to two samples, fXi g; i D 1; : : : ; n1 , in C1 and
fYi g; i D 1; : : : ; n2 , in C2 . PROC LOGISTIC uses the predicted probabilities in place of X and Y; however,
in general any criterion could be used. Denote the frequency of observation i in Ck as fki , and denote the
total frequency in Ck as Fk . The WEIGHTED option replaces fki with fki wki , where wki is the weight of
observation i in group Ck . The trapezoidal area under the curve is computed as
n1 X n2
1 X
cO D .Xi ; Yj /f1i f2j
F1 F2
i D1 j D1
8
< 1 Y <X
1
.X; Y / D Y DX
: 2
0 Y >X
Testing Linear Hypotheses about the Regression Coefficients F 5483
O D Pr.Y < X / C 12 Pr.Y D X /. Note that the concordance index, c, in the “Association of
so that E.c/
Predicted Probabilities and Observed Responses” table does not use weights unless both the WEIGHTED
and BINWIDTH=0 options are in effect.
To compare K empirical ROC curves, first compute the trapezoidal areas. Asymptotic normality of the
estimated area follows from U-statistic theory, and a covariance matrix S can be computed; For more
information, see DeLong, DeLong, and Clarke-Pearson (1988). A Wald confidence interval for the rth area,
1 r K, can be constructed as
cOr ˙ z1 ˛
2
sr;r
l 0 cO l 0c
Œl 0 Sl 1=2
has a standard normal distribution. The corresponding confidence interval is
0 1=2
l 0 cO ˙ z1 ˛
2
l Sl
H0 W Lˇ D c
where L is a matrix of coefficients for the linear hypotheses, and c is a vector of constants. The vector of
regression coefficients ˇ includes slope parameters as well as intercept parameters. The Wald chi-square
statistic for testing H0 is computed as
where bV.b̌/ is the estimated covariance matrix. Under H0 , 2W has an asymptotic chi-square distribution
with r degrees of freedom, where r is the rank of L.
Type III estimable functions, see Chapter 47, “The GLM Procedure,” and Chapter 15, “The Four Types of
Estimable Functions.” Also see Littell, Freund, and Spector (1991).
For models that use full-rank parameterization, all parameters are estimable when there are no missing
cells, so it is unnecessary to define estimable functions. The standard test of an effect of interest in this
case is the joint test that the values of the parameters associated with that effect are zero. For a model that
uses effects parameterization (as specified by the PARAM=EFFECT option in the CLASS statement), the
joint test for a main effect is equivalent to testing the equality of marginal means. For a model that uses
reference parameterization (as specified by the PARAM=REF option in the CLASS statement), the joint test
is equivalent to testing the equality of cell means at the reference level of the other model effects. For more
information about the coding scheme and the associated interpretation of results, see Muller and Fetterman
(2002, Chapter 14).
If there is no interaction term, the Type 3 test of an effect for a model that uses GLM parameterization is the
same as the joint test of the effect for the model that uses full-rank parameterization. In this situation, the
joint test is also called the Type 3 test. For a model that contains an interaction term and no missing cells,
the Type 3 test of a component main effect under GLM parameterization is the same as the joint test of the
component main effect under effect parameterization. Both test the equality of cell means. But this Type 3
test differs from the joint test under reference parameterization, which tests the equality of cell means at the
reference level of the other component main effect. If some cells are missing, you can obtain meaningful
tests only by testing a Type III estimation function, so in this case you should use GLM parameterization.
The results of a Type 3 test or a joint test do not depend on the order in which you specify the terms in the
MODEL statement.
Regression Diagnostics
For binary response data, regression diagnostics developed by Pregibon (1981) can be requested by specifying
the INFLUENCE option. For diagnostics available with conditional logistic regression, see the section
“Regression Diagnostic Details” on page 5493. These diagnostics can also be obtained from the OUTPUT
statement.
This section uses the following notation:
rj ; nj rj is the number of event responses out of nj trials for the jth observation. If events/trials syntax
is used, rj is the value of events and nj is the value of trials. For single-trial syntax, nj D 1, and
rj D 1 if the ordered response is 1, and rj D 0 if the ordered response is 2.
wj is the weight of the jth observation.
j is the probability of an event response for the jth observation given by j D F .˛ C ˇ 0 xj /, where
F ./ is the inverse link function defined on page 5458.
b̌ is the maximum likelihood estimate (MLE) of .˛; ˇ1 ; : : : ; ˇs /0 .
V.b̌/
b is the estimated covariance matrix of b̌.
pOj ; qO j pOj is the estimate of j evaluated at b̌, and qO j D 1 pOj .
Pregibon (1981) suggests using the index plots of several diagnostic statistics to identify influential ob-
servations and to quantify the effects on various aspects of the maximum likelihood fit. In an index plot,
the diagnostic statistic is plotted against the observation number. In general, the distributions of these
Regression Diagnostics F 5485
diagnostic statistics are not known, so cutoff values cannot be given for determining when the values are
large. However, the IPLOTS and INFLUENCE options in the MODEL statement and the PLOTS option in
the PROC LOGISTIC statement provide displays of the diagnostic values, allowing visual inspection and
comparison of the values across observations. In these plots, if the model is correctly specified and fits all
observations well, then no extreme points should appear.
The next five sections give formulas for these diagnostic statistics.
where
wj nj
w
ej D
pOj qO j Œg 0 .pOj /2
wj .rj nj pOj /ŒpOj qO j g 00 .pOj / C .qO j pOj /g 0 .pOj /
w
bj D w
ej C
.pOj qO j /2 Œg 0 .pOj /3
and g 0 ./ and g 00 ./ are the first and second derivatives of the link function g./, respectively.
For a binary response logit model, the hat matrix diagonal elements are
0 b b̌ 1
hj D wj nj pOj qO j .1; xj /V. /
xj
If the estimated probability is extreme (less than 0.1 or greater than 0.9, approximately), then the hat diagonal
might be greatly reduced in value. Consequently, when an observation has a very large or very small estimated
probability, its hat diagonal value is not a good indicator of the observation’s distance from the design space
(Hosmer and Lemeshow 2000, p. 171).
Residuals
Residuals are useful in identifying observations that are not explained well by the model. Pearson residuals
are components of the Pearson chi-square statistic and deviance residuals are components of the deviance.
The Pearson residual for the jth observation is
p
wj .rj nj pOj /
j D p
nj pOj qO j
The Pearson chi-square statistic is the sum of squares of the Pearson residuals.
The deviance residual for the jth observation is
p 8
q 2wj nj log.qO j /
ˆ
< if rj D 0
r n r
dj D ˙ 2wj Œrj log. n jpO / C .nj rj / log. nj qO j / if 0 < rj < nj
ˆ p j j j j
2wj nj log.pOj / if rj D nj
:
5486 F Chapter 73: The LOGISTIC Procedure
where the plus (minus) in ˙ is used if rj =nj is greater (less) than pOj . The deviance is the sum of squares of
the deviance residuals.
The STDRES option in the INFLUENCE and PLOTS=INFLUENCE options computes three more residuals
(Collett 2003). The Pearson and deviance residuals are standardized to have approximately unit variance:
j
e pj D p
1 hj
dj
edj D p
1 hj
The likelihood residuals, which estimate components of a likelihood ratio test of deleting an individual
observation, are a weighted combination of the standardized Pearson and deviance residuals
q
elj D sign.rj nj pOj / hj ep2 j C .1 hj /ed2
j
DFBETAS
For each parameter estimate, the procedure calculates a DFBETAS diagnostic for each observation. The
DFBETAS diagnostic for an observation is the standardized difference in the parameter estimate due to
deleting the observation, and it can be used to assess the effect of an individual observation on each estimated
parameter of the fitted model. Instead of reestimating the parameter every time an observation is deleted,
PROC LOGISTIC uses the one-step estimate. See the section “Predicted Probability of an Event for
Classification” on page 5476. For the jth observation, the DFBETAS are given by
DFBETASij D i b̌1 =O i j
where i D 0; 1; : : : ; s; O i is the standard error of the ith component of b̌, and i b̌1j is the ith component of
the one-step difference
1 wj .rj nj pOj / b b̌ 1
j D
b̌ V. /
1 hj xj
b̌1j is the approximate change (b̌ b̌1j ) in the vector of parameter estimates due to the omission of the jth
observation. The DFBETAS are useful in detecting observations that are causing instability in the selected
coefficients.
C and CBAR
C and CBAR are confidence interval displacement diagnostics that provide scalar measures of the influence
of individual observations on b̌. These diagnostics are based on the same idea as the Cook distance in
linear regression theory (Cook and Weisberg 1982), but use the one-step estimate. C and CBAR for the jth
observation are computed as
Cj D 2j hj =.1 hj /2
and
C j D 2j hj =.1 hj /
respectively.
Typically, to use these statistics, you plot them against an index and look for outliers.
Scoring Data Sets F 5487
DIFDEV D dj2 C C j
and
DIFCHISQ D C j = hj
Statistic Description
P
Total frequency F D Pi fi ni
Total weight W D iP fi wi ni
Log likelihood log L D i fi wi log.b i /
Full log likelihood P Lf D constant C log L
log
i 1fF_Yi ¤ I_Yi gfi ni
Misclassification (error) rate
F
AIC 2 log Lf C 2p
2pn
AICC 2 log Lf C
n p 1
BIC 2 log Lf C p log.n/
SC 2 log Lf C p log.F /
2=F
2 L0
R-square R D1
L
R2
Maximum-rescaled R-square 2=F
1 L0
AUC Area under the ROC curve
1 P
ij /2
P
Brier score (polytomous response) W Pi fi wi j .yij b
1
Brier score (binary response) W Pi fi wi .ri .1 i /2 C .ni
b 2i /
ri /b
1
Brier reliability (events/trials syntax) W i fi wi .ri =ni i /2
b
Suppose there are k C 1 response categories. Let Y be the response variable with levels 1; : : : ; k C 1. Let
x D .x0 ; x1 ; : : : ; xs /0 be a .s C 1/-vector of covariates, with x0 1. Let ˇ be the vector of intercept and
slope regression parameters.
Posterior probabilities are given by
po .Y D i jx/ ppno .Y
.Y Di /
Di /
p.Y D i jx/ D P i D 1; : : : ; k C 1
j po .Y D j jx/ ppno .Y
.Y Dj /
Dj /
where the old posterior probabilities (po .Y D i jx/; i D 1; : : : ; k C 1) are the conditional probabilities of
the response categories given x, the old priors (po .Y D i /; i D 1; : : : ; k C 1) are the sample proportions of
response categories of the training data, and the new priors (pn .Y D i /; i D 1; : : : ; k C 1) are specified in
the PRIOR= or PRIOREVENT= option. To simplify notation, absorb the old priors into the new priors; that
is
pn .Y D i /
p.Y D i / D i D 1; : : : ; k C 1
po .Y D i /
Note if the PRIOR= and PRIOREVENT= options are not specified, then p.Y D i / D 1.
The posterior probabilities are functions of ˇ and their estimates are obtained by substituting ˇ by its MLE
b̌. The variances of the estimated posterior probabilities are given by the delta method as follows:
@p.Y D i jx/ 0
@p.Y D ijx/
p .Y D i jx// D
Var.b Var.b̌/
@ˇ @ˇ
where
@po .Y Di jx/
po .Y D i jx/p.Y D i / j @po .Y@ˇDj jx/ p.Y D j /
P
@p.Y D i jx/ @ˇ
p.Y D i /
DP
Œ j po .Y D j jx/p.Y D j /2
P
@ˇ j po .Y D j jx/p.Y D j /
and the old posterior probabilities po .Y D i jx/ are described in the following sections.
A 100(1 ˛)% confidence interval for p.Y D i jx/ is
q
p .Y D i jx/ ˙ z1 ˛=2 Var.b
b c p .Y D i jx//
i D i .ˇ/ D ˛i C x0 ˇs ; i D 1; : : : ; k
For i D 1; : : : ; k, let ıi .x/ be a (k + 1) column vector with ith entry equal to 1, k + 1 entry equal to x, and all
other entries 0. The derivative of po .Y D i jx/ with respect to ˇ are
8 0
@po .Y D i jx/ < F .˛1 C x0 ˇs /ı1 .x/ i D1
D F 0 .˛i C x0 ˇs /ıi .x/ F 0 .˛i 0
1 C x ˇs /ıi 1 .x/ i D 2; : : : ; k
@ˇ
F 0 .˛k C x0 ˇs /ık .x/ i DkC1
:
In the delta-method equation for the variance, replace p.Y D jx/ with p.Y jx/.
Finally, for the cumulative response model, use
co .Y i jx/ D F .O i /
p i D 1; : : : ; k
co .Y k C 1jx/ D 1
p
@po .Y i jx/
D F 0 .˛i C x0 ˇs /ıi .x/ i D 1; : : : ; k
@ˇ
@po .Y k C 1jx/
D 0
@ˇ
Scoring Data Sets F 5491
i D i .ˇ/ D x0 ˇi i D 1; : : : ; k
1
co .Y D k C 1jx/ Pr.Y D k C 1jx/ D
p Pk
1 C lD1 exp.O l /
co .Y D i jx/ Pr.Y D i jx/ D p
p co .Y D k C 1jx/ exp.i /; i D 1; : : : ; k
where
@po .Y D ijx/ po .Y D i jx/.1 po .Y D i jx// j D i
D
@j po .Y D i jx/po .Y D j jx/ otherwise
D .ˇ/ D x0 ˇ
O D x0 Var.b̌/x
Var./
Estimates of po .Y D 1jx/ and confidence intervals for the po .Y D 1jx/ are obtained by back-transforming
O and the confidence intervals for , respectively. That is,
co .Y D 1jx/ D F ./
p O
Computational Details
For each stratum h, h D 1; : : : ; H , number the observations as i D 1; : : : ; nh so that hi indexes the ith
observation in stratum h. Denote the p covariates for the hith observation as xhi and its binary response
as yhi , and let y D .y11 ; : : : ; y1n1 ; : : : ; yH1 ; : : : ; yH nH /0 , Xh D .xh1 : : : xhnh /0 , and X D .X01 : : : X0H /0 .
Let the dummy variables zh ; h D 1; : : : ; H , be indicator functions for the strata (zh D 1 if the observation
is in stratum h), and denote zhi D .z1 ; : : : ; zH / for the hith observation, Zh D .zh1 : : : zhnh /0 , and Z D
.Z01 : : : Z0H /0 . Denote X D .ZjX) and xhi D .z0hi jx0hi /0 . Arrange the observations in each stratum h so
that yhi D 1 for i D 1; : : : ; mh , and yhi D 0 for i D mhC1 ; : : : ; nh . Suppose all observations have unit
frequency.
Consider the binary logistic regression model on page 5379 written as
logit./ D X
where the parameter vector D .˛0 ; ˇ 0 /0 consists of ˛ D .˛1 ; : : : ; ˛H /0 , ˛h is the intercept for stratum
h; h D 1; : : : ; H , and ˇ is the parameter vector for the p covariates.
From the section “Determining Observations for Likelihood Contributions” on page 5459, you can write the
likelihood contribution of observation hi; i D 1; : : : ; nh ; h D 1; : : : ; H; as
0
e yhi xhi
Lhi ./ D 0
x
hi
1Ce
where yhi D 1 when the response takes Ordered Value 1, and yhi D 0 otherwise.
The full likelihood is
nh
H Y
ey X
0
Y
L./ D Lhi ./ D Q 0
H Qnh x
hi
hD1 i D1 hD1 i D1 1 C e
When your nuisance parameters are the stratum-specific intercepts .˛1 ; : : : ; ˛H /0 , and the slopes ˇ are
your parameters of interest, “conditioning out” the nuisance parameters produces the conditional likelihood
(Lachin 2000)
H H Qmh 0
i D1 exp.xhi ˇ/
Y Y
L.ˇ/ D Lh .ˇ/ D P Qjmh 0
hD1 hD1 j Dj1 exp.xhj ˇ/
nh
where the summation is over all m h
subsets fj1 ; : : : ; jmh g of mh observations chosen from the nh observa-
tions in stratum h. Note that the nuisance parameters have been factored out of this equation.
For conditional asymptotic inference, maximum likelihood estimates b̌ of the regression parameters are
obtained by maximizing the conditional likelihood, and asymptotic results are applied to the conditional
likelihood function and the maximum likelihood estimators. A relatively fast method of computing this
conditional likelihood and its derivatives is given by Gail, Lubin, and Rubinstein (1981) and Howard (1972).
The optimization techniques can be controlled by specifying the NLOPTIONS statement.
Sometimes the log likelihood converges but the estimates diverge. This condition is flagged by having
inordinately large standard errors for some of your parameter estimates, and can be monitored by specifying
the ITPRINT option. Unfortunately, broad existence criteria such as those discussed in the section “Existence
of Maximum Likelihood Estimates” on page 5463 do not exist for this model. It might be possible to
circumvent such a problem by standardizing your independent variables before fitting the model.
and the h subscript on matrices indicates the submatrix for the stratum,
h D .
h1 ; : : : ;
hnh /0 , and
hi D
x0hi ˇ. Then the gradient and information matrix are
H
@lh
g.ˇ/ D D X0 .y /
@ˇ hD1
H
@2 lh
ƒ.ˇ/ D D X0 diag.U1 ; : : : ; UH /X
@ˇ 2 hD1
5494 F Chapter 73: The LOGISTIC Procedure
where
P Qjmh
@a.
h / j.i / j Dj1 exp.
hj /
hi D D P Qjmh
@
hi exp.
hj /
j Dj1
h D .h1 ; : : : ; hnh /
@2 a.
h /
2
@ a.
h /
Uh D D D faij g
@
h2 @
hi @
hj
P Qkmh
k.i;j / kDk1
exp.
hk / @a.
h / @a.
h /
aij D P Qkmh D hij hi hj
exp.
/ @
hi @
hj
kDk1 hk
and where hi is the conditional stratum-specific probability that subject i in stratum h is a case, the
summation on j.i / is over all subsets from f1; : : : ; nh g of size mh that contain the index i, and the summation
on k.i; j / is over all subsets from f1; : : : ; nh g of size mh that contain the indices i and j.
1
To produce the true one-step estimate ˇhi , start at the MLE b̌, delete the hith observation, and use this
reduced data set to compute the next Newton-Raphson step. Note that if there is only one event or one
nonevent in a stratum, deletion of that single observation is equivalent to deletion of the entire stratum. The
augmentation method does not take this into account.
The augmented model is
X0
0 0
g.ˇ / D .y / D
z0hi yhi hi
X0 ƒ.ˇ/ X0 Uzhi
ƒ.ˇ 0 / D U ŒX zhi D
z0hi z0hi UX z0hi Uzhi
Inserting the ˇ 0 and .X0 ; z0hi /0 into the Gail, Lubin, and Rubinstein (1981) algorithm provides the appropriate
estimates of g.ˇ 0 / and ƒ.ˇ 0 /. Indicate these estimates with b D .b̌/, bU D U.b̌/, bg, and ƒ.
b
hi ˇ D ˇ 0 ˇhi1
D b 1 .ˇ 0 /b
ƒ g.ˇ 0 /
D b 1 .b̌/.X0 b
ƒ Uzhi /M 1 0
zhi .y /
b
where
M D .z0hi b
Uzhi / .z0hi b
UX/ƒ
b 1
.b̌/.X0 b
Uzhi /
For each observation in the data set, a DFBETA statistic is computed for each parameter ˇj , 1 j p, and
standardized by the standard error of ˇj from the full data set to produce the estimate of DFBETAS.
Exact Conditional Logistic Regression F 5495
tracef.z0hi b
UX/ƒ
b 1 .b̌/.X0 b
Uzhi /g
hhi D
tracefz0hi b
Uzhi g
This definition of leverage produces different values from those defined by Pregibon (1984); Moolgavkar,
Lustbader, and Venzon (1985); Hosmer and Lemeshow (2000); however, it has the advantage that no extra
computations beyond those for the DFBETAS are required.
The estimated residuals ehi D yhi b g.ˇ 0 /, and the weights, or predicted probabilities,
hi are obtained from b
hi D yhi ehi . The residuals are standardized and reported as (estimated) Pearson residuals:
are then b
rhi nhi b
hi
p
nhi b
hi .1 hi /
b
where rhi is the number of events in the observation and nhi is the number of trials.
