SAS/STAT 922 User's Guide: The GENMOD Procedure (Book Excerpt)

SAS/STAT 922 User's Guide: The GENMOD Procedure (Book Excerpt)
®
SAS/STAT 9.22 User’s Guide
The GENMOD Procedure
(Book Excerpt)
SAS® Documentation
This document is an individual chapter from SAS/STAT® 9.22 User’s Guide.
The correct bibliographic citation for the complete manual is as follows: SAS Institute Inc. 2010. SAS/STAT® 9.22 User’s
Guide. Cary, NC: SAS Institute Inc.
Copyright © 2010, SAS Institute Inc., Cary, NC, USA
All rights reserved. Produced in the United States of America.
For a Web download or e-book: Your use of this publication shall be governed by the terms established by the vendor at
the time you acquire this publication.
U.S. Government Restricted Rights Notice: Use, duplication, or disclosure of this software and related documentation
by the U.S. government is subject to the Agreement with SAS Institute and the restrictions set forth in FAR 52.227-19,
Commercial Computer Software-Restricted Rights (June 1987).
SAS Institute Inc., SAS Campus Drive, Cary, North Carolina 27513.
1st electronic book, May 2010
SAS® Publishing provides a complete selection of books and electronic products to help customers use SAS software to
its fullest potential. For more information about our e-books, e-learning products, CDs, and hard-copy books, visit the
SAS Publishing Web site at support.sas.com/publishing or call 1-800-727-3228.
SAS® and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute
Inc. in the USA and other countries. ® indicates USA registration.
Other brand and product names are registered trademarks or trademarks of their respective companies.
Chapter 37
The GENMOD Procedure
Contents
Overview: GENMOD Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . .
2429
What Is a Generalized Linear Model? . . . . . . . . . . . . . . . . . . . . .
2430
Examples of Generalized Linear Models . . . . . . . . . . . . . . . . . . . . 2431
The GENMOD Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . .
2432
Getting Started: GENMOD Procedure . . . . . . . . . . . . . . . . . . . . . . . .
2435
Poisson Regression
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2435
Bayesian Analysis of a Linear Regression Model . . . . . . . . . . . . . . .
2440
Generalized Estimating Equations . . . . . . . . . . . . . . . . . . . . . . .
2453
Syntax: GENMOD Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2456
PROC GENMOD Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 2457
ASSESS Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2462
BAYES Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2463
BY Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2473
CLASS Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2474
CONTRAST Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2477
DEVIANCE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2479
EFFECTPLOT Statement . . . . . . . . . . . . . . . . . . . . . . . . . . .
2480
ESTIMATE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2481
EXACT Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
EXACTOPTIONS Statement . . . . . . . . . . . . . . . . . . . . . . . . .
2482
2484
FREQ Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2487
FWDLINK Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2487
INVLINK Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2488
LSMEANS Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2488
LSMESTIMATE Statement . . . . . . . . . . . . . . . . . . . . . . . . . .
2489
MODEL Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2491
OUTPUT Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2499
Programming Statements . . . . . . . . . . . . . . . . . . . . . . . . . . .
2502
REPEATED Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2503
SLICE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2507
STORE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2507
STRATA Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2507
VARIANCE Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2509
WEIGHT Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2509
2428 F Chapter 37: The GENMOD Procedure
ZEROMODEL Statement . . . . . . . . . . . . . . . . . . . . . . . . . . .
2510
Details: GENMOD Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2510
Generalized Linear Models Theory . . . . . . . . . . . . . . . . . . . . . .
2510
Specification of Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2522
Parameterization Used in PROC GENMOD . . . . . . . . . . . . . . . . . .
2523
Type 1 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2523
Type 3 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2524
Confidence Intervals for Parameters . . . . . . . . . . . . . . . . . . . . . .
2525
F Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2526
Lagrange Multiplier Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 2527
Predicted Values of the Mean . . . . . . . . . . . . . . . . . . . . . . . . . . 2527
Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2528
Multinomial Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2529
Zero-Inflated Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Generalized Estimating Equations . . . . . . . . . . . . . . . . . . . . . . .
2530
2532
Assessment of Models Based on Aggregates of Residuals . . . . . . . . . . . 2541
Case Deletion Diagnostic Statistics . . . . . . . . . . . . . . . . . . . . . .
2544
Bayesian Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2548
Exact Logistic and Poisson Regression . . . . . . . . . . . . . . . . . . . .
2553
Missing Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2556
Displayed Output for Classical Analysis . . . . . . . . . . . . . . . . . . .
2556
Displayed Output for Bayesian Analysis . . . . . . . . . . . . . . . . . . .
2564
Displayed Output for Exact Analysis . . . . . . . . . . . . . . . . . . . . . . 2567
ODS Table Names . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2568
ODS Graphics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2572
Examples: GENMOD Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . .
2574
Example 37.1: Logistic Regression . . . . . . . . . . . . . . . . . . . . . .
2574
Example 37.2: Normal Regression, Log Link
. . . . . . . . . . . . . . . .
2576
Example 37.3: Gamma Distribution Applied to Life Data . . . . . . . . . .
2579
Example 37.4: Ordinal Model for Multinomial Data . . . . . . . . . . . . .
2582
Example 37.5: GEE for Binary Data with Logit Link Function . . . . . . .
2586
Example 37.6: Log Odds Ratios and the ALR Algorithm . . . . . . . . . .
2589
Example 37.7: Log-Linear Model for Count Data . . . . . . . . . . . . . .
2592
Example 37.8: Model Assessment of Multiple Regression Using Aggregates
of Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2597
Example 37.9: Assessment of a Marginal Model for Dependent Data . . . .
2604
Example 37.10: Bayesian Analysis of a Poisson Regression Model . . . . .
2608
Example 37.11: Exact Poisson Regression . . . . . . . . . . . . . . . . . .
2622
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2626
Overview: GENMOD Procedure F 2429
Overview: GENMOD Procedure
The GENMOD procedure fits generalized linear models, as defined by Nelder and Wedderburn
(1972). The class of generalized linear models is an extension of traditional linear models that allows
the mean of a population to depend on a linear predictor through a nonlinear link function and allows
the response probability distribution to be any member of an exponential family of distributions.
Many widely used statistical models are generalized linear models. These include classical linear
models with normal errors, logistic and probit models for binary data, and log-linear models for
multinomial data. Many other useful statistical models can be formulated as generalized linear
models by the selection of an appropriate link function and response probability distribution.
See McCullagh and Nelder (1989) for a discussion of statistical modeling using generalized linear
models. The books by Aitkin et al. (1989) and Dobson (1990) are also excellent references with
many examples of applications of generalized linear models. Firth (1991) provides an overview
of generalized linear models. Myers, Montgomery, and Vining (2002) provide applications of
generalized linear models in the engineering and physical sciences. Collett (2003) and Hilbe (2009)
provide comprehensive accounts of generalized linear models when the responses are binary.