The STDRES option in the INFLUENCE option computes the standardized Pearson residual:
ehi
es;hi D p
1 hhi
For events/trials MODEL statement syntax, treat each observation as two observations (the first for the
nonevents and the second for the events) with frequencies fh;2i 1 D nhi rhi and fh;2i D rhi , and augment
the model with a matrix Zhi D Œzh;2i 1 zh;2i instead of a single zhi vector. Writing
hi D x0hi ˇfhi in the
preceding section results in the following gradient and information matrix:
2 3
0
g.ˇ 0 / D 4 fh;2i 1 .yh;2i 1 h;2i 1 / 5
fh;2i .yh;2i h;2i /
X0 diag.f /Udiag.f /Zhi
ƒ.ˇ/
ƒ.ˇ 0 / D
Z0hi diag.f /Udiag.f /X Z0hi diag.f /Udiag.f /Zhi
hi D yh;2i eh;2i =rh;2i , while the leverage and the DFBETAS are
The predicted probabilities are then b
produced from ƒ.ˇ 0 / in a fashion similar to that for the preceding single-trial equations.
C.t/ exp.t0 ˇ/
Pr.T D t/ D Qn 0
i D1 Œ1 C exp.xi ˇ/
0
In order to condition out the nuisance parameters, partition the parameter vector ˇ D .ˇN ; ˇI0 /0 , where ˇN is
a pN 1 vector of the nuisance parameters, and ˇI is the parameter vector for the remaining pI D p pN
parameters of interest. Likewise, partition X into XN and XI , T into TN and TI , and t into tN and tI . The
nuisance parameters can be removed from the analysis by conditioning on their sufficient statistics to create
the conditional likelihood of TI given TN D tN ,
Pr.T D t/
Pr.TI D tI jTN D tN / D
Pr.TN D tN /
C.tN ; tI / exp.t0I ˇI /
D fˇI .tI jtN / D P 0
u C.tN ; u/ exp.u ˇI /
where C.tN ; u/ is the number of vectors y such that y0 XN D tN and y0 XI D u. Note that the nuisance
parameters have factored out of this equation, and that C.tN ; tI / is a constant.
The goal of the exact conditional analysis is to determine how likely the observed response y0 is with respect
to all 2n possible responses y D .y1 ; : : : ; yn /0 . One way to proceed is to generate every y vector for which
y0 XN D tN , and count the number of vectors y for which y0 XI is equal to each unique tI . Generating the
conditional distribution from complete enumeration of the joint distribution is conceptually simple; however,
this method becomes computationally infeasible very quickly. For example, if you had only 30 observations,
you would have to scan through 230 different y vectors.
Several algorithms are available in PROC LOGISTIC to generate the exact distribution. All of the algorithms
are based on the following observation. Given any y D .y1 ; : : : ; yn /0 and a design X D .x1 ; : : : ; xn /0 , let
y.i / D .y1 ; : : : ; yi /0 and X.i / D .x1 ; : : : ; xi /0 be the first i rows of each matrix. Write the sufficient statistic
based on these i rows as t0.i / D y.i 0
X . A recursion relation results: t.i C1/ D t.i / C yi C1 xi C1 .
/ .i /
The multivariate shift algorithm developed by Hirji, Mehta, and Patel (1987), which steps through the
recursion relation by adding one observation at a time and building an intermediate distribution at each
step. If it determines that t.i / for the nuisance parameters could eventually equal t, then t.i / is added
to the intermediate distribution.
An extension of the multivariate shift algorithm to generalized logit models by Hirji (1992). Because
the generalized logit model fits a new set of parameters to each logit, the number of parameters in the
model can easily get too large for this algorithm to handle. Note for these models that the hypothesis
tests for each effect are computed across the logit functions, while individual parameters are estimated
for each logit function.
A network algorithm described in Mehta, Patel, and Senchaudhuri (1992), which builds a network for
each parameter that you are conditioning out in order to identify feasible yi for the y vector. These
Exact Conditional Logistic Regression F 5497
networks are combined and the set of feasible yi is further reduced, and then the multivariate shift
algorithm uses this knowledge to build the exact distribution without adding as many intermediate
t.iC1/ as the multivariate shift algorithm does.
A hybrid Monte Carlo and network algorithm described by Mehta, Patel, and Senchaudhuri (2000),
which extends their 1992 algorithm by sampling from the combined network to build the exact
distribution.
A Markov chain Monte Carlo algorithm described by Forster, McDonald, and Smith (2003), which
generates the exact distribution by repeatedly perturbing the response vector to obtain a new response
vector while maintaining the sufficient statistics for the nuisance parameters, and the resulting t are
added to the exact distribution.
The bulk of the computation time and memory for these algorithms is consumed by the creation of the
networks and the exact joint distribution. After the joint distribution for a set of effects is created, the
computational effort required to produce hypothesis tests and parameter estimates for any subset of the effects
is (relatively) trivial. See the section “Computational Resources for Exact Logistic Regression” on page 5506
for more computational notes about exact analyses.
N OTE : An alternative to using these exact conditional methods is to perform Firth’s bias-reducing penalized
likelihood method (see the FIRTH option in the MODEL statement); this method has the advantage of being
much faster and less memory intensive than exact algorithms, but it might not converge to a solution.
Hypothesis Tests
Consider testing the null hypothesis H0 W ˇI D 0 against the alternative HA W ˇI ¤ 0, conditional on TN D tN .
Under the null hypothesis, the test statistic for the exact probability test is just fˇI D0 .tI jtN /, while the
corresponding p-value is the probability of getting a less likely (more extreme) statistic,
X
p.tI jtN / D f0 .ujtN /
u2p
where p D fuW there exist y with y0 XI D u, y0 XN D tN , and f0 .ujtN / f0 .tI jtN /g.
The exact probability test is not necessarily a sum of tail areas and can be inflated if the distribution is skewed.
The more robust exact conditional scores test is a sum of tail areas and is usually preferred to the exact
probability test. To compute the exact conditional scores test, the conditional mean I and variance matrix
†I of the TI (conditional on TN D tN ) are calculated, and the score statistic for the observed value,
s D .tI I /0 †I 1 .tI I /
where, for i 2 fp; sg, < is i using strict inequalities, and D is i using equalities with the added
restriction that u ¤ tI . Letting .ı1 ; ı2 / D .0:5; 1:0/ yields Lancaster’s mid-p.
C AUTION : When the exact distribution has ties and you specify METHOD=NETWORKMC or
METHOD=MCMC, the sampling algorithm estimates p.tjtN / with error, and hence it cannot determine
precisely which values contribute to the reported p-values. For example, if the exact distribution has densities
f0:2; 0:2; 0:2; 0:4g and if the observed statistic has probability 0.2, then the exact probability p-value is
exactly 0.6. Under Monte Carlo sampling, if the densities after N samples are f0:18; 0:21; 0:23; 0:38g and the
observed probability is 0.21, then the resulting p-value is 0.39. Therefore, the exact probability test p-value
for this example fluctuates between 0.2, 0.4, and 0.6, and the reported p-values are actually lower bounds
for the true p-values. If you need more precise values, you can specify the OUTDIST= option, determine
appropriate cutoff values for the observed probability and score, and then construct p-value estimates from
the OUTDIST= data set and display them in the SAS log by using the following statements:
data _null_;
set outdist end=end;
retain pvalueProb 0 pvalueScore 0;
if prob < ProbCutOff then pvalueProb+prob;
if score > ScoreCutOff then pvalueScore+prob;
if end then put pvalueProb= pvalueScore=;
run;
Because the METHOD=MCMC samples are correlated, the covariance that is computed for the exact
conditional scores test is biased. Specifying the NTHIN= option might reduce this bias.
In this case, a median unbiased estimate (Hirji, Tsiatis, and Mehta 1989) b ˇ i is produced that satisfies
fb̌ .ti jtN / D 0:5, and a Newton-Raphson algorithm is used to perform the search.
i
The standard error of the exact conditional likelihood estimate is just the negative of the inverse of the second
derivative of the exact conditional log likelihood (Agresti 2002).
Likelihood ratio tests based on the conditional PDF are used to test the null H0 W ˇi D 0 against the alternative
HA W ˇi > 0. The critical region for this UMP test consists of the upper tail of values for Ti in the exact
distribution. Thus, the p-value pC .ti I 0/ for a one-tailed test is
X
pC .ti I 0/ D f0 .ujtN /
uti
Similarly, the p-value p .ti I 0/ for the one-tailed test of H0 against HA W ˇi < 0 is
X
p .ti I 0/ D f0 .ujtN /
uti
By default, the p-value that is reported for a single parameter in the “Exact Parameter Estimates” table is for
the two-tailed test. For median unbiased estimates, the p-value for a one-tailed test is always reported.
An upper 100.1 2/% exact confidence limit for b ˇ i corresponding to the observed ti is the solution ˇU .ti /
of D p .ti ; ˇU .ti //, while the lower exact confidence limit is the solution ˇL .ti / of D pC .ti ; ˇL .ti //.
Again, a Newton-Raphson procedure is used to search for the solutions. Note that one of the confidence
limits for a median unbiased estimate is set to infinity and the other is computed at 2, which results in the
display of a one-sided 100.1 2/% confidence interval.
Specifying the ONESIDED option displays only one p-value and one confidence interval, because small
values of pC .ti I 0/ and p .ti I 0/ support different alternative hypotheses and only one of these p-values can
be less than 0.50.
The mid-p confidence limits are the solutions to minfp .ti ; ˇ.ti //; pC .ti ; ˇ.ti //g .1 ı1 /fˇ .ti / .ti jtN / D
for D ˛=2; 1 ˛=2 (Vollset, Hirji, and Afifi 1991). ı1 D 1 produces the usual exact (or max-p) confidence
interval, ı1 D 0:5 yields the mid-p interval, and ı1 D 0 gives the min-p interval. The mean of the endpoints
of the max-p and min-p intervals provides the mean-p interval as defined by Hirji, Mehta, and Patel (1988).
Estimates and confidence intervals for the odds ratios are produced by exponentiating the estimates and
interval endpoints for the parameters.
_LINK_, a character variable of length 8 with four possible values: CLOGLOG for the complementary
log-log function, LOGIT for the logit function, NORMIT for the probit (alias normit) function, and
GLOGIT for the generalized logit function
_TYPE_, a character variable of length 8 with two possible values: PARMS for parameter estimates
or COV for covariance estimates. If an EXACT statement is also specified, then two other values are
possible: EPARMMLE for the exact maximum likelihood estimates and EPARMMUE for the exact
median unbiased estimates.
_NAME_, a character variable containing the name of the response variable when _TYPE_=PARMS,
EPARMMLE, and EPARMMUE, or the name of a model parameter when _TYPE_=COV
_STATUS_, a character variable that indicates whether the estimates have converged
one variable for each slope parameter and one variable for the offset variable if the OFFSET= option if
specified. If an effect is not included in the final model in a model building process, the corresponding
parameter estimates and covariances are set to missing values.
proc logistic;
class Gender;
model Net3 = Age Gender / link=glogit;
run;
Input and Output Data Sets F 5501
There are two logit functions, one that contrasts ABC with NBC and one that contrasts CBS with NBC.
For each logit, there are three parameters: an intercept parameter, a slope parameter for Age, and a slope
parameter for Gender (because there are only two gender levels and the EFFECT parameterization is used by
default). The names of the parameters and their descriptions are as follows:
Intercept_ABC intercept parameter for the logit that contrasts ABC with NBC
Intercept_CBS intercept parameter for the logit that contrasts CBS with NBC
Age_ABC Age slope parameter for the logit that contrasts ABC with NBC
Age_CBS Age slope parameter for the logit that contrasts CBS with NBC
GenderFemale_ABC Gender=Female slope parameter for the logit that contrasts ABC with NBC
GenderFemale_CBS Gender=Female slope parameter for the logit that contrasts CBS with NBC
In a polytomous response model, if you specify an effect in both the EQUALSLOPES and UNEQUAL-
SLOPES options, then its unequal slope parameter names are prefixed with “U_” if they are unconstrained or
“C_” if they are constrained.
If there are more than two response categories and you specify only the PREDPROBS= option, then each
input observation produces one observation in the OUT= data set. However, if you fit an ordinal (cumulative)
model and specify options other than the PREDPROBS= options, each input observation generates as many
output observations as one fewer than the number of response levels, and the predicted probabilities and their
confidence limits correspond to the cumulative predicted probabilities. If you fit a generalized logit model
and specify options other than the PREDPROBS= options, each input observation generates as many output
observations as the number of response categories; the predicted probabilities and their confidence limits
correspond to the probabilities of individual response categories.
For observations in which only the response variable is missing, values of the XBETA=, STDXBETA=,
PREDICTED=, UPPER=, LOWER=, and the PREDPROBS= options are computed even though these
observations do not affect the model fit. This enables, for instance, predicted probabilities to be computed for
new observations.
F_Y, the normalized levels of the response variable Y in the data set being scored. If the events/trials
syntax is used, the F_Y variable is not created.
I_Y, the normalized levels that the observations are classified into. Note that an observation is classified
into the level with the largest probability. If the events/trials syntax is used, the _INTO_ variable is
created instead, and it contains the values EVENT and NONEVENT.
P_Yi , the posterior probabilities of the normalized response level Yi
If the CLM option is specified in the SCORE statement, the OUT= data set also includes the following:
– LCL_Yi , the lower 100(1 ˛)% confidence limits for P_Yi
– UCL_Yi , the upper 100(1 ˛)% confidence limits for P_Yi
Input and Output Data Sets F 5503
data test;
input y x1 x2 count;
datalines;
0 0 0 1
1 0 0 1
0 1 1 2
1 1 1 1
1 0 2 3
1 1 2 1
1 2 0 3
1 2 1 2
1 2 2 1
;
The first nine observations in the dist data set contain an exact distribution for the parameters of the x2
effect (hence the values for the x1 parameter are missing), and the remaining five observations are for the
x1 parameter. If a joint distribution was created, there would be observations with values for both the x1
and x2 parameters. For CLASS variables, the corresponding parameters in the dist data set are identified by
concatenating the variable name with the appropriate classification level.
The data set contains the possible sufficient statistics of the parameters for the effects specified in the EXACT
statement, and the Count variable contains the number of different responses that yield these statistics. In
particular, there are six possible response vectors y for which the dot product y0 x1 was equal to 2, and for
which y0 x20, y0 x21, and y0 1 were equal to their actual observed values (displayed in the “Sufficient Statistics”
table).
When hypothesis tests are performed on the parameters, the Prob variable contains the probability of obtaining
that statistic (which is just the count divided by the total count), and the Score variable contains the score for
that statistic.
The OUTDIST= data set can contain a different exact conditional distribution for each specified EXACT
statement. For example, consider the following EXACT statements:
_STEP_, the model step number. This variable is not included if model selection is not requested.
_PROB_, the estimated probability of an event. These estimated probabilities serve as cutpoints for
predicting the response. Any observation with an estimated event probability that exceeds or equals
_PROB_ is predicted to be an event; otherwise, it is predicted to be a nonevent. Predicted probabilities
that are close to each other are grouped together, with the maximum allowable difference between the
largest and smallest values less than a constant that is specified by the ROCEPS= option. The smallest
estimated probability is used to represent the group.
_SENSIT_, the sensitivity, which is the proportion of event observations that were predicted to have an
event response
_1MSPEC_, one minus specificity, which is the proportion of nonevent observations that were predicted
to have an event response
Note that none of these statistics are affected by the bias-correction method discussed in the section “Classifi-
cation Table” on page 5475. An ROC curve is obtained by plotting _SENSIT_ against _1MSPEC_.
For more information, see the section “Receiver Operating Characteristic Curves” on page 5480.
Computational Resources
The memory needed to fit an unconditional model is approximately 8n.p C 2/ C 24.p C 2/2 bytes, where p
is the number of parameters estimated and n is the number of observations in the data set. For cumulative
response models with more than two response levels, a test of the parallel lines assumption requires an
additional memory of approximately 4k 2 .m C 1/2 C 24.m C 2/2 bytes, where k is the number of response
levels and m is the number of slope parameters. However, if this additional memory is not available, the
procedure skips the test and finishes the other computations. You might need more memory if you use the
SELECTION= option for model building.
The data that consist of relevant variables (including the design variables for model effects) and observations
for fitting the model are stored in a temporary utility file. If sufficient memory is available, such data will also
be kept in memory; otherwise, the data are reread from the utility file for each evaluation of the likelihood
function and its derivatives, with the resulting execution time of the procedure substantially increased.
5506 F Chapter 73: The LOGISTIC Procedure
Specifying the MULTIPASS option in the MODEL statement avoids creating this utility file and also does
not store the data in memory; instead, the DATA= data set is reread when needed. This saves approximately
8n.p C 2/ bytes of memory but increases the execution time.
If a conditional logistic regression is performed, then approximately 4.m2 C m C 4/ maxh .mh / C .8sH C
36/H C 12sH additional bytes of memory are needed, where mh is the number of events in stratum
h, H is the total number of strata, and sH is the number of variables used to define the strata. If the
CHECKDEPENDENCY=ALL option is specified in the STRATA statement, then an extra 4.m C H /.m C
H C 1/ bytes are required, and the resulting execution time of the procedure might be substantially increased.
data one;
do obs=1 to HalfNobs;
do Y=0 to 1;
X1=round(ranuni(0));
...
XN =round(ranuni(0));
output;
end;
end;
options fullstimer;
proc logistic exactonly;
exactoptions method=network maxtime=1200;
class X: / param=ref;
model Y=X:;
exact X1 / outdist=dist;
run;
Computational Resources F 5507
At any time while PROC LOGISTIC is deriving the distributions, you can terminate the computations by
pressing the system interrupt key sequence (see the SAS Companion for your system) and choosing to stop
computations. If you run out of memory, see the SAS Companion for your system to see how to allocate
more.
You can use the EXACTOPTIONS option MAXTIME= to limit the total amount of time PROC LOGISTIC
uses to derive all of the exact distributions. If PROC LOGISTIC does not finish within that time, the procedure
terminates.
5508 F Chapter 73: The LOGISTIC Procedure
Frequency calculation is performed in the log scale by default. This reduces the need to check for excessively
large frequencies but can be slower than not scaling. You can turn off the log scaling by specifying the
NOLOGSCALE option in the EXACTOPTIONS statement. If a frequency in the exact distribution is larger
than the largest integer that can be held in double precision, a warning is printed to the SAS log. Because
inaccuracies due to adding small numbers to these large frequencies might have little or no effect on the
statistics, the exact computations continue.
You can monitor the progress of the procedure by submitting your program with the EXACTOPTIONS option
STATUSTIME=. If the procedure is too slow, you can try another method by specifying the EXACTOPTIONS
option METHOD=, you can try reordering the variables in the MODEL statement (note that CLASS variables
are always processed before continuous covariates), or you can try reparameterizing your classification
variables as in the following statement:
If you condition out CLASS variables that use reference or GLM coding but you are not using the STRATA
statement, then you can speed up the analysis by specifying one of the nuisance CLASS variables in the
STRATA statement. This performance gain occurs because STRATA variables partition your data set into
smaller pieces. However, moving two (or more) nuisance CLASS variables into the STRATA statement
results in a different model, because the sufficient statistics for the second CLASS variable are actually
computed across the levels of the first CLASS variable.
Displayed Output
If you use the NOPRINT option in the PROC LOGISTIC statement, the procedure does not display any
output. Otherwise, the tables displayed by the LOGISTIC procedure are discussed in the following section
in the order in which they appear in the output. Some of the tables appear only in conjunction with certain
options or statements; for more information, see the section “ODS Table Names” on page 5513.
N OTE : The EFFECT, ESTIMATE, LSMEANS, LSMESTIMATE, and SLICE statements also create tables,
which are not listed in this section. For information about these tables, see the corresponding sections of
Chapter 19, “Shared Concepts and Topics.”
Table Summary
Model Information and the Number of Observations
See the section “Missing Values” on page 5456 for information about missing-value handling, and the
sections “FREQ Statement” on page 5420 and “WEIGHT Statement” on page 5456 for information about
valid frequencies and weights.
Response Profile
Displays the Ordered Value assigned to each response level. For more information, see the section “Response
Level Ordering” on page 5457.
Strata Summary
Shows the pattern of the number of events and the number of nonevents in a stratum. See the section
“STRATA Statement” on page 5453 for more information.