The analysis of correlated data arising from repeated measurements when the measurements are
assumed to be multivariate normal has been studied extensively. However, the normality assumption
might not always be reasonable; for example, different methodology must be used in the data analysis
when the responses are discrete and correlated. Generalized estimating equations (GEEs) provide a
practical method with reasonable statistical efficiency to analyze such data.
Liang and Zeger (1986) introduced GEEs as a method of dealing with correlated data when, except
for the correlation among responses, the data can be modeled as a generalized linear model. For
example, correlated binary and count data in many cases can be modeled in this way.
The GENMOD procedure can fit models to correlated responses by the GEE method. You can use
PROC GENMOD to fit models with most of the correlation structures from Liang and Zeger (1986)
by using GEEs. See Hardin and Hilbe (2003), Diggle, Liang, and Zeger (1994), and Lipsitz et al.
(1994) for more details on GEEs.
Bayesian analysis of generalized linear models can be requested by using the BAYES statement in the
GENMOD procedure. In Bayesian analysis, the model parameters are treated as random variables,
and inference about parameters is based on the posterior distribution of the parameters, given the
data. The posterior distribution is obtained using Bayes’ theorem as the likelihood function of the
data weighted with a prior distribution. The prior distribution enables you to incorporate knowledge
or experience of the likely range of values of the parameters of interest into the analysis. If you have
no prior knowledge of the parameter values, you can use a noninformative prior distribution, and
the results of the Bayesian analysis will be very similar to a classical analysis based on maximum
likelihood. A closed form of the posterior distribution is often not feasible, and a Markov chain Monte
Carlo method by Gibbs sampling is used to simulate samples from the posterior distribution. See
Chapter 7, “Introduction to Bayesian Analysis Procedures,” for an introduction to the basic concepts
of Bayesian statistics. Also see the section “Bayesian Analysis: Advantages and Disadvantages” on
page 147 for a discussion of the advantages and disadvantages of Bayesian analysis. See Ibrahim,
Chen, and Sinha (2001) for a detailed description of Bayesian analysis.
2430 F Chapter 37: The GENMOD Procedure
In a Bayesian analysis, a Gibbs chain of samples from the posterior distribution is generated for
the model parameters. Summary statistics (mean, standard deviation, quartiles, HPD and credible
intervals, correlation matrix) and convergence diagnostics (autocorrelations; Gelman-Rubin, Geweke,
Raftery-Lewis, and Heidelberger and Welch tests; the effective sample size; and Monte Carlo
standard errors) are computed for each parameter, as well as the correlation matrix and the covariance
matrix of the posterior sample. Trace plots, posterior density plots, and autocorrelation function plots
that are created using ODS Graphics are also provided for each parameter.
The GENMOD procedure enables you to perform exact logistic regression, also called exact conditional binary logistic regression, and exact Poisson regression, also called exact conditional Poisson
regression, by specifying one or more EXACT statements. You can test individual parameters or
conduct a joint test for several parameters. The procedure computes two exact tests: the exact
conditional score test and the exact conditional probability test. You can request exact estimation
of specific parameters and corresponding odds ratios where appropriate. Point estimates, standard
errors, and confidence intervals are provided.
The GENMOD procedure now uses ODS Graphics to create graphs as part of its output. For general
information about ODS Graphics, see Chapter 21, “Statistical Graphics Using ODS.”
What Is a Generalized Linear Model?
A traditional linear model is of the form
yi D x0i ˇ C "i
where yi is the response variable for the i th observation. The quantity xi is a column vector of
covariates, or explanatory variables, for observation i that is known from the experimental setting
and is considered to be fixed, or nonrandom. The vector of unknown coefficients ˇ is estimated by a
least squares fit to the data y. The "i are assumed to be independent, normal random variables with
zero mean and constant variance. The expected value of yi , denoted by i , is
i D x0i ˇ
While traditional linear models are used extensively in statistical data analysis, there are types of
problems such as the following for which they are not appropriate.
It might not be reasonable to assume that data are normally distributed. For example, the
normal distribution (which is continuous) might not be adequate for modeling counts or
measured proportions that are considered to be discrete.
If the mean of the data is naturally restricted to a range of values, the traditional linear model
might not be appropriate, since the linear predictor x0i ˇ can take on any value. For example,
the mean of a measured proportion is between 0 and 1, but the linear predictor of the mean in
a traditional linear model is not restricted to this range.
It might not be realistic to assume that the variance of the data is constant for all observations.
For example, it is not unusual to observe data where the variance increases with the mean of
the data.
Examples of Generalized Linear Models F 2431
A generalized linear model extends the traditional linear model and is therefore applicable to a wider
range of data analysis problems. A generalized linear model consists of the following components:
The linear component is defined just as it is for traditional linear models:
i D x0i ˇ
A monotonic differentiable link function g describes how the expected value of yi is related to
the linear predictor i :
g.i / D x0i ˇ
The response variables yi are independent for i = 1, 2,. . . and have a probability distribution
from an exponential family. This implies that the variance of the response depends on the
mean through a variance function V :
var.yi / D
V .i /
wi
where is a constant and wi is a known weight for each observation. The dispersion parameter
is either known (for example, for the binomial or Poisson distribution, D 1) or must be
estimated.
See the section “Response Probability Distributions” on page 2510 for the form of a probability
distribution from the exponential family of distributions.
As in the case of traditional linear models, fitted generalized linear models can be summarized through
statistics such as parameter estimates, their standard errors, and goodness-of-fit statistics. You can
also make statistical inference about the parameters by using confidence intervals and hypothesis
tests. However, specific inference procedures are usually based on asymptotic considerations, since
exact distribution theory is not available or is not practical for all generalized linear models.
Examples of Generalized Linear Models
You construct a generalized linear model by deciding on response and explanatory variables for your
data and choosing an appropriate link function and response probability distribution. Some examples
of generalized linear models follow. Explanatory variables can be any combination of continuous
variables, classification variables, and interactions.
Traditional Linear Model
response variable: a continuous variable
distribution: normal
link function: identity g./ D 2432 F Chapter 37: The GENMOD Procedure
Logistic Regression
response variable: a proportion
distribution: binomial
link function: logit g./ D log
1
Poisson Regression in Log-Linear Model
response variable: a count
distribution: Poisson
link function: log g./ D log./
Gamma Model with Log Link
response variable: a positive, continuous variable
distribution: gamma
link function: log g./ D log./
The GENMOD Procedure
The GENMOD procedure fits a generalized linear model to the data by maximum likelihood
estimation of the parameter vector ˇ. There is, in general, no closed form solution for the maximum
likelihood estimates of the parameters. The GENMOD procedure estimates the parameters of the
model numerically through an iterative fitting process. The dispersion parameter is also estimated
by maximum likelihood or, optionally, by the residual deviance or by Pearson’s chi-square divided by
the degrees of freedom. Covariances, standard errors, and p-values are computed for the estimated
parameters based on the asymptotic normality of maximum likelihood estimators.