Strata Information
Displays if you specify the INFO option in a STRATA statement.
the estimated standard error of the parameter estimate, computed as the square root of the corresponding
diagonal element of the estimated covariance matrix
the Wald chi-square statistic, computed by squaring the ratio of the parameter estimate divided by its
standard error estimate. For more information, see the section “Testing Linear Hypotheses about the
Regression Coefficients” on page 5483.
the p-value tests the null hypothesis H0 W ˇi D 0; small values reject the null.
the standardized estimate for the slope parameter, if you specify the STB option in the MODEL
statement. For more information, see the STB option on page 5438.
exponentiated values of the estimates of the slope parameters, if you specify the EXPB option in the
MODEL statement. For more information, see the EXPB option on page 5429.
the label of the variable, if you specify the PARMLABEL option in the MODEL statement and if
space permits. Due to constraints on the line size, the variable label might be suppressed in order to
display the table in one panel. Use the SAS system option LINESIZE= to specify a larger line size to
accommodate variable labels. A shorter line size can break the table into two panels allowing labels to
be displayed.
Classification Table
Displays if you use the CTABLE option in the MODEL statement. If you specify a list of cutpoints with the
PPROB= option, then the cutpoints are displayed in the Prob Level column. If you specify the prior event
probabilities with the PEVENT= option, then the probabilities are displayed in the Prob Event column. The
Correct column displays the number of correctly classified events and nonevents, the Incorrect Event column
displays the number of nonevents incorrectly classified as events, and the Incorrect Nonevent column gives
the number of events incorrectly classified as nonevents. For more information, see the section “Classification
Table” on page 5475.
Regression Diagnostics
Displays if you specify the INFLUENCE option in the MODEL statement. See the section “Regression
Diagnostics” on page 5484 for more information about diagnostics from an unconditional analysis, and the
section “Regression Diagnostic Details” on page 5493 for information about diagnostics from a conditional
analysis.
ODS Table Names F 5513
Sufficient Statistics
Displays if you request an OUTDIST= data set in an EXACT statement. The table lists the parameters
and their observed sufficient statistics.
ODS Graphics
Statistical procedures use ODS Graphics to create graphs as part of their output. ODS Graphics is described
in detail in Chapter 21, “Statistical Graphics Using ODS.”
Before you create graphs, ODS Graphics must be enabled (for example, by specifying the ODS GRAPH-
ICS ON statement). For more information about enabling and disabling ODS Graphics, see the section
“Enabling and Disabling ODS Graphics” on page 607 in Chapter 21, “Statistical Graphics Using ODS.”
The overall appearance of graphs is controlled by ODS styles. Styles and other aspects of using ODS
Graphics are discussed in the section “A Primer on ODS Statistical Graphics” on page 606 in Chapter 21,
“Statistical Graphics Using ODS.”
You must also specify the options in the PROC LOGISTIC statement that are indicated in Table 73.14.
When ODS Graphics is enabled, then the EFFECT, EFFECTPLOT, ESTIMATE, LSMEANS, LSMESTI-
MATE, and SLICE statements can produce plots that are associated with their analyses. For information
about these plots, see the corresponding sections of Chapter 19, “Shared Concepts and Topics.”
PROC LOGISTIC assigns a name to each graph it creates using ODS. You can use these names to reference
the graphs when using ODS. The names are listed in Table 73.14.
data Remission;
input remiss cell smear infil li blast temp;
label remiss='Complete Remission';
datalines;
1 .8 .83 .66 1.9 1.1 .996
1 .9 .36 .32 1.4 .74 .992
0 .8 .88 .7 .8 .176 .982
0 1 .87 .87 .7 1.053 .986
1 .9 .75 .68 1.3 .519 .98
0 1 .65 .65 .6 .519 .982
1 .95 .97 .92 1 1.23 .992
0 .95 .87 .83 1.9 1.354 1.02
0 1 .45 .45 .8 .322 .999
0 .95 .36 .34 .5 0 1.038
0 .85 .39 .33 .7 .279 .988
0 .7 .76 .53 1.2 .146 .982
0 .8 .46 .37 .4 .38 1.006
0 .2 .39 .08 .8 .114 .99
0 1 .9 .9 1.1 1.037 .99
1 1 .84 .84 1.9 2.064 1.02
0 .65 .42 .27 .5 .114 1.014
0 1 .75 .75 1 1.322 1.004
0 .5 .44 .22 .6 .114 .99
1 1 .63 .63 1.1 1.072 .986
0 1 .33 .33 .4 .176 1.01
0 .9 .93 .84 .6 1.591 1.02
1 1 .58 .58 1 .531 1.002
0 .95 .32 .3 1.6 .886 .988
1 1 .6 .6 1.7 .964 .99
1 1 .69 .69 .9 .398 .986
0 1 .73 .73 .7 .398 .986
;
The following invocation of PROC LOGISTIC illustrates the use of stepwise selection to identify the
prognostic factors for cancer remission. A significance level of 0.3 is required to allow a variable into
the model (SLENTRY=0.3), and a significance level of 0.35 is required for a variable to stay in the model
(SLSTAY=0.35). A detailed account of the variable selection process is requested by specifying the DETAILS
option. The Hosmer and Lemeshow goodness-of-fit test for the final selected model is requested by specifying
the LACKFIT option. The OUTEST= and COVOUT options in the PROC LOGISTIC statement create a
data set that contains parameter estimates and their covariances for the final selected model. The response
Example 73.1: Stepwise Logistic Regression and Predicted Values F 5519
variable option EVENT= chooses remiss=1 (remission) as the event so that the probability of remission is
modeled. The OUTPUT statement creates a data set that contains the cumulative predicted probabilities and
the corresponding confidence limits, and the individual and cross validated predicted probabilities for each
observation. The ODS OUTPUT statement writes the “Association” table from each selection step to a SAS
data set.
-2 Log L = 34.372
In Step 1 (Output 73.1.2), the variable li is selected into the model because it is the most significant variable
among those to be chosen (p D 0:0049 < 0:3). The intermediate model that contains an intercept and li is
then fitted. li remains significant (p D 0:0146 < 0:35) and is not removed.
In Step 2 (Output 73.1.3), the variable temp is added to the model. The model then contains an intercept and
the variables li and temp. Both li and temp remain significant at 0.35 level; therefore, neither li nor temp is
removed from the model.
Output 73.1.3 Step 2 of the Stepwise Analysis
Step 2. Effect temp entered:
In Step 3 (Output 73.1.4), the variable cell is added to the model. The model then contains an intercept and
the variables li, temp, and cell. None of these variables are removed from the model because all are significant
at the 0.35 level.
Example 73.1: Stepwise Logistic Regression and Predicted Values F 5523
Finally, none of the remaining variables outside the model meet the entry criterion, and the stepwise selection
is terminated. A summary of the stepwise selection is displayed in Output 73.1.5.
Results of the Hosmer and Lemeshow test are shown in Output 73.1.6. There is no evidence of a lack of fit in
the selected model .p D 0:5054/.
The data set betas created by the OUTEST= and COVOUT options is displayed in Output 73.1.7. The
data set contains parameter estimates and the covariance matrix for the final selected model. Note that all
explanatory variables listed in the MODEL statement are included in this data set; however, variables that are
not included in the final model have all missing values.
Obs _LINK_ _TYPE_ _STATUS_ _NAME_ Intercept cell smear infil li blast temp _LNLIKE_ _ESTTYPE_
1 LOGIT PARMS 0 Converged remiss 67.63 9.652 . . 3.8671 . -82.07 -10.9767 MLE
2 LOGIT COV 0 Converged Intercept 3236.19 157.097 . . 64.5726 . -3483.23 -10.9767 MLE
3 LOGIT COV 0 Converged cell 157.10 60.079 . . 6.9454 . -223.67 -10.9767 MLE
4 LOGIT COV 0 Converged smear . . . . . . . -10.9767 MLE
5 LOGIT COV 0 Converged infil . . . . . . . -10.9767 MLE
6 LOGIT COV 0 Converged li 64.57 6.945 . . 3.1623 . -75.35 -10.9767 MLE
7 LOGIT COV 0 Converged blast . . . . . . . -10.9767 MLE
8 LOGIT COV 0 Converged temp -3483.23 -223.669 . . -75.3513 . 3808.42 -10.9767 MLE
The data set pred created by the OUTPUT statement is displayed in Output 73.1.8. It contains all the variables
in the input data set, the variable phat for the (cumulative) predicted probability, the variables lcl and ucl
for the lower and upper confidence limits for the probability, and four other variables (IP_1, IP_0, XP_1,
and XP_0) for the PREDPROBS= option. The data set also contains the variable _LEVEL_, indicating
the response value to which phat, lcl, and ucl refer. For instance, for the first row of the OUTPUT data
set, the values of _LEVEL_ and phat, lcl, and ucl are 1, 0.72265, 0.16892, and 0.97093, respectively; this
means that the estimated probability that remiss=1 is 0.723 for the given explanatory variable values, and
the corresponding 95% confidence interval is (0.16892, 0.97093). The variables IP_1 and IP_0 contain the
predicted probabilities that remiss=1 and remiss=0, respectively. Note that values of phat and IP_1 are
identical because they both contain the probabilities that remiss=1. The variables XP_1 and XP_0 contain
the cross validated predicted probabilities that remiss=1 and remiss=0, respectively.
5526 F Chapter 73: The LOGISTIC Procedure
Obs remiss cell smear infil li blast temp _FROM_ _INTO_ IP_0 IP_1 XP_0 XP_1 _LEVEL_ phat lcl ucl
1 1 0.80 0.83 0.66 1.9 1.100 0.996 1 1 0.27735 0.72265 0.43873 0.56127 1 0.72265 0.16892 0.97093
2 1 0.90 0.36 0.32 1.4 0.740 0.992 1 1 0.42126 0.57874 0.47461 0.52539 1 0.57874 0.26788 0.83762
3 0 0.80 0.88 0.70 0.8 0.176 0.982 0 0 0.89540 0.10460 0.87060 0.12940 1 0.10460 0.00781 0.63419
4 0 1.00 0.87 0.87 0.7 1.053 0.986 0 0 0.71742 0.28258 0.67259 0.32741 1 0.28258 0.07498 0.65683
5 1 0.90 0.75 0.68 1.3 0.519 0.980 1 1 0.28582 0.71418 0.36901 0.63099 1 0.71418 0.25218 0.94876
6 0 1.00 0.65 0.65 0.6 0.519 0.982 0 0 0.72911 0.27089 0.67269 0.32731 1 0.27089 0.05852 0.68951
7 1 0.95 0.97 0.92 1.0 1.230 0.992 1 0 0.67844 0.32156 0.72923 0.27077 1 0.32156 0.13255 0.59516
8 0 0.95 0.87 0.83 1.9 1.354 1.020 0 1 0.39277 0.60723 0.09906 0.90094 1 0.60723 0.10572 0.95287
9 0 1.00 0.45 0.45 0.8 0.322 0.999 0 0 0.83368 0.16632 0.80864 0.19136 1 0.16632 0.03018 0.56123
10 0 0.95 0.36 0.34 0.5 0.000 1.038 0 0 0.99843 0.00157 0.99840 0.00160 1 0.00157 0.00000 0.68962
11 0 0.85 0.39 0.33 0.7 0.279 0.988 0 0 0.92715 0.07285 0.91723 0.08277 1 0.07285 0.00614 0.49982
12 0 0.70 0.76 0.53 1.2 0.146 0.982 0 0 0.82714 0.17286 0.63838 0.36162 1 0.17286 0.00637 0.87206
13 0 0.80 0.46 0.37 0.4 0.380 1.006 0 0 0.99654 0.00346 0.99644 0.00356 1 0.00346 0.00001 0.46530
14 0 0.20 0.39 0.08 0.8 0.114 0.990 0 0 0.99982 0.00018 0.99981 0.00019 1 0.00018 0.00000 0.96482
15 0 1.00 0.90 0.90 1.1 1.037 0.990 0 1 0.42878 0.57122 0.35354 0.64646 1 0.57122 0.25303 0.83973
16 1 1.00 0.84 0.84 1.9 2.064 1.020 1 1 0.28530 0.71470 0.47213 0.52787 1 0.71470 0.15362 0.97189
17 0 0.65 0.42 0.27 0.5 0.114 1.014 0 0 0.99938 0.00062 0.99937 0.00063 1 0.00062 0.00000 0.62665
18 0 1.00 0.75 0.75 1.0 1.322 1.004 0 0 0.77711 0.22289 0.73612 0.26388 1 0.22289 0.04483 0.63670
19 0 0.50 0.44 0.22 0.6 0.114 0.990 0 0 0.99846 0.00154 0.99842 0.00158 1 0.00154 0.00000 0.79644
20 1 1.00 0.63 0.63 1.1 1.072 0.986 1 1 0.35089 0.64911 0.42053 0.57947 1 0.64911 0.26305 0.90555
21 0 1.00 0.33 0.33 0.4 0.176 1.010 0 0 0.98307 0.01693 0.98170 0.01830 1 0.01693 0.00029 0.50475
22 0 0.90 0.93 0.84 0.6 1.591 1.020 0 0 0.99378 0.00622 0.99348 0.00652 1 0.00622 0.00003 0.56062
23 1 1.00 0.58 0.58 1.0 0.531 1.002 1 0 0.74739 0.25261 0.84423 0.15577 1 0.25261 0.06137 0.63597
24 0 0.95 0.32 0.30 1.6 0.886 0.988 0 1 0.12989 0.87011 0.03637 0.96363 1 0.87011 0.40910 0.98481
25 1 1.00 0.60 0.60 1.7 0.964 0.990 1 1 0.06868 0.93132 0.08017 0.91983 1 0.93132 0.44114 0.99573
26 1 1.00 0.69 0.69 0.9 0.398 0.986 1 0 0.53949 0.46051 0.62312 0.37688 1 0.46051 0.16612 0.78529
27 0 1.00 0.73 0.73 0.7 0.398 0.986 0 0 0.71742 0.28258 0.67259 0.32741 1 0.28258 0.07498 0.65683
If you want to order the selected models based on a statistic such as the AIC, R-square, or area under the
ROC curve (AUC), you can use the ODS OUTPUT statement to save the appropriate table to a data set and
then display the statistic along with the step number. For example, the following program orders the steps
according to the “c” statistic from the Association data set:
Next, a different variable selection method is used to select prognostic factors for cancer remission, and an
efficient algorithm is employed to eliminate insignificant variables from a model. The following statements
invoke PROC LOGISTIC to perform the backward elimination analysis:
Model Information
Data Set WORK.REMISSION
Response Variable remiss Complete Remission
Number of Response Levels 2
Model binary logit
Optimization Technique Fisher's scoring
Response Profile
Ordered Total
Value remiss Frequency
1 0 18
2 1 9
Probability modeled is remiss=1.
5528 F Chapter 73: The LOGISTIC Procedure
Note that you can also use the FAST option when SELECTION=STEPWISE. However, the FAST option
operates only on backward elimination steps. In this example, the stepwise process only adds variables, so
the FAST option would not be useful.
Results of the CTABLE option are shown in Output 73.1.12.
5530 F Chapter 73: The LOGISTIC Procedure
Classification Table
Correct Incorrect Percentages
Prob Non- Non- Sensi- Speci- False False
Level Event Event Event Event Correct tivity ficity POS NEG
0.060 9 0 18 0 33.3 100.0 0.0 66.7 .
0.080 9 2 16 0 40.7 100.0 11.1 64.0 0.0
0.100 9 4 14 0 48.1 100.0 22.2 60.9 0.0
0.120 9 4 14 0 48.1 100.0 22.2 60.9 0.0
0.140 9 7 11 0 59.3 100.0 38.9 55.0 0.0
0.160 9 10 8 0 70.4 100.0 55.6 47.1 0.0
0.180 9 10 8 0 70.4 100.0 55.6 47.1 0.0
0.200 8 13 5 1 77.8 88.9 72.2 38.5 7.1
0.220 8 13 5 1 77.8 88.9 72.2 38.5 7.1
0.240 8 13 5 1 77.8 88.9 72.2 38.5 7.1
0.260 6 13 5 3 70.4 66.7 72.2 45.5 18.8
0.280 6 13 5 3 70.4 66.7 72.2 45.5 18.8
0.300 6 13 5 3 70.4 66.7 72.2 45.5 18.8
0.320 6 14 4 3 74.1 66.7 77.8 40.0 17.6
0.340 5 14 4 4 70.4 55.6 77.8 44.4 22.2
0.360 5 14 4 4 70.4 55.6 77.8 44.4 22.2
0.380 5 15 3 4 74.1 55.6 83.3 37.5 21.1
0.400 5 15 3 4 74.1 55.6 83.3 37.5 21.1
0.420 5 15 3 4 74.1 55.6 83.3 37.5 21.1
0.440 5 15 3 4 74.1 55.6 83.3 37.5 21.1
0.460 4 16 2 5 74.1 44.4 88.9 33.3 23.8
0.480 4 16 2 5 74.1 44.4 88.9 33.3 23.8
0.500 4 16 2 5 74.1 44.4 88.9 33.3 23.8
0.520 4 16 2 5 74.1 44.4 88.9 33.3 23.8
0.540 3 16 2 6 70.4 33.3 88.9 40.0 27.3
0.560 3 16 2 6 70.4 33.3 88.9 40.0 27.3
0.580 3 16 2 6 70.4 33.3 88.9 40.0 27.3
0.600 3 16 2 6 70.4 33.3 88.9 40.0 27.3
0.620 3 16 2 6 70.4 33.3 88.9 40.0 27.3
0.640 3 16 2 6 70.4 33.3 88.9 40.0 27.3
0.660 3 16 2 6 70.4 33.3 88.9 40.0 27.3
0.680 3 16 2 6 70.4 33.3 88.9 40.0 27.3
0.700 3 16 2 6 70.4 33.3 88.9 40.0 27.3
0.720 2 16 2 7 66.7 22.2 88.9 50.0 30.4
0.740 2 16 2 7 66.7 22.2 88.9 50.0 30.4
0.760 2 16 2 7 66.7 22.2 88.9 50.0 30.4
0.780 2 16 2 7 66.7 22.2 88.9 50.0 30.4
0.800 2 17 1 7 70.4 22.2 94.4 33.3 29.2
0.820 2 17 1 7 70.4 22.2 94.4 33.3 29.2
0.840 0 17 1 9 63.0 0.0 94.4 100.0 34.6
0.860 0 17 1 9 63.0 0.0 94.4 100.0 34.6
0.880 0 17 1 9 63.0 0.0 94.4 100.0 34.6
0.900 0 17 1 9 63.0 0.0 94.4 100.0 34.6
0.920 0 17 1 9 63.0 0.0 94.4 100.0 34.6
0.940 0 17 1 9 63.0 0.0 94.4 100.0 34.6
0.960 0 18 0 9 66.7 0.0 100.0 . 33.3
Example 73.2: Logistic Modeling with Categorical Predictors F 5531
Each row of the “Classification Table” corresponds to a cutpoint applied to the predicted probabilities, which
is given in the Prob Level column. The 2 2 frequency tables of observed and predicted responses are given
by the next four columns. For example, with a cutpoint of 0.5, 4 events and 16 nonevents were classified
correctly. On the other hand, 2 nonevents were incorrectly classified as events and 5 events were incorrectly
classified as nonevents. For this cutpoint, the correct classification rate is 20/27 (=74.1%), which is given
in the sixth column. Accuracy of the classification is summarized by the sensitivity, specificity, and false
positive and negative rates, which are displayed in the last four columns. You can control the number of
cutpoints used, and their values, by using the PPROB= option.
data Neuralgia;
input Treatment $ Sex $ Age Duration Pain $ @@;
datalines;
P F 68 1 No B M 74 16 No P F 67 30 No
P M 66 26 Yes B F 67 28 No B F 77 16 No
A F 71 12 No B F 72 50 No B F 76 9 Yes
A M 71 17 Yes A F 63 27 No A F 69 18 Yes
B F 66 12 No A M 62 42 No P F 64 1 Yes
A F 64 17 No P M 74 4 No A F 72 25 No
P M 70 1 Yes B M 66 19 No B M 59 29 No
A F 64 30 No A M 70 28 No A M 69 1 No
B F 78 1 No P M 83 1 Yes B F 69 42 No
B M 75 30 Yes P M 77 29 Yes P F 79 20 Yes
A M 70 12 No A F 69 12 No B F 65 14 No
B M 70 1 No B M 67 23 No A M 76 25 Yes
P M 78 12 Yes B M 77 1 Yes B F 69 24 No
P M 66 4 Yes P F 65 29 No P M 60 26 Yes
A M 78 15 Yes B M 75 21 Yes A F 67 11 No
P F 72 27 No P F 70 13 Yes A M 75 6 Yes
B F 65 7 No P F 68 27 Yes P M 68 11 Yes
P M 67 17 Yes B M 70 22 No A M 65 15 No
P F 67 1 Yes A M 67 10 No P F 72 11 Yes
A F 74 1 No B M 80 21 Yes A F 69 3 No
;
The data set Neuralgia contains five variables: Treatment, Sex, Age, Duration, and Pain. The last variable,
Pain, is the response variable. A specification of Pain=Yes indicates there was pain, and Pain=No indicates
no pain. The variable Treatment is a categorical variable with three levels: A and B represent the two test
treatments, and P represents the placebo treatment. The gender of the patients is given by the categorical
variable Sex. The variable Age is the age of the patients, in years, when treatment began. The duration of
complaint, in months, before the treatment began is given by the variable Duration.