A number of popular link functions and probability distributions are available in the GENMOD
procedure. The built-in link functions are as follows:
identity: g./ D logit: g./ D log.=.1
//
probit: g./ D ˆ 1 ./, where ˆ is the standard normal cumulative distribution function
if ¤ 0
power: g./ D
log./ if D 0
log: g./ D log./
The GENMOD Procedure F 2433
complementary log-log: g./ D log. log.1
//
The available distributions and associated variance functions are as follows:
normal: V ./ D 1
binomial (proportion): V ./ D .1
/
Poisson: V ./ D gamma: V ./ D 2
inverse Gaussian: V ./ D 3
negative binomial: V ./ D C k2
geometric: V ./ D C 2
multinomial
zero-inflated Poisson
zero-inflated negative binomial
The negative binomial and zero-inflated negative binomial are distributions with an additional
parameter k in the variance function. PROC GENMOD estimates k by maximum likelihood, or you
can optionally set it to a constant value. See McCullagh and Nelder (1989), Hilbe (1994), Hilbe
(2007), Long (1997), Cameron and Trivedi (1998), or Lawless (1987) for discussions of the negative
binomial distribution.
The multinomial distribution is sometimes used to model a response that can take values from a
number of categories. The binomial is a special case of the multinomial with two categories. See
the section “Multinomial Models” on page 2529 and McCullagh and Nelder (1989, Chapter 5) for a
description of the multinomial distribution.
The zero-inflated Poisson and zero-inflated negative binomial are included in PROC GENMOD
even though they are not generalized linear models. They are useful extensions of generalized linear
models. See the section “Zero-Inflated Models” on page 2530 for information about the zero-inflated
distributions.
In addition, you can easily define your own link functions or distributions through DATA step
programming statements used within the procedure.
An important aspect of generalized linear modeling is the selection of explanatory variables in the
model. Changes in goodness-of-fit statistics are often used to evaluate the contribution of subsets of
explanatory variables to a particular model. The deviance, defined to be twice the difference between
the maximum attainable log likelihood and the log likelihood of the model under consideration, is
often used as a measure of goodness of fit. The maximum attainable log likelihood is achieved with
a model that has a parameter for every observation. See the section “Goodness of Fit” on page 2517
for formulas for the deviance.
One strategy for variable selection is to fit a sequence of models, beginning with a simple model with
only an intercept term, and then to include one additional explanatory variable in each successive
2434 F Chapter 37: The GENMOD Procedure
model. You can measure the importance of the additional explanatory variable by the difference in
deviances or fitted log likelihoods between successive models. Asymptotic tests computed by the
GENMOD procedure enable you to assess the statistical significance of the additional term.
The GENMOD procedure enables you to fit a sequence of models, up through a maximum number of
terms specified in a MODEL statement. A table summarizes twice the difference in log likelihoods
between each successive pair of models. This is called a Type 1 analysis in the GENMOD procedure,
because it is analogous to Type I (sequential) sums of squares in the GLM procedure. As with the
PROC GLM Type I sums of squares, the results from this process depend on the order in which the
model terms are fit.
The GENMOD procedure also generates a Type 3 analysis analogous to Type III sums of squares in
the GLM procedure. A Type 3 analysis does not depend on the order in which the terms for the model
are specified. A GENMOD procedure Type 3 analysis consists of specifying a model and computing
likelihood ratio statistics for Type III contrasts for each term in the model. The contrasts are defined
in the same way as they are in the GLM procedure. The GENMOD procedure optionally computes
Wald statistics for Type III contrasts. This is computationally less expensive than likelihood ratio
statistics, but it is thought to be less accurate because the specified significance level of hypothesis
tests based on the Wald statistic might not be as close to the actual significance level as it is for
likelihood ratio tests.
A Type 3 analysis generalizes the use of Type III estimable functions in linear models. Briefly, a
Type III estimable function (contrast) for an effect is a linear function of the model parameters that
involves the parameters of the effect and any interactions with that effect. A test of the hypothesis
that the Type III contrast for a main effect is equal to 0 is intended to test the significance of the main
effect in the presence of interactions. See Chapter 39, “The GLM Procedure,” and Chapter 15, “The
Four Types of Estimable Functions,” for more information about Type III estimable functions. Also
refer to Littell, Freund, and Spector (1991).
Additional features of the GENMOD procedure include the following:
likelihood ratio statistics for user-defined contrasts—that is, linear functions of the parameters
and p-values based on their asymptotic chi-square distributions
estimated values, standard errors, and confidence limits for user-defined contrasts and least
squares means
ability to create a SAS data set corresponding to most tables displayed by the procedure (see
Table 37.8 and Table 37.9)
confidence intervals for model parameters based on either the profile likelihood function or
asymptotic normality
syntax similar to that of PROC GLM for the specification of the response and model effects,
including interaction terms and automatic coding of classification variables
ability to fit GEE models for clustered response data
ability to perform Bayesian analysis by Gibbs sampling
Getting Started: GENMOD Procedure F 2435
Getting Started: GENMOD Procedure
Poisson Regression
You can use the GENMOD procedure to fit a variety of statistical models. A typical use of PROC
GENMOD is to perform Poisson regression.
You can use the Poisson distribution to model the distribution of cell counts in a multiway contingency
table. Aitkin et al. (1989) have used this method to model insurance claims data. Suppose the
following hypothetical insurance claims data are classified by two factors: age group (with two
levels) and car type (with three levels).
data insure;
input n c car$ age;
ln = log(n);
datalines;
500
42 small 1
1200 37 medium 1
100
1 large 1
400 101 small 2
500
73 medium 2
300
14 large 2
;
run;
In the preceding data set, the variable n represents the number of insurance policyholders and the
variable c represents the number of insurance claims. The variable car is the type of car involved
(classified into three groups) and the variable age is the age group of a policyholder (classified into
two groups).
You can use PROC GENMOD to perform a Poisson regression analysis of these data with a log link
function. This type of model is sometimes called a log-linear model.
Assume that the number of claims c has a Poisson probability distribution and that its mean, i , is
related to the factors car and age for observation i by
log.i / D log.ni / C x0i ˇ
D log.ni / C ˇ0 C
cari .1/ˇ1 C cari .2/ˇ2 C cari .3/ˇ3 C
agei .1/ˇ4 C agei .2/ˇ5
The indicator variables cari .j / and agei .j / are associated with the j th level of the variables car and
2436 F Chapter 37: The GENMOD Procedure
age for observation i
cari .j / D
1 if car D j
0 if car ¤ j
The ˇs are unknown parameters to be estimated by the procedure. The logarithm of the variable n is
used as an offset—that is, a regression variable with a constant coefficient of 1 for each observation.
A log-linear relationship between the mean and the factors car and age is specified by the log link
function. The log link function ensures that the mean number of insurance claims for each car and
age group predicted from the fitted model is positive.