The following statements use the LOGISTIC procedure to fit a two-way logit with interaction model for the
effect of Treatment and Sex, with Age and Duration as covariates. The categorical variables Treatment and
Sex are declared in the CLASS statement.
5532 F Chapter 73: The LOGISTIC Procedure
PROC LOGISTIC displays a table of the Type 3 analysis of effects based on the Wald test (Output 73.2.2).
Note that the Treatment*Sex interaction and the duration of complaint are not statistically significant (p =
0.9318 and p = 0.8752, respectively). This indicates that there is no evidence that the treatments affect pain
differently in men and women, and no evidence that the pain outcome is related to the duration of pain.
Joint Tests
Wald
Effect DF Chi-Square Pr > ChiSq
Treatment 2 11.9886 0.0025
Sex 1 5.3104 0.0212
Treatment*Sex 2 0.1412 0.9318
Age 1 7.2744 0.0070
Duration 1 0.0247 0.8752
Note: Under full-rank parameterizations, Type 3 effect tests are replaced by joint tests. The joint test for an effect is a test that all the
parameters associated with that effect are zero. Such joint tests might not be equivalent to Type 3 effect tests under GLM
parameterization.
Parameter estimates are displayed in Output 73.2.3. The Exp(Est) column contains the exponentiated
parameter estimates requested with the EXPB option. These values can, but do not necessarily, represent odds
ratios for the corresponding variables. For continuous explanatory variables, the Exp(Est) value corresponds
to the odds ratio for a unit increase of the corresponding variable. For CLASS variables that use effect coding,
the Exp(Est) values have no direct interpretation as a comparison of levels. However, when the reference
coding is used, the Exp(Est) values represent the odds ratio between the corresponding level and the reference
level. Following the parameter estimates table, PROC LOGISTIC displays the odds ratio estimates for those
variables that are not involved in any interaction terms. If the variable is a CLASS variable, the odds ratio
estimate comparing each level with the reference level is computed regardless of the coding scheme. In this
analysis, because the model contains the Treatment*Sex interaction term, the odds ratios for Treatment and
Example 73.2: Logistic Modeling with Categorical Predictors F 5533
Sex were not computed. The odds ratio estimates for Age and Duration are precisely the values given in the
Exp(Est) column in the parameter estimates table.
The following PROC LOGISTIC statements illustrate the use of forward selection on the data set Neuralgia
to identify the effects that differentiate the two Pain responses. The option SELECTION=FORWARD is
specified to carry out the forward selection. The term Treatment|Sex@2 illustrates another way to specify
main effects and two-way interactions. (Note that, in this case, the “@2” is unnecessary because no
interactions besides the two-way interaction are possible).
Output 73.2.5 shows the Type 3 analysis of effects, the parameter estimates, and the odds ratio estimates
for the selected model. All three variables, Treatment, Age, and Sex, are statistically significant at the 0.05
5534 F Chapter 73: The LOGISTIC Procedure
level (p=0.0018, p=0.0213, and p=0.0057, respectively). Because the selected model does not contain the
Treatment*Sex interaction, odds ratios for Treatment and Sex are computed. The estimated odds ratio is
24.022 for treatment A versus placebo, 41.528 for Treatment B versus placebo, and 6.194 for female patients
versus male patients. Note that these odds ratio estimates are not the same as the corresponding values in the
Exp(Est) column in the parameter estimates table because effect coding was used. From Output 73.2.5, it is
evident that both Treatment A and Treatment B are better than the placebo in reducing pain; females tend to
have better improvement than males; and younger patients are faring better than older patients.
Output 73.2.5 Type 3 Effects and Parameter Estimates with Effect Coding
Finally, the following statements refit the previously selected model, except that reference coding is used for
the CLASS variables instead of effect coding:
The ODDSRATIO statements compute the odds ratios for the covariates, and the NOOR option suppresses
the default odds ratio table. Four CONTRAST statements are specified; they provide another method of
producing the odds ratios. The three contrasts labeled ‘Pairwise’ specify a contrast vector, L, for each of
the pairwise comparisons between the three levels of Treatment. The contrast labeled ‘Female vs Male’
compares female to male patients. The option ESTIMATE=EXP is specified in all CONTRAST statements
to exponentiate the estimates of L0 ˇ. With the given specification of contrast coefficients, the first of the
‘Pairwise’ CONTRAST statements corresponds to the odds ratio of A versus P, the second corresponds to B
versus P, and the third corresponds to A versus B. You can also specify the ‘Pairwise’ contrasts in a single
contrast statement with three rows. The ‘Female vs Male’ CONTRAST statement corresponds to the odds
ratio that compares female to male patients.
The PLOTS(ONLY)= option displays only the requested odds ratio plot when ODS Graphics is enabled.
The EFFECTPLOT statements do not honor the ONLY option, and display the fitted model. The first
EFFECTPLOT statement by default produces a plot of the predicted values against the continuous Age
variable, grouped by the Treatment levels. The AT option produces one plot for males and another for females;
the NOOBS option suppresses the display of the observations. In the second EFFECTPLOT statement, a
SLICEFIT plot is specified to display the Age variable on the X axis, the fits are grouped by the Sex levels,
and the PLOTBY= option produces a panel of plots that displays each level of the Treatment variable.
The reference coding is shown in Output 73.2.6. The Type 3 analysis of effects and the parameter estimates for
the reference coding are displayed in Output 73.2.7. Although the parameter estimates are different because
of the different parameterizations, the “Type 3 Analysis of Effects” table remains the same as in Output 73.2.5.
With effect coding, the treatment A parameter estimate (0.8772) estimates the effect of treatment A compared
to the average effect of treatments A, B, and placebo. The treatment A estimate (3.1790) under the reference
coding estimates the difference in effect of treatment A and the placebo treatment.
Output 73.2.7 Type 3 Effects and Parameter Estimates with Reference Coding
The ODDSRATIO statement results are shown in Output 73.2.8, and the resulting plot is displayed in
Output 73.2.9. Note in Output 73.2.9 that the odds ratio confidence limits are truncated due to specifying the
RANGE=CLIP option; this enables you to see which intervals contain “1” more clearly. The odds ratios are
identical to those shown in the “Odds Ratio Estimates” table in Output 73.2.5 with the addition of the odds
ratio for “Treatment A vs B”. Both treatments A and B are highly effective over placebo in reducing pain, as
can be seen from the odds ratios comparing treatment A against P and treatment B against P (the second
and third rows in the table). However, the 95% confidence interval for the odds ratio comparing treatment A
to B is (0.0932, 3.5889), indicating that the pain reduction effects of these two test treatments are not very
different. Again, the ’Sex F vs M’ odds ratio shows that female patients fared better in obtaining relief from
pain than male patients. The odds ratio for Age shows that a patient one year older is 0.77 times as likely to
show no pain; that is, younger patients have more improvement than older patients.
Output 73.2.10 contains two tables: the “Contrast Test Results” table and the “Contrast Estimation and
Testing Results by Row” table. The former contains the overall Wald test for each CONTRAST statement.
The latter table contains estimates and tests of individual contrast rows. The estimates for the first two rows
of the ’Pairwise’ CONTRAST statements are the same as those given in the two preceding odds ratio tables
(Output 73.2.7 and Output 73.2.8). The third row estimates the odds ratio comparing A to B, agreeing with
Output 73.2.8, and the last row computes the odds ratio comparing pain relief for females to that for males.
ANCOVA-style plots of the model-predicted probabilities against the Age variable for each combination of
Treatment and Sex are displayed in Output 73.2.11 and Output 73.2.12. These plots confirm that females
always have a higher probability of pain reduction in each treatment group, the placebo treatment has a lower
probability of success than the other treatments, and younger patients respond to treatment better than older
patients.
data Cheese;
do Additive = 1 to 4;
do y = 1 to 9;
input freq @@;
output;
end;
end;
label y='Taste Rating';
datalines;
0 0 1 7 8 8 19 8 1
6 9 12 11 7 6 1 0 0
1 1 6 8 23 7 5 1 0
0 0 0 1 3 7 14 16 11
;
The response variable y is ordinally scaled. A cumulative logit model is used to investigate the effects of
the cheese additives on taste. The following statements invoke PROC LOGISTIC to fit this model with
y as the response variable and three indicator variables as explanatory variables, with the fourth additive
as the reference level. With this parameterization, each Additive parameter compares an additive to the
fourth additive. The COVB option displays the estimated covariance matrix, and the NOODDSRATIO
option suppresses the default odds ratio table. The ODDSRATIO statement computes odds ratios for all
combinations of the Additive levels. The PLOTS option produces a graphical display of the odds ratios, and
the EFFECTPLOT statement displays the predicted probabilities.
Additive1 indicates a tendency toward the lower-numbered categories of the first cheese additive relative
to the fourth. In other words, the fourth additive tastes better than the first additive. The second and third
additives are both less favorable than the fourth additive. The relative magnitudes of these slope estimates
imply the preference ordering: fourth, first, third, second.
Model Information
Data Set WORK.CHEESE
Response Variable y Taste Rating
Number of Response Levels 9
Frequency Variable freq
Model cumulative logit
Optimization Technique Fisher's scoring
Response Profile
Ordered Total
Value y Frequency
1 1 7
2 2 10
3 3 19
4 4 27
5 5 41
6 6 28
7 7 39
8 8 25
9 9 12
Probabilities modeled are cumulated over the lower Ordered Values.
Note: 8 observations having nonpositive frequencies or weights were excluded since they do not contribute to the
analysis.
The odds ratio results in Output 73.3.2 show the preferences more clearly. For example, the “Additive 1 vs
4” odds ratio says that the first additive has 5.017 times the odds of receiving a lower score than the fourth
additive; that is, the first additive is 5.017 times more likely than the fourth additive to receive a lower score.
Example 73.3: Ordinal Logistic Regression F 5543
Output 73.3.3 displays the odds ratios graphically; the range of the confidence limits is truncated by the
RANGE=CLIP option, so you can see that “1” is not contained in any of the intervals.
Output 73.3.5 displays the probability of each taste rating y within each additive. You can see that Additive=1
mostly receives ratings of 5 to 7, Additive=2 mostly receives ratings of 2 to 5, Additive=3 mostly receives
ratings of 4 to 6, and Additive=4 mostly receives ratings of 7 to 9, which also confirms the previously
discussed preference orderings.
Alternatively, you can use an adjacent-category logit model. The following statements invoke PROC
LOGISTIC to fit this model:
The odds ratio results in Output 73.3.7 show the same preferences that you obtain from the proportional odds
model. Output 73.3.8 displays the probability of each additive according to the taste rating.
The levels of the response variable (self, team, and class) have no essential ordering, so a logistic regression
is performed on the generalized logits. The model to be fit is
hij
log D ˛j C x0hi ˇj
hi r
where hij is the probability that a student in school h and program i prefers teaching style j, j ¤ r, and style
r is the baseline style (in this case, class). There are separate sets of intercept parameters ˛j and regression
parameters ˇj for each logit, and the vector xhi is the set of explanatory variables for the hith population.
Thus, two logits are modeled for each school and program combination: the logit comparing self to class and
the logit comparing team to class.
The following statements create the data set school and request the analysis. The LINK=GLOGIT option
forms the generalized logits. The response variable option ORDER=DATA means that the response variable
levels are ordered as they exist in the data set: self, team, and class; thus, the logits are formed by comparing
self to class and by comparing team to class. The ODDSRATIO statement produces odds ratios in the
presence of interactions, and a graphical display of the requested odds ratios is produced when ODS Graphics
is enabled.
data school;
length Program $ 9;
input School Program $ Style $ Count @@;
datalines;
1 regular self 10 1 regular team 17 1 regular class 26
1 afternoon self 5 1 afternoon team 12 1 afternoon class 50
2 regular self 21 2 regular team 17 2 regular class 26
2 afternoon self 16 2 afternoon team 12 2 afternoon class 36
3 regular self 15 3 regular team 15 3 regular class 16
3 afternoon self 12 3 afternoon team 12 3 afternoon class 20
;
Summary information about the model, the response variable, and the classification variables are displayed in
Output 73.4.1.
Model Information
Data Set WORK.SCHOOL
Response Variable Style
Number of Response Levels 3
Frequency Variable Count
Model generalized logit
Optimization Technique Newton-Raphson
Response Profile
Ordered Total
Value Style Frequency
1 self 79
2 team 85
3 class 174
Logits modeled use Style='class' as the reference category.
The “Testing Global Null Hypothesis: BETA=0” table in Output 73.4.2 shows that the parameters are
significantly different from zero.
However, the “Type 3 Analysis of Effects” table in Output 73.4.3 shows that the interaction effect is clearly
nonsignificant.
Joint Tests
Wald
Effect DF Chi-Square Pr > ChiSq
School 4 14.5522 0.0057
Program 2 10.4815 0.0053
School*Program 4 1.7439 0.7827
Note: Under full-rank parameterizations, Type 3 effect tests are replaced by joint tests. The joint test for an effect is a test that all the
parameters associated with that effect are zero. Such joint tests might not be equivalent to Type 3 effect tests under GLM
parameterization.
The table produced by the ODDSRATIO statement is displayed in Output 73.4.4. The differences between
the program preferences are small across all the styles (logits) compared to their variability as displayed by
the confidence limits in Output 73.4.5, confirming that the interaction effect is nonsignificant.
5550 F Chapter 73: The LOGISTIC Procedure
Because the interaction effect is clearly nonsignificant, a main-effects model is fit with the following
statements. The EFFECTPLOT statement creates a plot of the predicted values versus the levels of the School
variable at each level of the Program variables. The CLM option adds confidence bars, and the NOOBS
option suppresses the display of the observations.
The parameter estimates, tests for individual parameters, and odds ratios are displayed in Output 73.4.7. The
Program variable has nearly the same effect on both logits, while School=1 has the largest effect of the
schools.
Output 73.4.7 Estimates
The interaction plots in Output 73.4.8 show that School=1 and Program=afternoon have a preference for
the traditional classroom style. Of course, because these are not simultaneous confidence intervals, the
nonoverlapping 95% confidence limits do not take the place of an actual test.
Example 73.4: Nominal Response Data: Generalized Logits Model F 5553
data Screen;
do Disease='Present','Absent';
do Test=1,0;
input Count @@;
output;
end;
end;
datalines;
950 50
10 990
;
Classification Table
Correct Incorrect Percentages
Prob Prob Non- Non- Sensi- Speci- False False
Event Level Event Event Event Event Correct tivity ficity POS NEG
0.500 0.500 950 990 10 50 97.0 95.0 99.0 1.0 4.8
0.010 0.500 950 990 10 50 99.0 95.0 99.0 51.0 0.1
In the classification table, the column “Prob Level” represents the cutoff values (the settings of the PPROB=
option) for predicting whether an observation is an event. The “Correct” columns list the numbers of subjects
that are correctly predicted as events and nonevents, respectively, and the “Incorrect” columns list the number
of nonevents incorrectly predicted as events and the number of events incorrectly predicted as nonevents,
Example 73.6: Logistic Regression Diagnostics F 5555
respectively. For PEVENT=0.5, the false positive rate is 1% and the false negative rate is 4.8%. These results
ignore the fact that the samples were stratified and incorrectly assume that the overall sample proportion of
disease (which is 0.5) estimates the true disease rate. For a true disease rate of 0.01, the false positive rate and
the false negative rate are 51% and 0.1%, respectively, as shown in the second line of the classification table.
data vaso;
length Response $12;
input Volume Rate Response @@;
LogVolume=log(Volume);
LogRate=log(Rate);
datalines;
3.70 0.825 constrict 3.50 1.09 constrict
1.25 2.50 constrict 0.75 1.50 constrict
0.80 3.20 constrict 0.70 3.50 constrict
0.60 0.75 no_constrict 1.10 1.70 no_constrict
0.90 0.75 no_constrict 0.90 0.45 no_constrict
0.80 0.57 no_constrict 0.55 2.75 no_constrict
0.60 3.00 no_constrict 1.40 2.33 constrict
0.75 3.75 constrict 2.30 1.64 constrict
3.20 1.60 constrict 0.85 1.415 constrict
1.70 1.06 no_constrict 1.80 1.80 constrict
0.40 2.00 no_constrict 0.95 1.36 no_constrict
1.35 1.35 no_constrict 1.50 1.36 no_constrict
1.60 1.78 constrict 0.60 1.50 no_constrict
1.80 1.50 constrict 0.95 1.90 no_constrict
1.90 0.95 constrict 1.60 0.40 no_constrict
2.70 0.75 constrict 2.35 0.03 no_constrict
1.10 1.83 no_constrict 1.10 2.20 constrict
1.20 2.00 constrict 0.80 3.33 constrict
0.95 1.90 no_constrict 0.75 1.90 no_constrict
1.30 1.625 constrict
;
In the data set vaso, the variable Response represents the outcome of a test. The variable LogVolume
represents the log of the volume of air intake, and the variable LogRate represents the log of the rate of air
intake.
The following statements invoke PROC LOGISTIC to fit a logistic regression model to the vasoconstriction
data, where Response is the response variable, and LogRate and LogVolume are the explanatory variables.
Regression diagnostics are displayed when ODS Graphics is enabled, and the INFLUENCE option is specified
to display a table of the regression diagnostics.
5556 F Chapter 73: The LOGISTIC Procedure
Model Information
Data Set WORK.VASO
Response Variable Response
Number of Response Levels 2
Model binary logit
Optimization Technique Fisher's scoring
Response Profile
Ordered Total
Value Response Frequency
1 constrict 20
2 no_constrict 19
Probability modeled is Response='constrict'.
The INFLUENCE option displays the values of the explanatory variables (LogRate and LogVolume) for
each observation, a column for each diagnostic produced, and the case number that represents the sequence
number of the observation (Output 73.6.2).
5558 F Chapter 73: The LOGISTIC Procedure
Regression Diagnostics
Covariates
Confidence Confidence
Hat Interval Interval
Case Pearson Deviance Matrix Intercept LogRate LogVolume Displacement Displacement Delta Delta
Number LogRate LogVolume Residual Residual Diagonal DfBeta DfBeta DfBeta C CBar Deviance Chi-Square
1 -0.1924 1.3083 0.2205 0.3082 0.0927 -0.0165 0.0193 0.0556 0.00548 0.00497 0.1000 0.0536
2 0.0862 1.2528 0.1349 0.1899 0.0429 -0.0134 0.0151 0.0261 0.000853 0.000816 0.0369 0.0190
3 0.9163 0.2231 0.2923 0.4049 0.0612 -0.0492 0.0660 0.0589 0.00593 0.00557 0.1695 0.0910
4 0.4055 -0.2877 3.5181 2.2775 0.0867 1.0734 -0.9302 -1.0180 1.2873 1.1756 6.3626 13.5523
5 1.1632 -0.2231 0.5287 0.7021 0.1158 -0.0832 0.1411 0.0583 0.0414 0.0366 0.5296 0.3161
6 1.2528 -0.3567 0.6090 0.7943 0.1524 -0.0922 0.1710 0.0381 0.0787 0.0667 0.6976 0.4376
7 -0.2877 -0.5108 -0.0328 -0.0464 0.00761 -0.00280 0.00274 0.00265 8.321E-6 8.258E-6 0.00216 0.00109
8 0.5306 0.0953 -1.0196 -1.1939 0.0559 -0.1444 0.0613 0.0570 0.0652 0.0616 1.4870 1.1011
9 -0.2877 -0.1054 -0.0938 -0.1323 0.0342 -0.0178 0.0173 0.0153 0.000322 0.000311 0.0178 0.00911
10 -0.7985 -0.1054 -0.0293 -0.0414 0.00721 -0.00245 0.00246 0.00211 6.256E-6 6.211E-6 0.00172 0.000862
11 -0.5621 -0.2231 -0.0370 -0.0523 0.00969 -0.00361 0.00358 0.00319 0.000014 0.000013 0.00274 0.00138
12 1.0116 -0.5978 -0.5073 -0.6768 0.1481 -0.1173 0.0647 0.1651 0.0525 0.0447 0.5028 0.3021
13 1.0986 -0.5108 -0.7751 -0.9700 0.1628 -0.0931 -0.00946 0.1775 0.1395 0.1168 1.0577 0.7175
14 0.8459 0.3365 0.2559 0.3562 0.0551 -0.0414 0.0538 0.0527 0.00404 0.00382 0.1307 0.0693
15 1.3218 -0.2877 0.4352 0.5890 0.1336 -0.0940 0.1408 0.0643 0.0337 0.0292 0.3761 0.2186
16 0.4947 0.8329 0.1576 0.2215 0.0402 -0.0198 0.0234 0.0307 0.00108 0.00104 0.0501 0.0259
17 0.4700 1.1632 0.0709 0.1001 0.0172 -0.00630 0.00701 0.00914 0.000089 0.000088 0.0101 0.00511
18 0.3471 -0.1625 2.9062 2.1192 0.0954 0.9595 -0.8279 -0.8477 0.9845 0.8906 5.3817 9.3363
19 0.0583 0.5306 -1.0718 -1.2368 0.1315 -0.2591 0.2024 -0.00488 0.2003 0.1740 1.7037 1.3227
20 0.5878 0.5878 0.2405 0.3353 0.0525 -0.0331 0.0421 0.0518 0.00338 0.00320 0.1156 0.0610
21 0.6931 -0.9163 -0.1076 -0.1517 0.0373 -0.0180 0.0158 0.0208 0.000465 0.000448 0.0235 0.0120
22 0.3075 -0.0513 -0.4193 -0.5691 0.1015 -0.1449 0.1237 0.1179 0.0221 0.0199 0.3437 0.1956
23 0.3001 0.3001 -1.0242 -1.1978 0.0761 -0.1961 0.1275 0.0357 0.0935 0.0864 1.5212 1.1355
24 0.3075 0.4055 -1.3684 -1.4527 0.0717 -0.1281 0.0410 -0.1004 0.1558 0.1447 2.2550 2.0171
25 0.5766 0.4700 0.3347 0.4608 0.0587 -0.0403 0.0570 0.0708 0.00741 0.00698 0.2193 0.1190
26 0.4055 -0.5108 -0.1595 -0.2241 0.0548 -0.0366 0.0329 0.0373 0.00156 0.00147 0.0517 0.0269
27 0.4055 0.5878 0.3645 0.4995 0.0661 -0.0327 0.0496 0.0788 0.0101 0.00941 0.2589 0.1423
28 0.6419 -0.0513 -0.8989 -1.0883 0.0647 -0.1423 0.0617 0.1025 0.0597 0.0559 1.2404 0.8639
29 -0.0513 0.6419 0.8981 1.0876 0.1682 0.2367 -0.1950 0.0286 0.1961 0.1631 1.3460 0.9697
30 -0.9163 0.4700 -0.0992 -0.1400 0.0507 -0.0224 0.0227 0.0159 0.000554 0.000526 0.0201 0.0104
31 -0.2877 0.9933 0.6198 0.8064 0.2459 0.1165 -0.0996 0.1322 0.1661 0.1253 0.7755 0.5095
32 -3.5066 0.8544 -0.00073 -0.00103 0.000022 -3.22E-6 3.405E-6 2.48E-6 1.18E-11 1.18E-11 1.065E-6 5.324E-7
33 0.6043 0.0953 -1.2062 -1.3402 0.0510 -0.0882 -0.0137 -0.00216 0.0824 0.0782 1.8744 1.5331
34 0.7885 0.0953 0.5447 0.7209 0.0601 -0.0425 0.0877 0.0671 0.0202 0.0190 0.5387 0.3157
35 0.6931 0.1823 0.5404 0.7159 0.0552 -0.0340 0.0755 0.0711 0.0180 0.0170 0.5295 0.3091
36 1.2030 -0.2231 0.4828 0.6473 0.1177 -0.0867 0.1381 0.0631 0.0352 0.0311 0.4501 0.2641
37 0.6419 -0.0513 -0.8989 -1.0883 0.0647 -0.1423 0.0617 0.1025 0.0597 0.0559 1.2404 0.8639
38 0.6419 -0.2877 -0.4874 -0.6529 0.1000 -0.1395 0.1032 0.1397 0.0293 0.0264 0.4526 0.2639
39 0.4855 0.2624 0.7053 0.8987 0.0531 0.0326 0.0190 0.0489 0.0295 0.0279 0.8355 0.5254
Example 73.6: Logistic Regression Diagnostics F 5559
Because ODS Graphics is enabled, influence plots are displayed in Outputs 73.6.3 through 73.6.5. For general
information about ODS Graphics, see Chapter 21, “Statistical Graphics Using ODS.” For specific information
about the graphics available in the LOGISTIC procedure, see the section “ODS Graphics” on page 5516.
The vertical axis of an index plot represents the value of the diagnostic, and the horizontal axis represents the
sequence (case number) of the observation. The index plots are useful for identification of extreme values.
The index plots of the Pearson residuals and the deviance residuals (Output 73.6.3) indicate that case 4 and
case 18 are poorly accounted for by the model. The index plot of the diagonal elements of the hat matrix
(Output 73.6.3) suggests that case 31 is an extreme point in the design space. The index plots of DFBETAS
(Output 73.6.5) indicate that case 4 and case 18 are causing instability in all three parameter estimates. The
other four index plots in Outputs 73.6.3 and 73.6.4 also point to these two cases as having a large impact on
the coefficients and goodness of fit.
Other versions of diagnostic plots can be requested by specifying the appropriate options in the PLOTS=
option. For example, the following statements produce three other sets of influence diagnostic plots: the
PHAT option plots several diagnostics against the predicted probabilities (Output 73.6.6), the LEVERAGE
option plots several diagnostics against the leverage (Output 73.6.7), and the DPC option plots the deletion
diagnostics against the predicted probabilities and colors the observations according to the confidence interval
displacement diagnostic (Output 73.6.8). The LABEL option displays the observation numbers on the plots.
In all plots, you are looking for the outlying observations, and again cases 4 and 18 are noted.
data Data1;
input disease n age;
datalines;
0 14 25
0 20 35
0 19 45
7 18 55
6 12 65
17 17 75
;
Output 73.7.2 shows that the R-square for the model is 0.74. The odds of an event increases by a factor of
7.9 for each 10-year increase in age.
Because ODS Graphics is enabled, a graphical display of the ROC curve is produced as shown in Out-
put 73.7.3.
Note that the area under the ROC curve is estimated by the statistic c in the “Association of Predicted
Probabilities and Observed Responses” table. In this example, the area under the ROC curve is 0.953. By
default, the Y axis is labeled “Sensitivity” and the X axis is labeled “1–Specificity”.
Because there is only one continuous covariate and because ODS Graphics is enabled, the EFFECTPLOT
statement produces a graphical display of the predicted probability curve with bounding 95% confidence
limits as shown in Output 73.7.4.
5566 F Chapter 73: The LOGISTIC Procedure
data roc;
input alb tp totscore popind @@;
totscore = 10 - totscore;
datalines;
3.0 5.8 10 0 3.2 6.3 5 1 3.9 6.8 3 1 2.8 4.8 6 0
3.2 5.8 3 1 0.9 4.0 5 0 2.5 5.7 8 0 1.6 5.6 5 1
3.8 5.7 5 1 3.7 6.7 6 1 3.2 5.4 4 1 3.8 6.6 6 1
4.1 6.6 5 1 3.6 5.7 5 1 4.3 7.0 4 1 3.6 6.7 4 0
2.3 4.4 6 1 4.2 7.6 4 0 4.0 6.6 6 0 3.5 5.8 6 1
3.8 6.8 7 1 3.0 4.7 8 0 4.5 7.4 5 1 3.7 7.4 5 1
3.1 6.6 6 1 4.1 8.2 6 1 4.3 7.0 5 1 4.3 6.5 4 1
3.2 5.1 5 1 2.6 4.7 6 1 3.3 6.8 6 0 1.7 4.0 7 0
3.7 6.1 5 1 3.3 6.3 7 1 4.2 7.7 6 1 3.5 6.2 5 1
2.9 5.7 9 0 2.1 4.8 7 1 2.8 6.2 8 0 4.0 7.0 7 1
3.3 5.7 6 1 3.7 6.9 5 1 3.6 6.6 5 1
;
In the following statements, the NOFIT option is specified in the MODEL statement to prevent PROC
LOGISTIC from fitting the model with three covariates. Each ROC statement lists one of the covariates, and
PROC LOGISTIC then fits the model with that single covariate. Note that the original data set contains six
more records with missing values for one of the tests, but PROC LOGISTIC ignores all records with missing
values; hence there is a common sample size for each of the three models. The ROCCONTRAST statement
implements the nonparametric approach of DeLong, DeLong, and Clarke-Pearson (1988) to compare the
three ROC curves, the REFERENCE option specifies that the K-G Score curve is used as the reference curve
in the contrast, the E option displays the contrast coefficients, and the ESTIMATE option computes and
tests each comparison. With ODS Graphics enabled, the plots=roc(id=prob) specification in the PROC
LOGISTIC statement displays several plots, and the plots of individual ROC curves have certain points
labeled with their predicted probabilities.
Model Information
Data Set WORK.ROC
Response Variable popind
Number of Response Levels 2
Model binary logit
Optimization Technique Fisher's scoring
Response Profile
Ordered Total
Value popind Frequency
1 0 12
2 1 31
Probability modeled is popind=0.
For each ROC model, the model fitting details in Outputs 73.8.2, 73.8.4, and 73.8.6 can be suppressed with
the ROCOPTIONS(NODETAILS) option; however, the convergence status is always displayed.
The ROC curves for the three models are displayed in Outputs 73.8.3, 73.8.5, and 73.8.7. Note that the labels
on the ROC curve are produced by specifying the ID=PROB option, and are the predicted probabilities for
the cutpoints.
All ROC curves being compared are also overlaid on the same plot, as shown in Output 73.8.8.
5574 F Chapter 73: The LOGISTIC Procedure
Output 73.8.9 displays the association statistics, and displays the area under the ROC curve along with its
standard error and a confidence interval for each model in the comparison. The confidence interval for Total
Protein contains 0.50; hence it is not significantly different from random guessing, which is represented by
the diagonal line in the preceding ROC plots.
Output 73.8.10 shows that the contrast used ’K-G Score’ as the reference level. This table is produced by
specifying the E option in the ROCCONTRAST statement.
ROC Contrast
Coefficients
ROC Model Row1 Row2
Albumin 1 0
K-G Score -1 -1
Total Protein 0 1
Output 73.8.11 shows that the 2-degrees-of-freedom test that the ’K-G Score’ is different from at least one
other test is not significant at the 0.05 level.
Output 73.8.12 is produced by specifying the ESTIMATE option in the ROCCONTRAST statement. Each
row shows that the curves are not significantly different.
data One;
do A=0,1;
do B=0,1;
do Y=1,2;
input F @@;
output;
end;
end;
end;
datalines;
23 63 31 70 67 100 70 104
;
The following statements fit a full model to examine the main effects of A and B as well as the interaction
effect of A and B:
Model Information
Data Set WORK.ONE
Response Variable Y
Number of Response Levels 2
Frequency Variable F
Model binary logit
Optimization Technique Fisher's scoring
Response Profile
Ordered Total
Value Y Frequency
1 1 191
2 2 337
Probability modeled is Y=1.
Pearson and deviance goodness-of-fit tests cannot be obtained for this model because a full model containing
four parameters is fit, leaving no residual degrees of freedom. For a binary response model, the goodness-of-
fit tests have m q degrees of freedom, where m is the number of subpopulations and q is the number of
model parameters. In the preceding model, m D q D 4, resulting in zero degrees of freedom for the tests.
The following statements fit a reduced model containing only the A effect, so two degrees of freedom become
available for testing goodness of fit. Specifying the SCALE=NONE option requests the Pearson and deviance
statistics. With single-trial syntax, the AGGREGATE= option is needed to define the subpopulations in the
study. Specifying AGGREGATE=(A B) creates subpopulations of the four combinations of levels of A and B.
Although the B effect is being dropped from the model, it is still needed to define the original subpopulations
in the study. If AGGREGATE=(A) were specified, only two subpopulations would be created from the levels
of A, resulting in m D q D 2 and zero degrees of freedom for the tests.
The goodness-of-fit tests in Output 73.9.2 show that dropping the B main effect and the A*B interaction
simultaneously does not result in significant lack of fit of the model. The tests’ large p-values indicate
insufficient evidence for rejecting the null hypothesis that the model fits.
data seeds;
input pot n r cult soil;
datalines;
1 16 8 0 0
2 51 26 0 0
3 45 23 0 0
4 39 10 0 0
5 36 9 0 0
6 81 23 1 0
7 30 10 1 0
8 39 17 1 0
9 28 8 1 0
10 62 23 1 0
11 51 32 0 1
12 72 55 0 1
13 41 22 0 1
14 12 3 0 1
15 13 10 0 1
16 79 46 1 1
17 30 15 1 1
18 51 32 1 1
19 74 53 1 1
20 56 12 1 1
;
PROC LOGISTIC is used as follows to fit a logit model to the data, with cult, soil, and cult soil interaction
as explanatory variables. The option SCALE=NONE is specified to display goodness-of-fit statistics.
Example 73.10: Overdispersion F 5579
Output 73.10.1 Results of the Model Fit for the Two-Way Layout
Full Model With SCALE=NONE
If the link function and the model specification are correct and if there are no outliers, then the lack
of fit might be due to overdispersion. Without adjusting for the overdispersion, the standard errors are
likely to be underestimated, causing the Wald tests to be too sensitive. In PROC LOGISTIC, there are
three SCALE= options to accommodate overdispersion. With unequal sample sizes for the observations,
5580 F Chapter 73: The LOGISTIC Procedure
SCALE=WILLIAMS is preferred. The Williams model estimates a scale parameter by equating the value
of Pearson 2 for the full model to its approximate expected value. The full model considered in the following
statements is the model with cultivar, soil condition, and their interaction. Using a full model reduces the risk
of contaminating with lack of fit due to incorrect model specification.
Model Information
Data Set WORK.SEEDS
Response Variable (Events) r
Response Variable (Trials) n
Weight Variable 1 / ( 1 + 0.075941 * (n - 1) )
Model binary logit
Optimization Technique Fisher's scoring
Response Profile
Ordered Binary Total Total
Value Outcome Frequency Weight
1 Event 437 92.95346
2 Nonevent 469 105.36819
Because neither cult nor cult soil is statistically significant (p = 0.5290 and p = 0.9274, respectively),
a reduced model that contains only the soil condition factor is fitted, with the observations weighted by
1=.1 C 0:075941.N 1//. This can be done conveniently in PROC LOGISTIC by including the scale
estimate in the SCALE=WILLIAMS option as follows:
data Data1;
do ID=1 to 63;
do Outcome = 1 to 0 by -1;
input Gall Hyper @@;
output;
end;
end;
datalines;
0 0 0 0 0 0 0 0 0 1 0 1 0 0 1 0 1 0 0 1
0 1 0 0 1 0 0 0 1 1 0 1 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 1 1 0 0 1 1 0 1 0 1 0 0 1
0 1 0 0 0 0 1 1 0 0 1 1 0 0 0 1 0 1 0 0
0 0 1 1 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0
0 0 0 1 1 0 0 1 0 0 0 1 1 0 0 0 0 1 0 0
0 1 0 0 0 1 0 0 0 1 0 0 0 0 0 0 1 1 1 1
0 0 0 1 0 1 0 0 0 1 0 1 0 1 0 1 0 1 0 0
0 0 0 0 0 1 1 0 0 0 0 1 0 0 0 0 1 0 0 0
0 0 0 0 1 1 0 0 0 1 0 0 0 0 0 0 0 1 0 1
0 0 0 0 0 1 0 1 0 1 0 0 0 1 0 0 1 0 0 0
0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 0 1 1 0 0
1 0 1 0 0 1 0 0 1 0 0 0
;
There are several ways to approach this problem with PROC LOGISTIC:
Specify EXACT and STRATA statements to perform an exact logistic regression on the original data
set, if you believe the data set is too small or too sparse for the usual asymptotics to hold.
Transform each matched pair into a single observation, and then specify a PROC LOGISTIC statement
on this transformed data without a STRATA statement; this also performs a conditional logistic
regression and produces essentially the same results.
SAS statements and selected results for the first two approaches are given in the remainder of this example.
Example 73.11: Conditional Logistic Regression for Matched Pairs Data F 5583
Conditional Analysis
Model Information
Data Set WORK.DATA1
Response Variable Outcome
Number of Response Levels 2
Number of Strata 63
Model binary logit
Optimization Technique Newton-Raphson ridge
Response Profile
Ordered Total
Value Outcome Frequency
1 0 63
2 1 63
Probability modeled is Outcome=1.
Strata Summary
Outcome
Response Number of
Pattern 0 1 Strata Frequency
1 1 1 63 126
Note that the score statistic in the “Conditional Exact Tests” table in Output 73.11.2 is identical to the score
statistic in Output 73.11.1 from the conditional analysis. The exact odds ratio confidence interval is much
wider than its conditional analysis counterpart, but the parameter estimates are similar. The exact analysis
confirms the marginal significance of Gall as a predictor variable.
5586 F Chapter 73: The LOGISTIC Procedure
data one;
length Diagnosis $ 9;
input Diagnosis $ Friendships $ Recovered Total @@;
datalines;
Anxious Poor 0 0 Anxious Good 13 21
Depressed Poor 0 8 Depressed Good 15 20
;
Notice that no children in the study are both anxious and have poor friendships and that no children who are
depressed and have poor friendships have recovered. The following statements fit an unconditional logistic
regression model to these data.
The sparseness of the data and the separability of the data set make this a good candidate for an exact logistic
regression. In the following code, the EXACTONLY option suppresses the unconditional logistic regression
results, the EXACT statement requests an exact analysis of the two covariates, the OUTDIST= option outputs
the exact distribution into a SAS data set, the JOINT option computes a joint test for the significance of the
two covariates in the model, and the ESTIMATE option produces parameter estimates for the two covariates.
In Output 73.12.2, the joint probability statistic (0.000081) is the probability of the observed sufficient
statistics (Friendships=28 and Diagnosis=13) in the joint exact distribution (part of which is displayed in
Output 73.12.3). Note that the joint exact distribution has sufficient statistics displayed for both covariates;
the marginal distributions have sufficient statistics displayed for only one covariate. The associated exact
p-value (0.0007) in Output 73.12.2 is the sum of the probabilities of all sufficient statistics in the joint exact
distribution that have probabilities less than or equal to 0.000081. The mid-p-value (0.0007) adjusts the
exact p-value for the discreteness of the exact distribution by subtracting half the probability of the observed
sufficient statistic (Hirji 2006). The score statistic (13.1905) is a weighted distance of the observed sufficient
statistics from the mean sufficient statistics, and the score p-value (0.0008) is the sum of all probabilities in
the Dist data set for sufficient statistics that are no closer to the mean.
5588 F Chapter 73: The LOGISTIC Procedure
For the univariate tests of the Diagnosis variable, PROC LOGISTIC extracts the part of the joint exact
distribution for which Friendships=28; this is displayed in Output 73.12.4. In addition, PROC LOGISTIC
computes the probability and score tests in Output 73.12.2 in the same fashion as the joint tests. Both of the
tests do not reject the null hypothesis that the parameter is zero.