The following statements invoke the GENMOD procedure to perform this analysis:
proc genmod data=insure;
class car age;
model c = car age / dist
= poisson
link
= log
offset = ln;
run;
The variables car and age are specified as CLASS variables so that PROC GENMOD automatically
generates the indicator variables associated with car and age.
The MODEL statement specifies c as the response variable and car and age as explanatory variables.
An intercept term is included by default. Thus, the model matrix X (the matrix that has as its i th row
the transpose of the covariate vector for the i th observation) consists of a column of 1s representing
the intercept term and columns of 0s and 1s derived from indicator variables representing the levels
of the car and age variables.
That is, the model matrix is
2
1 1 0 0
6 1 0 1 0
6
6 1 0 0 1
XD6
6 1 1 0 0
6
4 1 0 1 0
1 0 0 1
1
1
1
0
0
0
0
0
0
1
1
1
3
7
7
7
7
7
7
5
where the first column corresponds to the intercept, the next three columns correspond to the variable
car, and the last two columns correspond to the variable age.
The response distribution is specified as Poisson, and the link function is chosen to be log. That is,
the Poisson mean parameter is related to the linear predictor by
log./ D x0i ˇ
The logarithm of n is specified as an offset variable, as is common in this type of analysis. In this
case, the offset variable serves to normalize the fitted cell means to a per-policyholder basis, since
the total number of claims, not individual policyholder claims, is observed.
Poisson Regression F 2437
PROC GENMOD produces the following default output from the preceding statements.
Figure 37.1 Model Information
The GENMOD Procedure
Model Information
Data Set
Distribution
Link Function
Dependent Variable
Offset Variable
WORK.INSURE
Poisson
Log
c
ln
The “Model Information” table displayed in Figure 37.1 provides information about the specified
model and the input data set.
Figure 37.2 Class Level Information
Class Level Information
Class
Levels
car
age
Values
3
2
large medium small
1 2
Figure 37.2 displays the “Class Level Information” table, which identifies the levels of the classification variables that are used in the model. Note that car is a character variable, and the values are
sorted in alphabetical order. This is the default sort order, but you can select different sort orders
with the ORDER= option in the PROC GENMOD statement.
Figure 37.3 Goodness of Fit
Criteria For Assessing Goodness Of Fit
Criterion
Deviance
Scaled Deviance
Pearson Chi-Square
Scaled Pearson X2
Log Likelihood
Full Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)
DF
Value
Value/DF
2
2
2
2
2.8207
2.8207
2.8416
2.8416
837.4533
-16.4638
40.9276
80.9276
40.0946
1.4103
1.4103
1.4208
1.4208
The “Criteria For Assessing Goodness Of Fit” table displayed in Figure 37.3 contains statistics that
summarize the fit of the specified model. These statistics are helpful in judging the adequacy of
a model and in comparing it with other models under consideration. If you compare the deviance
2438 F Chapter 37: The GENMOD Procedure
of 2.8207 with its asymptotic chi-square with 2 degrees of freedom distribution, you find that the
p-value is 0.24. This indicates that the specified model fits the data reasonably well.
Figure 37.4 Analysis of Parameter Estimates
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
Intercept
car
car
car
age
age
Scale
large
medium
small
1
2
DF
Estimate
Standard
Error
1
1
1
0
1
0
0
-1.3168
-1.7643
-0.6928
0.0000
-1.3199
0.0000
1.0000
0.0903
0.2724
0.1282
0.0000
0.1359
0.0000
0.0000
Wald 95% Confidence
Limits
-1.4937
-2.2981
-0.9441
0.0000
-1.5863
0.0000
1.0000
-1.1398
-1.2304
-0.4414
0.0000
-1.0536
0.0000
1.0000
Wald
Chi-Square
212.73
41.96
29.18
.
94.34
.
Analysis Of Maximum Likelihood
Parameter Estimates
Parameter
Intercept
car
car
car
age
age
Scale
Pr > ChiSq
large
medium
small
1
2
<.0001
<.0001
<.0001
.
<.0001
.
NOTE: The scale parameter was held fixed.
Figure 37.4 displays the “Analysis Of Parameter Estimates” table, which summarizes the results
of the iterative parameter estimation process. For each parameter in the model, PROC GENMOD
displays columns with the parameter name, the degrees of freedom associated with the parameter,
the estimated parameter value, the standard error of the parameter estimate, the confidence intervals,
and the Wald chi-square statistic and associated p-value for testing the significance of the parameter
to the model. If a column of the model matrix corresponding to a parameter is found to be linearly
dependent, or aliased, with columns corresponding to parameters preceding it in the model, PROC
GENMOD assigns it zero degrees of freedom and displays a value of zero for both the parameter
estimate and its standard error.
This table includes a row for a scale parameter, even though there is no free scale parameter in the
Poisson distribution. See the section “Response Probability Distributions” on page 2510 for the
form of the Poisson probability distribution. PROC GENMOD allows the specification of a scale
parameter to fit overdispersed Poisson and binomial distributions. In such cases, the SCALE row
indicates the value of the overdispersion scale parameter used in adjusting output statistics. See
the section “Overdispersion” on page 2521 for more about overdispersion and the meaning of the
SCALE parameter output by the GENMOD procedure. PROC GENMOD displays a note indicating
that the scale parameter is fixed—that is, not estimated by the iterative fitting process.
Poisson Regression F 2439
It is usually of interest to assess the importance of the main effects in the model. Type 1 and Type 3
analyses generate statistical tests for the significance of these effects. You can request these analyses
with the TYPE1 and TYPE3 options in the MODEL statement, as follows:
proc genmod data=insure;
class car age;
model c = car age / dist
= poisson
link
= log
offset = ln
type1
type3;
run;
The results of these analyses are summarized in the figures that follow.
Figure 37.5 Type 1 Analysis
The GENMOD Procedure
LR Statistics For Type 1 Analysis
Source
Deviance
DF
ChiSquare
Pr > ChiSq
Intercept
car
age
175.1536
107.4620
2.8207
2
1
67.69
104.64
<.0001
<.0001
In the table for Type 1 analysis displayed in Figure 37.5, each entry in the deviance column represents
the deviance for the model containing the effect for that row and all effects preceding it in the table.
For example, the deviance corresponding to car in the table is the deviance of the model containing
an intercept and car. As more terms are included in the model, the deviance decreases.
Entries in the chi-square column are likelihood ratio statistics for testing the significance of the
effect added to the model containing all the preceding effects. The chi-square value of 67.69 for
car represents twice the difference in log likelihoods between fitting a model with only an intercept
term and a model with an intercept and car. Since the scale parameter is set to 1 in this analysis, this
is equal to the difference in deviances. Since two additional parameters are involved, this statistic
can be compared with a chi-square distribution with two degrees of freedom. The resulting p-value
(labeled Pr>Chi) of less than 0.0001 indicates that this variable is highly significant. Similarly, the
chi-square value of 104.64 for age represents the difference in log likelihoods between the model
with the intercept and car and the model with the intercept, car, and age. This effect is also highly
significant, as indicated by the small p-value.