For the univariate tests of the Friendships variable, PROC LOGISTIC extracts the part of the exact distribution
for which Diagnosis=13; this is displayed in Output 73.12.5. Both tests in Output 73.12.2 reject the null
hypothesis that the parameter is zero.
Example 73.12: Exact Conditional Logistic Regression F 5589
The parameter estimates are displayed in Output 73.12.6. Similar to the univariate tests, the parameter
estimates are derived from the marginal exact distributions.
The Diagnosis parameter estimate (–0.5981) is computed by an iterative search for the parameter that
maximizes the univariate conditional probability density function, as described in the section “Inference for a
Single Parameter” on page 5498. The Diagnosis parameter is not significantly different from zero.
Because the observed sufficient statistic for the Friendships parameter is on the edge of its distribution, the
Friendships parameter estimate is the value for which the univariate conditional probability density function
is equal to 0.5. In a similar fashion, the confidence limits for both the Diagnosis and Friendships parameters
are created by finding values to make the two tail probabilities equal to 0.025.
To conclude, you find that the Friendships variable is a significant effect for explaining how mothers perceive
the recovery of their children.
You can also use the exact parameter estimates to compute predicted probabilities for any data. This facility
is not built into PROC LOGISTIC for exact logistic regression, because exact methods can be very expensive
and the computations can fail. But this example is well behaved, so you can use the following statements to
score the data:
5590 F Chapter 73: The LOGISTIC Procedure
%let beta0=-15;
%let beta1=16;
data one;
keep sample X y pry;
do sample=1 to 1000;
do i=1 to 100;
X=rantbl(987987,.4,.6)-1;
xb= &beta0 + X*&beta1;
exb=exp(xb);
pry= exb/(1+exb);
cut= ranuni(393993);
if (pry < cut) then y=1; else y=0;
output;
end;
end;
run;
The following statements perform the bias-corrected and exact logistic regression on each of the 1000
different data sets, output the odds ratio tables by using the ODS OUTPUT statement, and compute various
statistics across the data sets by using the MEANS procedure:
5592 F Chapter 73: The LOGISTIC Procedure
%let beta0=1;
%let beta1=2;
data one;
do sample=1 to 1000;
do pair=1 to 20;
ran=ranuni(939393);
a=3*ranuni(9384984)-1;
pdf0= pdf('NORMAL',a,.4,1);
pdf1= pdf('NORMAL',a,1,1);
pry0= pdf0/(pdf0+pdf1);
pry1= 1-pry0;
xb= log(pry0/pry1);
x= (xb-&beta0*pair/100) / &beta1;
y=0;
output;
x= (-xb-&beta0*pair/100) / &beta1;
y=1;
output;
end;
end;
run;
Unconditional, conditional, exact, and Firth-adjusted analyses are performed on the data sets, and the mean,
minimum, and maximum odds ratios and the mean upper and lower limits for the odds ratios are displayed in
Table 73.17. C AUTION : Due to the exact analyses, this program takes a long time and a lot of resources to
run. You might want to reduce the number of samples generated.
Further examination of the data set all shows that the differences between the square root of the unconditional
odds ratio estimates and the conditional estimates have mean –0.00019 and standard deviation 0.0008,
verifying that the unconditional odds ratio is about the square of the conditional odds ratio. The conditional
and exact conditional odds ratios are also nearly equal, with their differences having mean 3E–7 and standard
deviation 6E–6. The differences between the Firth and the conditional odds ratios can be large (mean 0.69,
Firth Conditional
standard deviation 5.40), but their relative differences, , have mean 0.20 with standard
Conditional
deviation 0.19, so the largest differences occur with the larger estimates.
Example 73.14: Complementary Log-Log Model for Infection Rates F 5595
D1 e
The following statements create the data set sero, which contains the results of a serological survey of malarial
infection. Individuals of nine age groups (Group) were tested. The variable A represents the midpoint of the
age range for each age group. The variable N represents the number of individuals tested in each age group,
and the variable R represents the number of individuals that are seropositive.
data sero;
input Group A N R;
X=log(A);
label X='Log of Midpoint of Age Range';
datalines;
1 1.5 123 8
2 4.0 132 6
3 7.5 182 18
4 12.5 140 14
5 17.5 138 20
6 25.0 161 39
7 35.0 133 19
8 47.0 92 25
9 60.0 74 44
;
For the ith group with the age midpoint Ai , the probability of being seropositive is pi D 1 exp. Ai /. It
follows that
By fitting a binomial model with a complementary log-log link function and by using X=log(A) as an offset
term, you can estimate ˛ D log./ as an intercept parameter. The following statements invoke PROC
LOGISTIC to compute the maximum likelihood estimate of ˛. The LINK=CLOGLOG option is specified to
request the complementary log-log link function. Also specified is the CLPARM=PL option, which requests
the profile-likelihood confidence limits for ˛.
5596 F Chapter 73: The LOGISTIC Procedure
Model Information
Data Set WORK.SERO
Response Variable (Events) R
Response Variable (Trials) N
Offset Variable X Log of Midpoint of Age Range
Model binary cloglog
Optimization Technique Fisher's scoring
Response Profile
Ordered Binary Total
Value Outcome Frequency
1 Event 193
2 Nonevent 982
-2 Log L = 967.1158
Output 73.14.1 shows that the maximum likelihood estimate of ˛ D log./ and its estimated standard error
˛ D 4:6605 and b
are b b˛
D 0:0725, respectively. The infection rate is estimated as
eb̌0 e 4:6605
D1
b e b D1 e D1 e D 0:00942
The 95% confidence interval for
, obtained by back-transforming the 95% confidence interval for ˛, is
(0.0082, 0.0108); that is, there is a 95% chance that, in repeated sampling, the interval of 8 to 11 infections
per thousand individuals contains the true infection rate.
The goodness-of-fit statistics for the constant risk model are statistically significant (p < 0:0001), indicating
that the assumption of constant risk of infection is not correct. You can fit a more extensive model by
allowing a separate risk of infection for each age group. Suppose i is the mean number of infections per
year for the ith age group. The probability of seropositive for the ith group with the age midpoint Ai is
pi D 1 exp. i Ai /, so that
In the following statements, a complementary log-log model is fit containing Group as an explanatory
classification variable with the GLM coding (so that a dummy variable is created for each age group), no
intercept term, and X=log(A) as an offset term. The ODS OUTPUT statement saves the estimates and
their 95% profile-likelihood confidence limits to the ClparmPL data set. Note that log.i / is the regression
parameter associated with GroupD i .
For the first age group (Group=1), the point estimate of log.1 / is –3.1048, which transforms into an infection
rate of 1 exp. exp. 3:1048// D 0:0438. A 95% confidence interval for this infection rate is obtained by
transforming the 95% confidence interval for log.1 /. For the first age group, the lower and upper confidence
limits are 1 exp. exp. 3:8880/ D 0:0203 and 1 exp. exp. 2:4833// D 0:0801, respectively; that is,
there is a 95% chance that, in repeated sampling, the interval of 20 to 80 infections per thousand individuals
contains the true infection rate. The following statements perform this transformation on the estimates and
confidence limits saved in the ClparmPL data set; the resulting estimated infection rates in one year’s time
for each age group are displayed in Table 73.18. Note that the infection rate for the first age group is high
compared to that of the other age groups.
data ClparmPL;
set ClparmPL;
Estimate=round( 1000*( 1-exp(-exp(Estimate)) ) );
LowerCL =round( 1000*( 1-exp(-exp(LowerCL )) ) );
UpperCL =round( 1000*( 1-exp(-exp(UpperCL )) ) );
run;
Example 73.15: Complementary Log-Log Model for Interval-Censored Survival Times F 5599
i t D Pr.Ti D t j Ti t; xi /; t D 1; 2; : : :
Suppose ti is the observed survival time of the ith subject. Suppose ıi D 1 if Ti D ti is an event time and 0
otherwise. The likelihood for the grouped survival data is given by
Y
L D ŒPr.Ti D ti /ıi ŒPr.Ti > ti /1 ıi
i
ti
Y i t ıi Y
i
D .1 ij /
1 i ti
i j D1
ti yij
YY ij
D .1 ij /
1 ij
i j D1
where yij D 1 if the ith subject experienced an event at time Ti D j and 0 otherwise.
5600 F Chapter 73: The LOGISTIC Procedure
Note that the likelihood L for the grouped survival data is the same as the likelihood of a binary response
model with event probabilities ij . If the data are generated by a continuous-time proportional hazards
model, Prentice and Gloeckler (1978) have shown that
where the coefficient vector ˇ is identical to that of the continuous-time proportional hazards model, and
˛j is a constant related to the conditional survival probability in the interval defined by Ti D j at xi D 0.
The grouped data survival model is therefore equivalent to the binary response model with complementary
log-log link function. To fit the grouped survival model by using PROC LOGISTIC, you must treat each
discrete time unit for each subject as a separate observation. For each of these observations, the response is
dichotomous, corresponding to whether or not the subject died in the time unit.
Consider a study of the effect of insecticide on flour beetles. Four different concentrations of an insecticide
were sprayed on separate groups of flour beetles. The following DATA step saves the number of male and
female flour beetles dying in successive intervals in the data set Beetles:
To use PROC LOGISTIC with the grouped survival data, you must expand the data so that each beetle has a
separate record for each day of survival. A beetle that died in the third day (time=3) would contribute three
observations to the analysis, one for each day it was alive at the beginning of the day. A beetle that survives
the 13-day duration of the experiment (time=14) would contribute 13 observations.
The following DATA step creates a new data set named Days containing the beetle-day observations from the
data set Beetles. In addition to the variables sex, conc, and freq, the data set contains an outcome variable y
and a classification variable day. The variable y has a value of 1 if the observation corresponds to the day that
the beetle died, and it has a value of 0 otherwise. An observation for the first day will have a value of 1 for
day; an observation for the second day will have a value of 2 for day, and so on. For instance, Output 73.15.1
shows an observation in the Beetles data set with time=3, and Output 73.15.2 shows the corresponding
beetle-day observations in the data set Days.
data Days;
set Beetles;
do day=1 to time;
if (day < 14) then do;
y= (day=time);
output;
end;
end;
run;
The following statements invoke PROC LOGISTIC to fit a complementary log-log model for binary data with
the response variable Y and the explanatory variables day, sex, and Variableconc. Specifying the EVENT=
option ensures that the event (y=1) probability is modeled. The GLM coding in the CLASS statement creates
an indicator column in the design matrix for each level of day. The coefficients of the indicator effects for
day can be used to estimate the baseline survival function. The NOINT option is specified to prevent any
redundancy in estimating the coefficients of day. The Newton-Raphson algorithm is used for the maximum
likelihood estimation of the parameters.
Results of the model fit are given in Output 73.15.3. Both sex and conc are statistically significant for the
survival of beetles sprayed by the insecticide. Female beetles are more resilient to the chemical than male
beetles, and increased concentration of the insecticide increases its effectiveness.
Output 73.15.3 Parameter Estimates for the Grouped Proportional Hazards Model
The coefficients of parameters for the day variable are the maximum likelihood estimates of ˛1 ; : : : ; ˛13 ,
respectively. The baseline survivor function S0 .t / is estimated by
Y
SO0 .t/ D Pr.T
b > t/ D exp. exp.b ˛ j //
j t
and the survivor function for a given covariate pattern (sex=x1 and conc=x2 ) is estimated by
O 0:5651x1 C3:0918x2 /
S.t/ D ŒSO0 .t /exp.
The following statements compute the survival curves for male and female flour beetles exposed to the
insecticide in concentrations of 0.20 mg/cm2 and 0.80 mg/cm2 :
Example 73.15: Complementary Log-Log Model for Interval-Censored Survival Times F 5603
%modstyle(name=LogiStyle,parent=htmlblue,markers=circlefilled);
ods listing style=LogiStyle;
proc sgplot data=one;
title 'Flour Beetles Sprayed with Insecticide';
xaxis grid integer;
yaxis grid label='Survival Function';
pbspline y=s_m20 x=day /
legendlabel = "Male at 0.20 conc." name="pred1";
pbspline y=s_m80 x=day /
legendlabel = "Male at 0.80 conc." name="pred2";
pbspline y=s_f20 x=day /
legendlabel = "Female at 0.20 conc." name="pred3";
pbspline y=s_f80 x=day /
legendlabel = "Female at 0.80 conc." name="pred4";
discretelegend "pred1" "pred2" "pred3" "pred4" / across=2;
run;
5604 F Chapter 73: The LOGISTIC Procedure
Output 73.15.4 Predicted Survival at Insecticide Concentrations of 0.20 and 0.80 mg/cm2
The probability of survival is displayed on the vertical axis. Notice that most of the insecticide effect occurs
by day 6 for both the high and low concentrations.
Example 73.16: Scoring Data Sets F 5605
data Crops;
length Crop $ 10;
infile datalines truncover;
input Crop $ @@;
do i=1 to 3;
input x1-x4 @@;
if (x1 ^= .) then output;
end;
input;
datalines;
Corn 16 27 31 33 15 23 30 30 16 27 27 26
Corn 18 20 25 23 15 15 31 32 15 32 32 15
Corn 12 15 16 73
Soybeans 20 23 23 25 24 24 25 32 21 25 23 24
Soybeans 27 45 24 12 12 13 15 42 22 32 31 43
Cotton 31 32 33 34 29 24 26 28 34 32 28 45
Cotton 26 25 23 24 53 48 75 26 34 35 25 78
Sugarbeets 22 23 25 42 25 25 24 26 34 25 16 52
Sugarbeets 54 23 21 54 25 43 32 15 26 54 2 54
Clover 12 45 32 54 24 58 25 34 87 54 61 21
Clover 51 31 31 16 96 48 54 62 31 31 11 11
Clover 56 13 13 71 32 13 27 32 36 26 54 32
Clover 53 08 06 54 32 32 62 16
;
In the following statements, you specify a SCORE statement to use the fitted model to score the Crops
data. The data together with the predicted values are saved in the data set Score1. The output from the
EFFECTPLOT statement is discussed at the end of this section.
data Prior;
length Crop $10.;
input Crop _PRIOR_;
datalines;
Clover 11
Corn 7
Cotton 6
Soybeans 6
Sugarbeets 6
;
The data sets Score1, Score2, Score3, Score4, and Score5 are identical. The following statements display
the scoring results in Output 73.16.2:
The following statements use the previously fitted and saved model in the sasuser.CropModel data set to
score the observations in a new data set, Test. The results of scoring the test data are saved in the ScoredTest
data set and displayed in Output 73.16.3.
data Test;
input Crop $ 1-10 x1-x4;
datalines;
Corn 16 27 31 33
Soybeans 21 25 23 24
Cotton 29 24 26 28
Sugarbeets 54 23 21 54
Clover 32 32 62 16
;
The EFFECTPLOT statement that is specified in the first PROC LOGISTIC invocation produces a plot
of the model-predicted probabilities versus X3 while holding the other three covariates at their means
(Output 73.16.4). This plot shows how the value of X3 affects the probabilities of the various crops when
the other prognostic factors are fixed at their means. If you are interested in the effect of X3 when the other
covariates are fixed at a certain level—say, 10—specify the following EFFECTPLOT statement.
Then the odds ratio for Treatment A versus B at Sex=F is computed as exp..lA0 lB0 /ˇ/. Different l vectors
must be similarly constructed when Sex=M because the resulting odds ratio will be different due to the
interaction.
Output 73.17.1 Odds Ratios from the ODDSRATIO Statement
The results from the LSMEANS statement are displayed in Output 73.17.2 through Output 73.17.4.
The LS-means are computed by constructing each of the l coefficient vectors shown in Output 73.17.2,
and then computing l 0 ˇ. The LS-means are not estimates of the event probabilities; they are estimates
of the linear predictors on the logit scale and therefore are estimated log odds. In order to obtain event
probabilities, you need to apply the inverse-link transformation by specifying the ILINK option in the
LSMEANS statement. Notice in Output 73.17.2 that the Sex rows of the coefficient vectors do not select
either Sex=F or Sex=M. Instead, the LS-means are computed at an average of these two levels, so only
one result needs to be reported. For more information about the construction of LS-means, see the section
“Construction of Least Squares Means” on page 3707 in Chapter 47, “The GLM Procedure.”
The Treatment LS-means shown in Output 73.17.3 are all significantly nonzero at the 0.05 level. These
LS-means are predicted population margins of the logits; that is, they estimate the marginal means over a
balanced population, and they are effectively the within-Treatment means appropriately adjusted for the other
effects in the model. The LS-means are not event probabilities; in order to obtain event probabilities, you
need to apply the inverse-link transformation by specifying the ILINK option in the LSMEANS statement.
For more information about LS-means, see the section “LSMEANS Statement” on page 458 in Chapter 19,
“Shared Concepts and Topics.”
Pairwise differences between the Treatment LS-means, requested with the DIFF option, are displayed in
Output 73.17.4. The LS-mean for the level that is displayed in the _Treatment column is subtracted from the
LS-mean for the level in the Treatment column, so the first row displays the LS-mean for Treatment level A
minus the LS-mean for Treatment level B. The difference (–0.6669) is the estimated difference in log odds,
5612 F Chapter 73: The LOGISTIC Procedure
or equivalently the log odds ratio of the two treatments. The Pr > |z| column indicates that the A and B levels
are not significantly different; however, both of these levels are significantly different from level P. If the
inverse-link transformation is specified by the ILINK option, then these differences do not transform back to
differences in probabilities.
There are two odds ratios for Treatment level A versus B in Output 73.17.1; these are constructed at each
level of the interacting covariate Sex. In contrast, there is only one LS-means odds ratio for Treatment
level A versus B in Output 73.17.4. This odds ratio is computed at an average of the interacting effects by
creating the l vectors shown in Output 73.17.2 (the Row1 column corresponds to lA , and the Row2 column
corresponds to lB ) and computing exp.lA0 ˇ lB0 ˇ/.
Because multiple tests are performed, you can protect yourself from falsely significant results by adjusting
your p-values for multiplicity. The ADJUST=BON option performs the very conservative Bonferroni
adjustment, and adds the columns labeled with ‘Adj’ to Output 73.17.4. By comparing the Pr > |z| column to
the Adj P column, you can see that the p-values are adjusted upwards; in this case, there is no change in your
conclusions. The confidence intervals are also adjusted for multiplicity—all adjusted intervals are wider than
the unadjusted intervals, but again your conclusions in this example are unchanged. You can specify other
adjustment methods by using the ADJUST= option.
Output 73.17.4 Differences and Odds Ratios for the Treatment LS-Means
If you want to jointly test whether the active treatments are different from the placebo, you can specify
a custom hypothesis test among LS-means by using the LSMESTIMATE statement. In the following
statements, the LS-means for the two treatments are contrasted against the LS-mean of the placebo, and the
JOINT option performs a joint test that the two treatments are not different from placebo.
Output 73.17.5 displays the results from the LSMESTIMATE statement. The “Least Squares Means Estimates”
table displays the differences of the two active treatments against the placebo, and the results are identical
to the second and third rows of Output 73.17.4. The “Chi-Square Test for Least Squares Means Estimates”
table displays the joint test. In all of these tests, you reject the null hypothesis that the treatment has the same
effect as the placebo.
If you want to work with LS-means but you prefer to compute the Treatment odds ratios within the Sex
levels in the same fashion as the ODDSRATIO statement does, you can specify the SLICE statement. In
the following statements, you specify the same options in the SLICE statement as you do in the LSMEANS
statement, except that you also specify the SLICEBY= option to perform an LS-means analysis partitioned
into sets that are defined by the Sex variable:
Similarly, the results for Sex=M are shown in Output 73.17.8 and Output 73.17.9.
data docvisit;
input sex age agesq income levyplus freepoor freerepa
illness actdays hscore chcond1 chcond2 dvisits;
if ( dvisits > 2) then dvisits = 2;
datalines;
1 0.19 0.0361 0.55 1 0 0 1 4 1 0 0 1
1 0.19 0.0361 0.45 1 0 0 1 2 1 0 0 1
0 0.19 0.0361 0.90 0 0 0 3 0 0 0 0 1
logit.Pr.Y i j x// D ˛i C ˇ 0 x; i D 1; 2
The test of the proportional odds assumption in Output 73.18.1 rejects the null hypothesis that all the slopes
are equal across the two response functions. This test is very anticonservative; that is, it tends to reject the
null hypothesis even when the proportional odds assumption is reasonable.
The proportional odds assumption for ordinal response models can be relaxed by specifying the
UNEQUALSLOPES option in the MODEL statement. A general (nonproportional odds) model has different
slope parameters ˇi D .ˇ1;i ; : : : ; ˇ12;i / for every logit i:
The preceding general model fits 12 2 D 24 slope parameters, and according to the “Linear Hypotheses
Testing Results” table in Output 73.18.2, several variables are unnecessarily allowing nonproportional odds.
You can obtain a more parsimonious model by specifying a subset of the parameters to have nonproportional
odds. The following statements allow the parameters for the variables in the “Linear Hypotheses Testing
Results” table that have p-values less than 0.1 (actdays, agesq, and income) to vary across the response
functions:
You can also specify the following code to let stepwise selection determine which parameters have unequal
slopes:
data a;
label p='Pr>ChiSq';
format p 8.6;
input Test $10. ChiSq1 DF1 ChiSq2 DF2;
ChiSq= ChiSq1-ChiSq2;
DF= DF1-DF2;
p=1-probchi(ChiSq,DF);
keep Test Chisq DF p;
datalines;
Gen vs PO 761.4797 24 734.2971 12
PPO vs PO 752.5512 15 734.2971 12
Gen vs PPO 761.4797 24 752.5512 15
;
Therefore, you reject the proportional odds model in favor of both the general model and the partial
proportional odds model, and the partial proportional odds model fits as well as the general model. The
likelihood ratio test of the general model versus the proportional odds model is very similar to the score test
of the proportional odds assumption in Output 73.18.1 because of the large sample size (Stokes, Davis, and
Koch 2000, p. 249).
N OTE : The proportional odds model has increasing intercepts, which ensures the increasing nature of
the cumulative response functions. However, none of the parameters in the partial proportional odds or
general models are constrained. Because of this, sometimes during the optimization process a predicted
individual probability can be negative; the optimization continues because it might recover from this situation.
Sometimes your final model will predict negative individual probabilities for some of the observations; in
this case a message is displayed, and you should check your data for outliers and possibly redefine your
model. Other times the model fits your data well, but if you try to score new data you can get negative
individual probabilities. This means the model is not appropriate for the data you are trying to score, a
message is displayed, and the estimates are set to missing. These optimization issues do not arise when you
fit an adjacent-category or generalized logit model that has equal and unequal slope effects.
References F 5621
References
Agresti, A. (1990). Categorical Data Analysis. New York: John Wiley & Sons.
Agresti, A. (1992). “A Survey of Exact Inference for Contingency Tables.” Statistical Science 7:131–177.
Agresti, A. (2002). Categorical Data Analysis. 2nd ed. New York: John Wiley & Sons.
Agresti, A. (2010). Analysis of Ordinal Categorical Data. 2nd ed. New York: John Wiley & Sons.
Agresti, A. (2013). Categorical Data Analysis. 3rd ed. Hoboken, NJ: John Wiley & Sons.
Aitchison, J., and Silvey, S. (1957). “The Generalization of Probit Analysis to the Case of Multiple Responses.”
Biometrika 44:131–140.
Albert, A., and Anderson, J. A. (1984). “On the Existence of Maximum Likelihood Estimates in Logistic
Regression Models.” Biometrika 71:1–10.
Allison, P. D. (1982). “Discrete-Time Methods for the Analysis of Event Histories.” In Sociological Methods
and Research, edited by S. Leinhardt, 61–98. San Francisco: Jossey-Bass.
Allison, P. D. (1999). Logistic Regression Using the SAS System: Theory and Application. Cary, NC: SAS
Institute Inc.
Allison, P. D. (2012). Logistic Regression Using SAS: Theory and Application. 2nd ed. Cary, NC: SAS
Institute Inc.
Ashford, J. R. (1959). “An Approach to the Analysis of Data for Semi-quantal Responses in Biology
Response.” Biometrics 15:573–581.
Bartolucci, A. A., and Fraser, M. D. (1977). “Comparative Step-Up and Composite Test for Selecting
Prognostic Indicator Associated with Survival.” Biometrical Journal 19:437–448.
Breslow, N. E., and Day, N. E. (1980). The Analysis of Case-Control Studies. Statistical Methods in Cancer
Research, IARC Scientific Publications, vol. 1, no. 32. Lyon: International Agency for Research on
Cancer.
Brier, G. W. (1950). “Verification of Forecasts Expressed in Terms of Probability.” Monthly Weather Review
78:1–3.
Burnham, K. P., and Anderson, D. R. (1998). Model Selection and Inference: A Practical Information-
Theoretic Approach. New York: Springer-Verlag.
Cameron, A. C., and Trivedi, P. K. (1998). Regression Analysis of Count Data. Cambridge: Cambridge
University Press.
Collett, D. (2003). Modelling Binary Data. 2nd ed. London: Chapman & Hall.
Cook, R. D., and Weisberg, S. (1982). Residuals and Influence in Regression. New York: Chapman & Hall.
Cox, D. R. (1970). The Analysis of Binary Data. New York: Chapman & Hall.
5622 F Chapter 73: The LOGISTIC Procedure
Cox, D. R. (1972). “Regression Models and Life Tables.” Journal of the Royal Statistical Society, Series B
20:187–220. With discussion.
Cox, D. R., and Snell, E. J. (1989). The Analysis of Binary Data. 2nd ed. London: Chapman & Hall.
DeLong, E. R., DeLong, D. M., and Clarke-Pearson, D. L. (1988). “Comparing the Areas under Two or
More Correlated Receiver Operating Characteristic Curves: A Nonparametric Approach.” Biometrics
44:837–845.
Demler, O. V., Pencina, M. J., and D’Agostino, R. B. (2012). “Misuse of DeLong Test to Compare AUCs for
Nested Models.” Statistics in Medicine 31:2577–2587.
Derr, B. (2013). “Ordinal Response Modeling with the LOGISTIC Procedure.” In Proceedings of the
SAS Global Forum 2013 Conference. Cary, NC: SAS Institute Inc. https://support.sas.com/
resources/papers/proceedings13/446-2013.pdf.
Draper, C. C., Voller, A., and Carpenter, R. G. (1972). “The Epidemiologic Interpretation of Serologic Data
in Malaria.” American Journal of Tropical Medicine and Hygiene 21:696–703.
Finney, D. J. (1947). “The Estimation from Individual Records of the Relationship between Dose and Quantal
Response.” Biometrika 34:320–334.
Fleiss, J. L. (1981). Statistical Methods for Rates and Proportions. 2nd ed. New York: John Wiley & Sons.
Forster, J. J., McDonald, J. W., and Smith, P. W. F. (2003). “Markov Chain Monte Carlo Exact Inference for
Binomial and Multinomial Logistic Regression Models.” Statistics and Computing 13:169–177.
Freeman, D. H., Jr. (1987). Applied Categorical Data Analysis. New York: Marcel Dekker.
Furnival, G. M., and Wilson, R. W. (1974). “Regression by Leaps and Bounds.” Technometrics 16:499–511.
Gail, M. H., Lubin, J. H., and Rubinstein, L. V. (1981). “Likelihood Calculations for Matched Case-Control
Studies and Survival Studies with Tied Death Times.” Biometrika 68:703–707.
Hand, D. J., Daly, F., Lunn, A. D., McConway, K. J., and Ostrowski, E. (1994). A Handbook of Small Data
Sets. London: Chapman & Hall.
Hanley, J. A., and McNeil, B. J. (1982). “The Meaning and Use of the Area under a Receiver Operating
Characteristic (ROC) Curve.” Radiology 143:29–36.
Harrell, F. E. (1986). “The LOGIST Procedure.” In SUGI Supplemental Library Guide, Version 5 Edition.
Cary, NC: SAS Institute Inc.
Heinze, G. (1999). The Application of Firth’s Procedure to Cox and Logistic Regression. Technical Report
10, updated January 2001, Department of Medical Computer Sciences, Section of Clinical Biometrics,
University of Vienna.
Heinze, G. (2006). “A Comparative Investigation of Methods for Logistic Regression with Separated or
Nearly Separated Data.” Statistics in Medicine 25:4216–4226.
Heinze, G., and Schemper, M. (2002). “A Solution to the Problem of Separation in Logistic Regression.”
Statistics in Medicine 21:2409–2419.
References F 5623
Hirji, K. F. (1992). “Computing Exact Distributions for Polytomous Response Data.” Journal of the American
Statistical Association 87:487–492.
Hirji, K. F. (2006). Exact Analysis of Discrete Data. Boca Raton, FL: Chapman & Hall/CRC.
Hirji, K. F., Mehta, C. R., and Patel, N. R. (1987). “Computing Distributions for Exact Logistic Regression.”
Journal of the American Statistical Association 82:1110–1117.
Hirji, K. F., Mehta, C. R., and Patel, N. R. (1988). “Exact Inference for Matched Case-Control Studies.”
Biometrics 44:803–814.
Hirji, K. F., Tsiatis, A. A., and Mehta, C. R. (1989). “Median Unbiased Estimation for Binary Data.”
American Statistician 43:7–11.
Hosmer, D. W., Jr., and Lemeshow, S. (2000). Applied Logistic Regression. 2nd ed. New York: John Wiley
& Sons.
Hosmer, D. W., Jr., and Lemeshow, S. (2013). Applied Logistic Regression. 3rd ed. New York: John Wiley &
Sons.
Howard, S. (1972). “Remark on the Paper by Cox (‘Regression Methods and Life Tables’).” Journal of the
Royal Statistical Society, Series B 34:187–220.
Hurvich, C. M., and Tsai, C.-L. (1993). “A Corrected Akaike Information Criterion for Vector Autoregressive
Model Selection.” Journal of Time Series Analysis 14:271–279.
Izrael, D., Battaglia, A. A., Hoaglin, D. C., and Battaglia, M. P. (2002). “Use of the ROC Curve and the
Bootstrap in Comparing Weighted Logistic Regression Models.” In Proceedings of the Twenty-Seventh
Annual SAS Users Group International Conference. Cary, NC: SAS Institute Inc. http://www2.sas.
com/proceedings/sugi27/p248-27.pdf.
Lachin, J. M. (2000). Biostatistical Methods: The Assessment of Relative Risks. New York: John Wiley &
Sons.
Lancaster, H. O. (1961). “Significance Tests in Discrete Distributions.” Journal of the American Statistical
Association 56:223–234.
Lawless, J. F., and Singhal, K. (1978). “Efficient Screening of Nonnormal Regression Models.” Biometrics
34:318–327.
Lee, E. T. (1974). “A Computer Program for Linear Logistic Regression Analysis.” Computer Programs in
Biomedicine 4:80–92.
Littell, R. C., Freund, R. J., and Spector, P. C. (1991). SAS System for Linear Models. 3rd ed. Cary, NC: SAS
Institute Inc.
McCullagh, P., and Nelder, J. A. (1989). Generalized Linear Models. 2nd ed. London: Chapman & Hall.
Mehta, C. R., and Patel, N. R. (1995). “Exact Logistic Regression: Theory and Examples.” Statistics in
Medicine 14:2143–2160.
Mehta, C. R., Patel, N. R., and Senchaudhuri, P. (1992). “Exact Stratified Linear Rank Tests for Ordered
Categorical and Binary Data.” Journal of Computational and Graphical Statistics 1:21–40.
Mehta, C. R., Patel, N. R., and Senchaudhuri, P. (2000). “Efficient Monte Carlo Methods for Conditional
Logistic Regression.” Journal of the American Statistical Association 95:99–108.
Moolgavkar, S. H., Lustbader, E. D., and Venzon, D. J. (1985). “Assessing the Adequacy of the Logistic
Regression Model for Matched Case-Control Studies.” Statistics in Medicine 4:425–435.
Muller, K. E., and Fetterman, B. A. (2002). Regression and ANOVA: An Integrated Approach Using SAS
Software. Cary, NC: SAS Institute Inc.
Murphy, A. H. (1973). “A New Vector Partition of the Probability Score.” Journal of Applied Meteorology
12:595–600.
Naessens, J. M., Offord, K. P., Scott, W. F., and Daood, S. L. (1986). “The MCSTRAT Procedure.” In SUGI
Supplemental Library User’s Guide, Version 5 Edition, 307–328. Cary, NC: SAS Institute Inc.
Nelder, J. A., and Wedderburn, R. W. M. (1972). “Generalized Linear Models.” Journal of the Royal
Statistical Society, Series A 135:370–384.
Peterson, B. L., and Harrell, F. E. (1990). “Partial Proportional Odds Models for Ordinal Response Variables.”
Journal of the Royal Statistical Society, Series B 39:205–217.
Pregibon, D. (1984). “Data Analytic Methods for Matched Case-Control Studies.” Biometrics 40:639–651.
Prentice, R. L., and Gloeckler, L. A. (1978). “Regression Analysis of Grouped Survival Data with Applica-
tions to Breast Cancer Data.” Biometrics 34:57–67.
Press, S. J., and Wilson, S. (1978). “Choosing between Logistic Regression and Discriminant Analysis.”
Journal of the American Statistical Association 73:699–705.
Santner, T. J., and Duffy, D. E. (1986). “A Note on A. Albert and J. A. Anderson’s Conditions for the
Existence of Maximum Likelihood Estimates in Logistic Regression Models.” Biometrika 73:755–758.
SAS Institute Inc. (1995). Logistic Regression Examples Using the SAS System. Cary, NC: SAS Institute Inc.
SAS Institute Inc. (2015). “Sample 54866: Logistic Model Selection Using Area under Curve (AUC) or
R-Square Selection Criteria.” SAS Institute Inc., Cary, NC. http://support.sas.com/kb/54/866.
html.
Stokes, M. E., Davis, C. S., and Koch, G. G. (2000). Categorical Data Analysis Using the SAS System. 2nd
ed. Cary, NC: SAS Institute Inc.
Stokes, M. E., Davis, C. S., and Koch, G. G. (2012). Categorical Data Analysis Using SAS. 3rd ed. Cary,
NC: SAS Institute Inc.
References F 5625
Storer, B. E., and Crowley, J. (1985). “A Diagnostic for Cox Regression and General Conditional Likelihoods.”
Journal of the American Statistical Association 80:139–147.
Venzon, D. J., and Moolgavkar, S. H. (1988). “A Method for Computing Profile-Likelihood-Based Confidence
Intervals.” Journal of the Royal Statistical Society, Series C 37:87–94.
Vollset, S. E., Hirji, K. F., and Afifi, A. A. (1991). “Evaluation of Exact and Asymptotic Interval Estimators
in Logistic Analysis of Matched Case-Control Studies.” Biometrics 47:1311–1325.
Walker, S. H., and Duncan, D. B. (1967). “Estimation of the Probability of an Event as a Function of Several
Independent Variables.” Biometrika 54:167–179.
Williams, D. A. (1982). “Extra-binomial Variation in Logistic Linear Models.” Journal of the Royal Statistical
Society, Series C 31:144–148.
Zamar, D., McNeney, B., , and Graham, J. (2007). “elrm: Software Implementing Exact-Like Inference for
Logistic Regression Models.” Journal of Statistical Software 21:1–18.