2440 F Chapter 37: The GENMOD Procedure
Figure 37.6 Type 3 Analysis
LR Statistics For Type 3 Analysis
Source
car
age
DF
ChiSquare
Pr > ChiSq
2
1
72.82
104.64
<.0001
<.0001
The Type 3 analysis results in the same conclusions as the Type 1 analysis. The Type 3 chi-square
value for the car variable, for example, is twice the difference between the log likelihood for the
model with the variables Intercept, car, and age included and the log likelihood for the model with the
car variable excluded. The hypothesis tested in this case is the significance of the variable car given
that the variable age is in the model. In other words, it tests the additional contribution of car in the
model.
The values of the Type 3 likelihood ratio statistics for the car and age variables indicate that both
of these factors are highly significant in determining the claims performance of the insurance
policyholders.
Bayesian Analysis of a Linear Regression Model
Neter et al. (1996) describe a study of 54 patients undergoing a certain kind of liver operation in
a surgical unit. The data set Surg contains survival time and certain covariates for each patient.
Observations for the first 20 patients in the data set Surg are shown in Figure 37.7.
Bayesian Analysis of a Linear Regression Model F 2441
Figure 37.7 Surgical Unit Data
Obs
x1
x2
x3
x4
y
logy
Logx1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
6.7
5.1
7.4
6.5
7.8
5.8
5.7
3.7
6.0
3.7
6.3
6.7
5.8
5.8
7.7
7.4
6.0
3.7
7.3
5.6
62
59
57
73
65
38
46
68
67
76
84
51
96
83
62
74
85
51
68
57
81
66
83
41
115
72
63
81
93
94
83
43
114
88
67
68
28
41
74
87
2.59
1.70
2.16
2.01
4.30
1.42
1.91
2.57
2.50
2.40
4.13
1.86
3.95
3.95
3.40
2.40
2.98
1.55
3.56
3.02
200
101
204
101
509
80
80
127
202
203
329
65
830
330
168
217
87
34
215
172
2.3010
2.0043
2.3096
2.0043
2.7067
1.9031
1.9031
2.1038
2.3054
2.3075
2.5172
1.8129
2.9191
2.5185
2.2253
2.3365
1.9395
1.5315
2.3324
2.2355
1.90211
1.62924
2.00148
1.87180
2.05412
1.75786
1.74047
1.30833
1.79176
1.30833
1.84055
1.90211
1.75786
1.75786
2.04122
2.00148
1.79176
1.30833
1.98787
1.72277
Consider the model
Y D ˇ0 C ˇ1 LogX1 C ˇ2 X 2 C ˇ3 X 3 C ˇ4 X 4 C where Y is the survival time, LogX1 is log(blood-clotting score), X2 is a prognostic index, X3 is an
enzyme function test score, X4 is a liver function test score, and is an N.0; 2 / error term.
A question of scientific interest is whether blood clotting score has a positive effect on survival
time. Using PROC GENMOD, you can obtain a maximum likelihood estimate of the coefficient and
construct a null point hypothesis to test whether ˇ1 is equal to 0. However, if you are interested in
finding the probability that the coefficient is positive, Bayesian analysis offers a convenient alternative.
You can use Bayesian analysis to directly estimate the conditional probability, Pr.ˇ1 > 0jY/, using
the posterior distribution samples, which are produced as part of the output by PROC GENMOD.
The example that follows shows how to use PROC GENMOD to carry out a Bayesian analysis of the
linear model with a normal error term. The SEED= option is specified to maintain reproducibility;
no other options are specified in the BAYES statement. By default, a uniform prior distribution
is assumed on the regression coefficients. The uniform prior is a flat prior on the real line with
a distribution that reflects ignorance of the location of the parameter, placing equal likelihood on
all possible values the regression coefficient can take. Using the uniform prior in the following
example, you would expect the Bayesian estimates to resemble the classical results of maximizing
the likelihood. If you can elicit an informative prior distribution for the regression coefficients, you
should use the COEFFPRIOR= option to specify it. A default noninformative gamma prior is used
for the scale parameter .
2442 F Chapter 37: The GENMOD Procedure
You should make sure that the posterior distribution samples have achieved convergence before
using them for Bayesian inference. PROC GENMOD produces three convergence diagnostics by
default. If ODS Graphics is enabled as specified in the following SAS statements, diagnostic plots
are also displayed. See the section “Assessing Markov Chain Convergence” on page 155 for more
information about convergence diagnostics and their interpretation.
Summary statistics of the posterior distribution samples are produced by default. However, these
statistics might not be sufficient for carrying out your Bayesian inference, and further processing
of the posterior samples might be necessary. The following SAS statements request the Bayesian
analysis, and the OUTPOST= option saves the samples in the SAS data set PostSurg for further
processing:
ods graphics on;
proc genmod data=Surg;
model y = Logx1 X2 X3 X4 / dist=normal;
bayes seed=1 OutPost=PostSurg;
run;
ods graphics off;
The results of this analysis are shown in the following figures.
The “Model Information” table in Figure 37.8 summarizes information about the model you fit and
the size of the simulation.
Figure 37.8 Model Information
The GENMOD Procedure
Bayesian Analysis
Model Information
Data Set
Burn-In Size
MC Sample Size
Thinning
Sampling Algorithm
Distribution
Link Function
Dependent Variable
WORK.SURG
2000
10000
1
Conjugate
Normal
Identity
y
Survival Time
Bayesian Analysis of a Linear Regression Model F 2443
The “Analysis of Maximum Likelihood Parameter Estimates” table in Figure 37.9 summarizes
maximum likelihood estimates of the model parameters.
Figure 37.9 Maximum Likelihood Parameter Estimates
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
DF
Estimate
Standard
Error
Intercept
Logx1
x2
x3
x4
Scale
1
1
1
1
1
1
-730.559
171.8758
4.3019
4.0309
18.1377
59.8591
85.4333
38.2250
0.5566
0.4996
12.0721
5.7599
Wald 95% Confidence
Limits
-898.005
96.9561
3.2109
3.0517
-5.5232
49.5705
-563.112
246.7954
5.3929
5.0100
41.7986
72.2832
NOTE: The scale parameter was estimated by maximum likelihood.
Since no prior distributions for the regression coefficients were specified, the default noninformative
uniform distributions shown in the “Uniform Prior for Regression Coefficients” table in Figure 37.10
are used. Noninformative priors are appropriate if you have no prior knowledge of the likely range
of values of the parameters, and if you want to make probability statements about the parameters or
functions of the parameters. See, for example, Ibrahim, Chen, and Sinha (2001) for more information
about choosing prior distributions.