Subject Index
Akaike’s information criterion LOGISTIC procedure, 5462
LOGISTIC procedure, 5465 Fisher scoring algorithm
LOGISTIC procedure, 5437, 5439, 5460
backward elimination forward selection
LOGISTIC procedure, 5437, 5464 LOGISTIC procedure, 5437, 5464
Bayes’ theorem frequency variable
LOGISTIC procedure, 5435, 5477 LOGISTIC procedure, 5420
best subset selection
LOGISTIC procedure, 5427, 5437, 5464 gradient
branch-and-bound algorithm LOGISTIC procedure, 5467
LOGISTIC procedure, 5464
hat matrix
classification table LOGISTIC procedure, 5485
LOGISTIC procedure, 5435, 5475, 5476, 5554 Hessian matrix
complete separation LOGISTIC procedure, 5437, 5467
LOGISTIC procedure, 5463 hierarchy
conditional logistic regression LOGISTIC procedure, 5430
LOGISTIC procedure, 5453, 5492 Hosmer-Lemeshow test
confidence intervals LOGISTIC procedure, 5432, 5479
profile likelihood (LOGISTIC), 5435, 5469 test statistic (LOGISTIC), 5480
Wald (LOGISTIC), 5440, 5470
confidence limits infinite parameter estimates
LOGISTIC procedure, 5474 LOGISTIC procedure, 5433, 5463
convergence criterion initial values
profile likelihood (LOGISTIC), 5435 LOGISTIC procedure, 5501
correct classification rate
LOGISTIC procedure, 5476 leverage
LOGISTIC procedure, 5485
descriptive statistics likelihood residuals
LOGISTIC procedure, 5399 LOGISTIC procedure, 5485
deviance link function
LOGISTIC procedure, 5427, 5437, 5478 LOGISTIC procedure, 5379, 5432, 5458, 5468
deviance residuals log likelihood
LOGISTIC procedure, 5485 output data sets (LOGISTIC), 5393
DFBETAS statistics log odds
LOGISTIC procedure, 5486 LOGISTIC procedure, 5471
dispersion parameter LOGISTIC procedure
LOGISTIC procedure, 5478 Akaike’s information criterion, 5465
Bayes’ theorem, 5435
estimability checking best subset selection, 5427
LOGISTIC procedure, 5409 branch-and-bound algorithm, 5464
exact conditional logistic regression, see exact logistic classification table, 5435, 5475, 5476, 5554
regression conditional logistic regression, 5453, 5492
exact logistic regression confidence intervals, 5435, 5440, 5469, 5470
LOGISTIC procedure, 5414, 5495 confidence limits, 5474
convergence criterion, 5416, 5417, 5427
false negative, false positive rate customized odds ratio, 5455
LOGISTIC procedure, 5435, 5476, 5555 descriptive statistics, 5399
Firth’s penalized likelihood deviance, 5427, 5437, 5478
DFBETAS diagnostic, 5486 stratified exact logistic regression, 5453
dispersion parameter, 5478 subpopulation, 5427, 5437, 5478
displayed output, 5508 testing linear hypotheses, 5454, 5483
estimability checking, 5409 Williams’ method, 5478
exact logistic regression, 5414, 5495 logistic regression, see also LOGISTIC procedure
existence of MLEs, 5463
Firth’s penalized likelihood, 5462 maximum likelihood
Fisher scoring algorithm, 5437, 5439, 5460 algorithms (LOGISTIC), 5460
frequency variable, 5420 estimates (LOGISTIC), 5463
goodness of fit, 5427, 5437 missing values
gradient, 5467 LOGISTIC procedure, 5456
hat matrix, 5485 model
Hessian matrix, 5437, 5467 fitting criteria (LOGISTIC), 5465
hierarchy, 5430 hierarchy (LOGISTIC), 5381, 5430
Hosmer-Lemeshow test, 5432, 5479, 5480 model selection
infinite parameter estimates, 5433 LOGISTIC procedure, 5425, 5437, 5464
initial values, 5501 multiple classifications
introductory example, 5382 cutpoints (LOGISTIC), 5435
leverage, 5485
link function, 5379, 5432, 5458, 5468 Newton-Raphson algorithm
log odds, 5471 LOGISTIC procedure, 5437, 5439, 5460, 5461
maximum likelihood algorithms, 5460
odds ratio
missing values, 5456
confidence limits (LOGISTIC), 5427, 5428, 5436
model fitting criteria, 5465
customized (LOGISTIC), 5455
model hierarchy, 5381, 5430
estimation (LOGISTIC), 5470
model selection, 5425, 5437, 5464
with interactions (LOGISTIC), 5441
multiple classifications, 5435
ODS graph names
Newton-Raphson algorithm, 5437, 5439, 5460,
LOGISTIC procedure, 5516
5461
optimization
odds ratio confidence limits, 5427, 5428, 5436
LOGISTIC procedure, 5441
odds ratio estimation, 5470
options summary
odds ratios with interactions, 5441
EFFECT statement, 5411
ODS graph names, 5516
ESTIMATE statement, 5413
ODS table names, 5513
output data sets
optimization, 5441
LOGISTIC procedure, 5499, 5501–5503, 5505
output data sets, 5393, 5499, 5501–5503, 5505
overdispersion
overdispersion, 5436, 5477, 5478
LOGISTIC procedure, 5436, 5477, 5478
parallel lines assumption, 5429, 5439, 5615
overlap of data points
partial proportional odds model, 5429, 5439, 5615
LOGISTIC procedure, 5463
Pearson’s chi-square, 5427, 5437, 5478
predicted probabilities, 5474 parallel lines assumption
prior event probability, 5435, 5477, 5554 LOGISTIC procedure, 5429, 5439, 5615
profile-likelihood convergence criterion, 5435 partial proportional odds model
rank correlation, 5473 LOGISTIC procedure, 5429, 5439, 5615
regression diagnostics, 5484 Pearson residuals
residuals, 5485 LOGISTIC procedure, 5485
response level ordering, 5391, 5424, 5457 Pearson’s chi-square
ROC curve, 5434, 5448, 5480, 5505 LOGISTIC procedure, 5427, 5437, 5478
ROC curve, comparing, 5449, 5481 predicted probabilities
Schwarz criterion, 5465 LOGISTIC procedure, 5474
score statistics, 5467 prior event probability
scoring data sets, 5450, 5487 LOGISTIC procedure, 5435, 5477, 5554
selection methods, 5425, 5437, 5464
singular contrast matrix, 5409 quasi-complete separation
LOGISTIC procedure, 5463
R-square statistic
LOGISTIC procedure, 5436, 5466
rank correlation
LOGISTIC procedure, 5473
receiver operating characteristic, see ROC curve
regression diagnostics
LOGISTIC procedure, 5484
residuals
LOGISTIC procedure, 5485
response level ordering
LOGISTIC procedure, 5391, 5424, 5457
reverse response level ordering
LOGISTIC procedure, 5457
ROC curve
comparing (LOGISTIC), 5449, 5481
LOGISTIC procedure, 5434, 5448, 5480, 5505
Schwarz criterion
LOGISTIC procedure, 5465
score statistics
LOGISTIC procedure, 5467
selection methods, see model selection
singularity criterion
contrast matrix (LOGISTIC), 5409
standardized deviance residuals
LOGISTIC procedure, 5485
standardized Pearson residuals
LOGISTIC procedure, 5485
stepwise selection
LOGISTIC procedure, 5437, 5464, 5518
stratified exact logistic regression
LOGISTIC procedure, 5453
subpopulation
LOGISTIC procedure, 5437
survivor function
estimates (LOGISTIC), 5602
Williams’ method
overdispersion (LOGISTIC), 5478
Syntax Index
ABSFCONV option COV option
MODEL statement (LOGISTIC), 5416, 5427 ROCCONTRAST statement (LOGISTIC), 5449
ADJACENTPAIRS option COVB option
ROCCONTRAST statement (LOGISTIC), 5449 MODEL statement (LOGISTIC), 5428
AGGREGATE= option COVOUT option
MODEL statement (LOGISTIC), 5427 PROC LOGISTIC statement, 5391
ALPHA= option CPREFIX= option
CONTRAST statement (LOGISTIC), 5408 CLASS statement (LOGISTIC), 5404
EXACT statement (LOGISTIC), 5414 CTABLE option
MODEL statement (LOGISTIC), 5427 MODEL statement (LOGISTIC), 5428
OUTPUT statement (LOGISTIC), 5446 CUMULATIVE option
PROC LOGISTIC statement, 5391 SCORE statement (LOGISTIC), 5451
SCORE statement (LOGISTIC), 5450
AT option DATA= option
ODDSRATIO statement (LOGISTIC), 5442 PROC LOGISTIC statement, 5391
SCORE statement (LOGISTIC), 5451
BEST= option DEFAULT= option
MODEL statement (LOGISTIC), 5427 UNITS statement (LOGISTIC), 5455
BINWIDTH= option DESCENDING option
MODEL statement (LOGISTIC), 5427 CLASS statement (LOGISTIC), 5404
BY statement MODEL statement, 5424
LOGISTIC procedure, 5403 PROC LOGISTIC statement, 5391
DETAILS option
C= option MODEL statement (LOGISTIC), 5429
OUTPUT statement (LOGISTIC), 5444 DFBETAS= option
CBAR= option OUTPUT statement (LOGISTIC), 5444
OUTPUT statement (LOGISTIC), 5444 DIFCHISQ= option
CHECKDEPENDENCY= option OUTPUT statement (LOGISTIC), 5445
STRATA statement (LOGISTIC), 5453 DIFDEV= option
CL option OUTPUT statement (LOGISTIC), 5445
MODEL statement (LOGISTIC), 5440 DIFF= option
CL= option ODDSRATIO statement (LOGISTIC), 5442
ODDSRATIO statement (LOGISTIC), 5442
CLASS statement E option
LOGISTIC procedure, 5404 CONTRAST statement (LOGISTIC), 5408
CLM option ROCCONTRAST statement (LOGISTIC), 5449
SCORE statement (LOGISTIC), 5450 EFFECT statement
CLODDS= option LOGISTIC procedure, 5410
MODEL statement (LOGISTIC), 5428 EFFECTPLOT statement
CLPARM= option LOGISTIC procedure, 5412
MODEL statement (LOGISTIC), 5428 EQUALSLOPES option
CLTYPE= option MODEL statement (LOGISTIC), 5429
EXACT statement (LOGISTIC), 5414 ESTIMATE option
CODE statement EXACT statement (LOGISTIC), 5415
LOGISTIC procedure, 5407 ROCCONTRAST statement (LOGISTIC), 5449
CONTRAST statement ESTIMATE statement
LOGISTIC procedure, 5407 LOGISTIC procedure, 5413
CORRB option ESTIMATE= option
MODEL statement (LOGISTIC), 5428 CONTRAST statement (LOGISTIC), 5408
EVENT= option OUTPUT statement (LOGISTIC), 5444
MODEL statement, 5424
EXACT statement LACKFIT option
LOGISTIC procedure, 5414 MODEL statement (LOGISTIC), 5432
EXACTONLY option LINK= option
PROC LOGISTIC statement, 5391 MODEL statement (LOGISTIC), 5432
EXACTOPTIONS option ROC statement (LOGISTIC), 5448
PROC LOGISTIC statement, 5392 LOGISTIC procedure, 5389
EXACTOPTIONS statement ID statement, 5420
LOGISTIC procedure, 5416 NLOPTIONS statement, 5441
EXPEST option syntax, 5389
MODEL statement (LOGISTIC), 5429 LOGISTIC procedure, BY statement, 5403
LOGISTIC procedure, CLASS statement, 5404
FAST option CPREFIX= option, 5404
MODEL statement (LOGISTIC), 5429 DESCENDING option, 5404
FCONV= option LPREFIX= option, 5404
MODEL statement (LOGISTIC), 5417, 5430 MISSING option, 5404
FIRTH option ORDER= option, 5404
MODEL statement (LOGISTIC), 5430 PARAM= option, 5405
FITSTAT option REF= option, 5406
SCORE statement (LOGISTIC), 5451 TRUNCATE option, 5406
FREQ statement LOGISTIC procedure, CODE statement, 5407
LOGISTIC procedure, 5420 LOGISTIC procedure, CONTRAST statement, 5407
ALPHA= option, 5408
GCONV= option E option, 5408
MODEL statement (LOGISTIC), 5430 ESTIMATE= option, 5408
SINGULAR= option, 5409
H= option LOGISTIC procedure, EFFECT statement, 5410
OUTPUT statement (LOGISTIC), 5445 LOGISTIC procedure, EFFECTPLOT statement, 5412
HIERARCHY= option LOGISTIC procedure, ESTIMATE statement, 5413
MODEL statement (LOGISTIC), 5430 LOGISTIC procedure, EXACT statement, 5414
ALPHA= option, 5414
ID statement
CLTYPE= option, 5414
LOGISTIC procedure, 5420
ESTIMATE option, 5415
INCLUDE= option
JOINT option, 5415
MODEL statement (LOGISTIC), 5431
JOINTONLY option, 5415
INEST= option
MIDPFACTOR= option, 5415
PROC LOGISTIC statement, 5392
ONESIDED option, 5415
INFLUENCE option
OUTDIST= option, 5415
MODEL statement (LOGISTIC), 5432
LOGISTIC procedure, EXACTOPTIONS statement,
INFO option
5416
STRATA statement (LOGISTIC), 5454
LOGISTIC procedure, FREQ statement, 5420
INMODEL= option
LOGISTIC procedure, LSMEANS statement, 5420
PROC LOGISTIC statement, 5392
LOGISTIC procedure, LSMESTIMATE statement,
IPLOTS option
5422
MODEL statement (LOGISTIC), 5432
LOGISTIC procedure, MODEL statement, 5423
ITPRINT option
ABSFCONV option, 5416, 5427
MODEL statement (LOGISTIC), 5432
AGGREGATE= option, 5427
JOINT option ALPHA= option, 5427
EXACT statement (LOGISTIC), 5415 BEST= option, 5427
JOINTONLY option BINWIDTH= option, 5427
EXACT statement (LOGISTIC), 5415 CL option, 5440
CLODDS= option, 5428
keyword= option CLPARM= option, 5428
CORRB option, 5428 STB option, 5438
COVB option, 5428 STOP= option, 5439
CTABLE option, 5428 STOPRES option, 5439
DESCENDING option, 5424 TECHNIQUE= option, 5439
DETAILS option, 5429 UNEQUALSLOPES option, 5439
EQUALSLOPES option, 5429 WALDCL option, 5440
EVENT= option, 5424 WALDRL option, 5436
EXPEST option, 5429 XCONV= option, 5419, 5440
FAST option, 5429 LOGISTIC procedure, ODDSRATIO statement, 5441
FCONV= option, 5417, 5430 AT option, 5442
FIRTH option, 5430 CL= option, 5442
GCONV= option, 5430 DIFF= option, 5442
HIERARCHY= option, 5430 PLCONV= option, 5442
INCLUDE= option, 5431 PLMAXITER= option, 5442
INFLUENCE option, 5432 PLSINGULAR= option, 5442
IPLOTS option, 5432 LOGISTIC procedure, OUTPUT statement, 5443
ITPRINT option, 5432 ALPHA= option, 5446
LACKFIT option, 5432 C= option, 5444
LINK= option, 5432 CBAR= option, 5444
MAXFUNCTION= option, 5433 DFBETAS= option, 5444
MAXITER= option, 5433 DIFCHISQ= option, 5445
MAXSTEP= option, 5433 DIFDEV= option, 5445
NOCHECK option, 5433 H= option, 5445
NODESIGNPRINT= option, 5433 keyword= option, 5444
NODUMMYPRINT= option, 5433 LOWER= option, 5445
NOFIT option, 5434 OUT= option, 5444
NOINT option, 5434 PREDICTED= option, 5445
NOLOGSCALE option, 5419, 5434 PREDPROBS= option, 5445
NOODDSRATIO option, 5434 RESCHI= option, 5446
NOOR option, 5434 RESDEV= option, 5446
OFFSET= option, 5434 RESLIK= option, 5446
ORDER= option, 5424 STDRESCHI= option, 5446
ORPVALUE option, 5434 STDRESDEV= option, 5446
OUTROC= option, 5434 STDXBETA = option, 5446
PARMLABEL option, 5434 UPPER= option, 5446
PCORR option, 5434 XBETA= option, 5446
PEVENT= option, 5435 LOGISTIC procedure, PROC LOGISTIC statement,
PLCL option, 5435 5390
PLCONV= option, 5435 ALPHA= option, 5391
PLRL option, 5435 COVOUT option, 5391
PPROB= option, 5435 DATA= option, 5391
REFERENCE= option, 5424 DESCENDING option, 5391
RIDGING= option, 5435 EXACTONLY option, 5391
RISKLIMITS option, 5436 EXACTOPTIONS option, 5392
ROCCI option, 5436 INEST= option, 5392
ROCEPS= option, 5436 INMODEL= option, 5392
RSQUARE option, 5436 MULTIPASS option, 5392
SCALE= option, 5436 NAMELEN= option, 5393
SELECTION= option, 5437 NOCOV option, 5393
SEQUENTIAL option, 5437 NOPRINT option, 5393
SINGULAR= option, 5437 ORDER= option, 5393
SLENTRY= option, 5437 OUTDESIGN= option, 5393
SLSTAY= option, 5438 OUTDESIGNONLY option, 5393
START= option, 5438 OUTEST= option, 5393
OUTMODEL= option, 5393 MODEL statement (LOGISTIC), 5433
PLOTS option, 5394 MIDPFACTOR= option
ROCOPTIONS option, 5397 EXACT statement (LOGISTIC), 5415
SIMPLE option, 5399 MISSING option
TRUNCATE option, 5399 CLASS statement (LOGISTIC), 5404
LOGISTIC procedure, ROC statement, 5448 STRATA statement (LOGISTIC), 5453
LINK= option, 5448 MODEL statement
NOINT option, 5448 LOGISTIC procedure, 5423
NOOFFSET option, 5448 MULTIPASS option
LOGISTIC procedure, ROCCONTRAST statement, PROC LOGISTIC statement, 5392
5449
ADJACENTPAIRS option, 5449 NAMELEN= option
COV option, 5449 PROC LOGISTIC statement, 5393
E option, 5449 NLOPTIONS statement
ESTIMATE option, 5449 LOGISTIC procedure, 5441
REFERENCE option, 5449 NOCHECK option
LOGISTIC procedure, SCORE statement, 5450 MODEL statement (LOGISTIC), 5433
ALPHA= option, 5450 NOCOV option
CLM option, 5450 PROC LOGISTIC statement, 5393
CUMULATIVE option, 5451 NODESIGNPRINT= option
DATA= option, 5451 MODEL statement (LOGISTIC), 5433
FITSTAT option, 5451 NODUMMYPRINT= option
OUT= option, 5451 MODEL statement (LOGISTIC), 5433
OUTROC= option, 5451 NOFIT option
PRIOR= option, 5452 MODEL statement (LOGISTIC), 5434
PRIOREVENT= option, 5452 NOINT option
ROCEPS= option, 5452 MODEL statement (LOGISTIC), 5434
LOGISTIC procedure, SLICE statement, 5452 ROC statement (LOGISTIC), 5448
LOGISTIC procedure, STORE statement, 5452 NOLOGSCALE option
LOGISTIC procedure, STRATA statement, 5453 MODEL statement (LOGISTIC), 5419, 5434
CHECKDEPENDENCY= option, 5453 NOODDSRATIO option
INFO option, 5454 MODEL statement (LOGISTIC), 5434
MISSING option, 5453 NOOFFSET option
NOSUMMARY option, 5454 ROC statement (LOGISTIC), 5448
LOGISTIC procedure, TEST statement, 5454 NOOR option
PRINT option, 5455 MODEL statement (LOGISTIC), 5434
LOGISTIC procedure, UNITS statement, 5455 NOPRINT option
DEFAULT= option, 5455 PROC LOGISTIC statement, 5393
LOGISTIC procedure, WEIGHT statement, 5456 NORMALIZE option
NORMALIZE option, 5456 WEIGHT statement (LOGISTIC), 5456
LOWER= option NOSUMMARY option
OUTPUT statement (LOGISTIC), 5445 STRATA statement (LOGISTIC), 5454
LPREFIX= option
CLASS statement (LOGISTIC), 5404 ODDSRATIO statement
LSMEANS statement LOGISTIC procedure, 5441
LOGISTIC procedure, 5420 OFFSET= option
LSMESTIMATE statement MODEL statement (LOGISTIC), 5434
LOGISTIC procedure, 5422 ONESIDED option
EXACT statement (LOGISTIC), 5415
MAXFUNCTION= option ORDER= option
MODEL statement (LOGISTIC), 5433 CLASS statement (LOGISTIC), 5404
MAXITER= option MODEL statement, 5424
MODEL statement (LOGISTIC), 5433 PROC LOGISTIC statement, 5393
MAXSTEP= option ORPVALUE option
MODEL statement (LOGISTIC), 5434 PROC LOGISTIC statement, see LOGISTIC
OUT= option procedure
OUTPUT statement (LOGISTIC), 5444
SCORE statement (LOGISTIC), 5451 REF= option
OUTDESIGN= option CLASS statement (LOGISTIC), 5406
PROC LOGISTIC statement, 5393 REFERENCE option
OUTDESIGNONLY option ROCCONTRAST statement (LOGISTIC), 5449
PROC LOGISTIC statement, 5393 REFERENCE= option
OUTDIST= option MODEL statement, 5424
EXACT statement (LOGISTIC), 5415 RESCHI= option
OUTEST= option OUTPUT statement (LOGISTIC), 5446
PROC LOGISTIC statement, 5393 RESDEV= option
OUTMODEL= option OUTPUT statement (LOGISTIC), 5446
PROC LOGISTIC statement, 5393 RESLIK= option
OUTPUT statement OUTPUT statement (LOGISTIC), 5446
LOGISTIC procedure, 5443 RIDGING= option
OUTROC= option MODEL statement (LOGISTIC), 5435
MODEL statement (LOGISTIC), 5434 RISKLIMITS option
SCORE statement (LOGISTIC), 5451 MODEL statement (LOGISTIC), 5436
ROC statement
PARAM= option LOGISTIC procedure, 5448
CLASS statement (LOGISTIC), 5405 ROCCI option
PARMLABEL option MODEL statement (LOGISTIC), 5436
MODEL statement (LOGISTIC), 5434 ROCCONTRAST statement
PCORR option LOGISTIC procedure, 5449
MODEL statement (LOGISTIC), 5434 ROCEPS= option
PEVENT= option MODEL statement (LOGISTIC), 5436
MODEL statement (LOGISTIC), 5435 SCORE statement (LOGISTIC), 5452
PLCL option ROCOPTIONS option
MODEL statement (LOGISTIC), 5435 PROC LOGISTIC statement, 5397
PLCONV= option RSQUARE option
MODEL statement (LOGISTIC), 5435 MODEL statement (LOGISTIC), 5436
ODDSRATIO statement (LOGISTIC), 5442
PLMAXITER= option SCALE= option
ODDSRATIO statement (LOGISTIC), 5442 MODEL statement (LOGISTIC), 5436
PLOTS option SCORE statement
PROC LOGISTIC statement, 5394 LOGISTIC procedure, 5450
PLRL option SELECTION= option
MODEL statement (LOGISTIC), 5435 MODEL statement (LOGISTIC), 5437
PLSINGULAR= option SEQUENTIAL option
ODDSRATIO statement (LOGISTIC), 5442 MODEL statement (LOGISTIC), 5437
PPROB= option SIMPLE option
MODEL statement (LOGISTIC), 5435 PROC LOGISTIC statement, 5399
PREDICTED= option SINGULAR= option
OUTPUT statement (LOGISTIC), 5445 CONTRAST statement (LOGISTIC), 5409
PREDPROBS= option MODEL statement (LOGISTIC), 5437
OUTPUT statement (LOGISTIC), 5445 SLENTRY= option
PRINT option MODEL statement (LOGISTIC), 5437
TEST statement (LOGISTIC), 5455 SLICE statement
PRIOR= option LOGISTIC procedure, 5452
SCORE statement (LOGISTIC), 5452 SLSTAY= option
PRIOREVENT= option MODEL statement (LOGISTIC), 5438
SCORE statement (LOGISTIC), 5452 START= option
MODEL statement (LOGISTIC), 5438
STB option
MODEL statement (LOGISTIC), 5438
STDRESCHI= option
OUTPUT statement (LOGISTIC), 5446
STDRESDEV= option
OUTPUT statement (LOGISTIC), 5446
STDXBETA= option
OUTPUT statement (LOGISTIC), 5446
STOP= option
MODEL statement (LOGISTIC), 5439
STOPRES option
MODEL statement (LOGISTIC), 5439
STORE statement
LOGISTIC procedure, 5452
STRATA statement
LOGISTIC procedure, 5453
TECHNIQUE= option
MODEL statement (LOGISTIC), 5439
TEST statement
LOGISTIC procedure, 5454
TRUNCATE option
CLASS statement (LOGISTIC), 5406
PROC LOGISTIC statement, 5399
UNEQUALSLOPES option
MODEL statement (LOGISTIC), 5439
UNITS statement, LOGISTIC procedure, 5455
UPPER= option
OUTPUT statement (LOGISTIC), 5446
WALDCL option
MODEL statement (LOGISTIC), 5440
WALDRL option
MODEL statement (LOGISTIC), 5436
WEIGHT statement
LOGISTIC procedure, 5456
XBETA= option
OUTPUT statement (LOGISTIC), 5446
XCONV= option
MODEL statement (LOGISTIC), 5419, 5440