Figure 37.10 Regression Coefficient Priors
The GENMOD Procedure
Bayesian Analysis
Uniform Prior for Regression Coefficients
Parameter
Prior
Intercept
Logx1
x2
x3
x4
Constant
Constant
Constant
Constant
Constant
2444 F Chapter 37: The GENMOD Procedure
The default noninformative gamma prior distribution for the normal scale parameter is shown in the
“Independent Prior Distributions for Model Parameters” table in Figure 37.11.
Figure 37.11 Scale Parameter Prior
Independent Prior Distributions for Model Parameters
Parameter
Prior
Distribution
Hyperparameters
Shape
Scale
Dispersion
Inverse Gamma
2.001
0.0001
By default, the maximum likelihood estimates of the regression parameters are used as the starting
values for the simulation when noninformative prior distributions are used. These are listed in the
“Initial Values and Seeds” table in Figure 37.12.
Figure 37.12 MCMC Initial Values and Seeds
Initial Values of the Chain
Chain
Seed
Intercept
Logx1
x2
x3
x4
1
1
-730.559
171.8758
4.301896
4.030878
18.1377
Initial Values of the Chain
Dispersion
3223.694
Summary statistics for the posterior sample are displayed in the “Fit Statistics,” “Descriptive Statistics
for the Posterior Sample,” “Interval Statistics for the Posterior Sample,” and “Posterior Correlation
Matrix” tables in Figure 37.13, Figure 37.14, Figure 37.15, and Figure 37.16, respectively.
Figure 37.13 Fit Statistics
Fit Statistics
DIC (smaller is better)
pD (effective number of parameters)
608.411
6.571
Bayesian Analysis of a Linear Regression Model F 2445
Figure 37.14 Descriptive Statistics
The GENMOD Procedure
Bayesian Analysis
Posterior Summaries
Parameter
Intercept
Logx1
x2
x3
x4
Dispersion
N
Mean
Standard
Deviation
25%
10000
10000
10000
10000
10000
10000
-730.1
171.7
4.3000
4.0310
18.0888
3795.9
91.0133
40.3792
0.5989
0.5354
12.8949
770.4
-789.6
144.3
3.8990
3.6645
9.4919
3247.6
Percentiles
50%
-729.6
171.8
4.2932
4.0265
18.0430
3694.7
75%
-670.5
198.6
4.6951
4.3910
26.7881
4238.2
Figure 37.15 Interval Statistics
Posterior Intervals
Parameter
Alpha
Equal-Tail Interval
Intercept
Logx1
x2
x3
x4
Dispersion
0.050
0.050
0.050
0.050
0.050
0.050
-908.7
92.4773
3.1062
2.9812
-7.2646
2569.0
HPD Interval
-551.0
251.6
5.4839
5.1041
43.6506
5548.5
-906.2
94.2813
3.1747
2.9532
-5.9839
2389.4
-549.2
253.0
5.5328
5.0612
44.6427
5308.8
Figure 37.16 Posterior Sample Correlation Matrix
Posterior Correlation Matrix
Parameter
Intercept
Logx1
x2
x3
x4
Dispersion
Intercept
Logx1
x2
x3
x4
Dispersion
1.000
-0.856
-0.580
-0.712
0.579
-0.002
-0.856
1.000
0.285
0.490
-0.636
0.009
-0.580
0.285
1.000
0.302
-0.492
-0.007
-0.712
0.490
0.302
1.000
-0.616
-0.004
0.579
-0.636
-0.492
-0.616
1.000
0.002
-0.002
0.009
-0.007
-0.004
0.002
1.000
Since noninformative prior distributions were used, the posterior sample means, standard deviations,
and interval statistics shown in Figure 37.13 and Figure 37.14 are consistent with the maximum
likelihood estimates shown in Figure 37.9.
By default, PROC GENMOD computes three convergence diagnostics: the lag1, lag5, lag10, and
lag50 autocorrelations (Figure 37.17); Geweke diagnostic statistics (Figure 37.18); and effective
sample sizes (Figure 37.19). There is no indication that the Markov chain has not converged.
2446 F Chapter 37: The GENMOD Procedure
See the section “Assessing Markov Chain Convergence” on page 155 for more information about
convergence diagnostics and their interpretation.
Figure 37.17 Posterior Sample Autocorrelations
The GENMOD Procedure
Bayesian Analysis
Posterior Autocorrelations
Parameter
Intercept
Logx1
x2
x3
x4
Dispersion
Lag 1
Lag 5
Lag 10
Lag 50
-0.0050
0.0030
-0.0113
0.0019
-0.0001
-0.0019
-0.0023
-0.0063
-0.0046
0.0064
-0.0084
0.0088
-0.0138
-0.0070
-0.0235
-0.0073
0.0050
-0.0297
0.0032
-0.0034
-0.0139
0.0047
-0.0084
0.0025
Figure 37.18 Geweke Diagnostic Statistics
Geweke Diagnostics
Parameter
Intercept
Logx1
x2
x3
x4
Dispersion
z
Pr > |z|
-0.8783
1.4800
-0.0438
0.1000
-0.8893
0.1011
0.3798
0.1389
0.9651
0.9204
0.3739
0.9195
Figure 37.19 Effective Sample Sizes
Effective Sample Sizes
Parameter
Intercept
Logx1
x2
x3
x4
Dispersion
ESS
Autocorrelation
Time
Efficiency
10000.0
10000.0
10232.2
10000.0
10000.0
10000.0
1.0000
1.0000
0.9773
1.0000
1.0000
1.0000
1.0000
1.0000
1.0232
1.0000
1.0000
1.0000
Trace, autocorrelation, and density plots for the seven model parameters, shown in Figure 37.20
through Figure 37.25, are useful in diagnosing whether the Markov chain of posterior samples has
converged. These plots show no evidence that the chain has not converged. See the section “Visual
Analysis via Trace Plots” on page 155 for help with interpreting these diagnostic plots.
Bayesian Analysis of a Linear Regression Model F 2447
Figure 37.20 Diagnostic Plots for Intercept
2448 F Chapter 37: The GENMOD Procedure
Figure 37.21 Diagnostic Plots for logX1
Bayesian Analysis of a Linear Regression Model F 2449
Figure 37.22 Diagnostic Plots for X2
2450 F Chapter 37: The GENMOD Procedure
Figure 37.23 Diagnostic Plots for X3
Bayesian Analysis of a Linear Regression Model F 2451
Figure 37.24 Diagnostic Plots for X4
2452 F Chapter 37: The GENMOD Procedure
Figure 37.25 Diagnostic Plots for X5
Suppose, for illustration, a question of scientific interest is whether blood clotting score has a positive
effect on survival time. Since the model parameters are regarded as random quantities in a Bayesian
analysis, you can answer this question by estimating the conditional probability of ˇ1 being positive,
given the data, Pr.ˇ1 > 0jY/, from the posterior distribution samples. The following SAS statements
compute the estimate of the probability of ˇ1 being positive:
data Prob;
set PostSurg;
Indicator = (logX1 > 0);
label Indicator= 'log(Blood Clotting Score) > 0';
run;
proc Means data = Prob(keep=Indicator) n mean;
run;
As shown in Figure 37.26, there is a 1.00 probability of a positive relationship between the logarithm
of a blood clotting score and survival time, adjusted for the other covariates.
Generalized Estimating Equations F 2453
Figure 37.26 Probability That ˇ1 > 0
The MEANS Procedure
Analysis Variable : Indicator log(Blood Clotting Score) > 0
N
Mean
--------------------10000
0.9999000
---------------------
Generalized Estimating Equations
This section illustrates the use of the REPEATED statement to fit a GEE model, using repeated
measures data from the “Six Cities” study of the health effects of air pollution (Ware et al. 1984). The
data analyzed are the 16 selected cases in Lipsitz et al. (1994). The binary response is the wheezing
status of 16 children at ages 9, 10, 11, and 12 years. The mean response is modeled as a logistic
regression model by using the explanatory variables city of residence, age, and maternal smoking
status at the particular age. The binary responses for individual children are assumed to be equally
correlated, implying an exchangeable correlation structure.
The data set and SAS statements that fit the model by the GEE method are as follows:
data six;
input case city$ @@;
do i=1 to 4;
input age smoke wheeze @@;
output;
end;
datalines;
1 portage
9 0 1 10 0 1 11 0 1
2 kingston 9 1 1 10 2 1 11 2 0
3 kingston 9 0 1 10 0 0 11 1 0
4 portage
9 0 0 10 0 1 11 0 1
5 kingston 9 0 0 10 1 0 11 1 0
6 portage
9 0 0 10 1 0 11 1 0
7 kingston 9 1 0 10 1 0 11 0 0
8 portage
9 1 0 10 1 0 11 1 0
9 portage
9 2 1 10 2 0 11 1 0
10 kingston 9 0 0 10 0 0 11 0 0
11 kingston 9 1 1 10 0 0 11 0 1
12 portage
9 1 0 10 0 0 11 0 0
13 kingston 9 1 0 10 0 1 11 1 1
14 portage
9 1 0 10 2 0 11 1 0
15 kingston 9 1 0 10 1 0 11 1 0
16 portage
9 1 1 10 1 1 11 2 0
;
run;
12 0 0
12 2 0
12 1 0
12 1 0
12 1 0
12 1 0
12 0 0
12 2 0
12 1 0
12 1 0
12 0 1
12 0 0
12 1 1
12 2 1
12 2 1
12 1 0
2454 F Chapter 37: The GENMOD Procedure
proc genmod data=six ;
class case city ;
model wheeze = city age smoke / dist=bin;
repeated subject=case / type=exch covb corrw;
run;
The CLASS statement and the MODEL statement specify the model for the mean of the wheeze
variable response as a logistic regression with city, age, and smoke as independent variables, just as
for an ordinary logistic regression.
The REPEATED statement invokes the GEE method, specifies the correlation structure, and controls
the displayed output from the GEE model. The option SUBJECT=CASE specifies that individual
subjects be identified in the input data set by the variable case. The SUBJECT= variable case must be
listed in the CLASS statement. Measurements on individual subjects at ages 9, 10, 11, and 12 are in
the proper order in the data set, so the WITHINSUBJECT= option is not required. The TYPE=EXCH
option specifies an exchangeable working correlation structure, the COVB option specifies that the
parameter estimate covariance matrix be displayed, and the CORRW option specifies that the final
working correlation be displayed.
Initial parameter estimates for iterative fitting of the GEE model are computed as in an ordinary
generalized linear model, as described previously. Results of the initial model fit displayed as part
of the generated output are not shown here. Statistics for the initial model fit such as parameter
estimates, standard errors, deviances, and Pearson chi-squares do not apply to the GEE model and
are valid only for the initial model fit. The following figures display information that applies to the
GEE model fit.
Figure 37.27 displays general information about the GEE model fit.
Figure 37.27 GEE Model Information
The GENMOD Procedure
GEE Model Information
Correlation Structure
Subject Effect
Number of Clusters
Correlation Matrix Dimension
Maximum Cluster Size
Minimum Cluster Size
Exchangeable
case (16 levels)
16
4
4
4
Figure 37.28 displays the parameter estimate covariance matrices specified by the COVB option.
Both model-based and empirical covariances are produced.
Generalized Estimating Equations F 2455
Figure 37.28 GEE Parameter Estimate Covariance Matrices
Covariance Matrix (Model-Based)
Prm1
Prm2
Prm4
Prm5
Prm1
Prm2
Prm4
Prm5
5.74947
-0.22257
-0.53472
0.01655
-0.22257
0.45478
-0.002410
0.01876
-0.53472
-0.002410
0.05300
-0.01658
0.01655
0.01876
-0.01658
0.19104
Covariance Matrix (Empirical)
Prm1
Prm2
Prm4
Prm5
Prm1
Prm2
Prm4
Prm5
9.33994
-0.85104
-0.83253
-0.16534
-0.85104
0.47368
0.05736
0.04023
-0.83253
0.05736
0.07778
-0.002364
-0.16534
0.04023
-0.002364
0.13051
The exchangeable working correlation matrix specified by the CORRW option is displayed in
Figure 37.29.
Figure 37.29 GEE Working Correlation Matrix
Working Correlation Matrix
Row1
Row2
Row3
Row4
Col1
Col2
Col3
Col4
1.0000
0.1648
0.1648
0.1648
0.1648
1.0000
0.1648
0.1648
0.1648
0.1648
1.0000
0.1648
0.1648
0.1648
0.1648
1.0000
The parameter estimates table, displayed in Figure 37.30, contains parameter estimates, standard
errors, confidence intervals, Z scores, and p-values for the parameter estimates. Empirical standard
error estimates are used in this table. A table that displays model-based standard errors can be created
by using the REPEATED statement option MODELSE.
Figure 37.30 GEE Parameter Estimates Table
Analysis Of GEE Parameter Estimates
Empirical Standard Error Estimates
Parameter
Intercept
city
kingston
city
portage
age
smoke
Estimate
Standard
Error
-1.2751
-0.1223
0.0000
0.2036
0.0935
3.0561
0.6882
0.0000
0.2789
0.3613
95% Confidence
Limits
-7.2650
-1.4713
0.0000
-0.3431
-0.6145
4.7148
1.2266
0.0000
0.7502
0.8016
Z Pr > |Z|
-0.42
-0.18
.
0.73
0.26
0.6765
0.8589
.
0.4655
0.7957
2456 F Chapter 37: The GENMOD Procedure
Syntax: GENMOD Procedure
You can specify the following statements in the GENMOD procedure. Items within the < > are
optional.
PROC GENMOD < options > ;
ASSESS | ASSESSMENT VAR=(effect) | LINK < / options > ;
BAYES < options > ;
BY variables ;
CLASS variable < (options) >: : :< variable < (options) > > < / options > ;
CONTRAST ’label’ contrast-specification < / options > ;
DEVIANCE variable = expression ;
EFFECTPLOT < plot-type < (plot-definition-options) > > < / options > ;
ESTIMATE ’label’ effect values < ,. . . effect values > < / options > ;
EXACT < ’label’ > < INTERCEPT > < effects > < / options > ;
EXACTOPTIONS options ;
FREQ | FREQUENCY variable ;
FWDLINK variable = expression ;
INVLINK variable = expression ;
LSMEANS < model-effects > < / options > ;
LSMESTIMATE model-effect < ‘label’ > values < divisor =n > < , . . . < ‘label’ > values
< divisor =n > > < / options > ;
MODEL response = < effects > < / options > ;
OUTPUT < OUT=SAS-data-set > < keyword=name. . . keyword=name > ;
Programming statements ;
REPEATED SUBJECT=subject-effect < / options > ;
SLICE model-effect < / options > ;
STORE < OUT= >item-store-name < / LABEL=‘label’ > ;
STRATA variable < (option) > : : : < variable < (option) > > < / options > ;
WEIGHT | SCWGT variable ;
VARIANCE variable = expression ;
ZEROMODEL < effects > < / options > ;
The ASSESS, BAYES, BY, CLASS, CONTRAST, DEVIANCE, ESTIMATE, FREQUENCY,
FWDLINK, INVLINK, MODEL, OUTPUT, programming statements, REPEATED, VARIANCE,
WEIGHT, and ZEROMODEL statements are described in full after the PROC GENMOD statement
in alphabetical order. The EFFECTPLOT, LSMEANS, LSMESTIMATE, SLICE, and STORE
statements are common to many procedures. Summary descriptions of functionality and syntax for
these statements are also given after the PROC GENMOD statement in alphabetical order, and full
documentation about them is available in Chapter 19, “Shared Concepts and Topics.”
The PROC GENMOD statement invokes the procedure. All statements other than the MODEL
statement are optional. The CLASS statement, if present, must precede the MODEL statement, and
the CONTRAST and EXACT statements must come after the MODEL statement.
PROC GENMOD Statement F 2457
PROC GENMOD Statement
PROC GENMOD < options > ;
The PROC GENMOD statement invokes the procedure. You can specify the following options.
DATA=SAS-data-set
specifies 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.
DESCENDING
DESCEND
DESC
specifies that the levels of the response variable for the ordinal multinomial model and the
binomial model with single variable response syntax be sorted in the reverse of the default
order. For example, if RORDER=FORMATTED (the default), the DESCENDING option
causes the levels to be sorted from highest to lowest instead of from lowest to highest. If
RORDER=FREQ, the DESCENDING option causes the levels to be sorted from lowest
frequency count to highest instead of from highest to lowest.
EXACTONLY
requests only the exact analyses. The asymptotic analysis that PROC GENMOD usually
performs is suppressed.
NAMELEN=n
specifies the length of effect names in tables and output data sets to be n characters long, where
n is a value between 20 and 200 characters. The default length is 20 characters.
ORDER=DATA | FORMATTED | FREQ | INTERNAL
specifies the order in which to sort the levels of the classification variables (which are specified
in the CLASS statement). The ORDER= option can be useful when you use the CONTRAST
or ESTIMATE statement because it determines which parameters in the model correspond to
each level in the data.
This option applies to the levels for all classification variables, except when you use the
(default) ORDER=FORMATTED option with numeric classification variables that have no
explicit format. With this option, the levels of such variables are ordered by their internal
value.
The ORDER= option can take the following values:
Value of ORDER=
Levels Sorted By
DATA
Order of appearance in the input data set
FORMATTED
External formatted value, except for numeric variables with
no explicit format, which are sorted by their unformatted
(internal) value
2458 F Chapter 37: The GENMOD Procedure
Table 37.0
continued
Value of ORDER=
Levels Sorted By
FREQ
Descending frequency count; levels with the most observations come first in the order
INTERNAL
Unformatted value
By default, ORDER=FORMATTED. For FORMATTED and INTERNAL, the sort order is
machine-dependent. For more information about sorting 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.
PLOTS < (global-plot-option) >= plot-request < (options) >
PLOTS < (global-plot-options) > < = (plot-request < (options) > < ... plot-request < (options) > >) >
specifies plots to be created using ODS Graphics. Many of the observational statistics in the
output data set can be plotted using this option. You are not required to create an output data
set in order to produce a plot. When you specify only one plot request, you can omit the
parentheses around the plot request. Here are some examples:
PLOTS=ALL
PLOTS=PREDICTED
PLOTS=(PREDICTED RESCHI)
PLOTS(UNPACK)=DFBETA
You must enable ODS Graphics before requesting plots, for example, like this:
ods graphics on;
proc genmod plots=all;
model y = x;
run;
ods graphics off;
Any specified global plot options apply to all plots that are specified with plot requests. The
following global plot options are available.
CLUSTERLABEL
displays formatted levels of the SUBJECT= effect instead of plot symbols. This option
applies only to diagnostic statistics for models fit by GEEs that are plotted against cluster
number, and provides a way to identify cluster level names with corresponding ordered
cluster numbers.
UNPACK
displays multiple plots individually. The default is to display related multiple plots in a
panel.
See the section “OUTPUT Statement” on page 2499 for definitions of the statistics specified
with the plot requests. The plot requests include the following:
PROC GENMOD Statement F 2459
ALL
produces all available plots.
COOKSD
DOBS
plots the Cook’s distance statistic as a function of observation number.
DFBETA
plots the ˇ deletion statistic as a function of observation number for each regression
parameter in the model.
DFBETAS
plots the standardized ˇ deletion statistic as a function of observation number for each
regression parameter in the model.
LEVERAGE
plots the leverage as a function of observation number.
PREDICTED< (option) >
plots predicted values with confidence limits as a function of observation number. The
PREDICTED plot request has the following option:
CLM
includes confidence limits in the predicted value plot.
PZERO
plots the zero inflation probability for zero-inflated Poisson and negative binomial
models as a function of observation number.
RESCHI< (options) >
The RESCHI plot request has the following options:
INDEX
plots as a function of observation number.
XBETA
plots as a function of linear predictor.
If you do not specify an option, Pearson residuals are plotted as a function of observation
number.
RESDEV< (options) >
plots deviance residuals. The RESDEV plot request has the following options:
INDEX
plots as a function of observation number.
XBETA
plots as a function of linear predictor.
2460 F Chapter 37: The GENMOD Procedure
If you do not specify an option, deviance residuals are plotted as a function of observation
number.
RESLIK< (options) >
plots likelihood residuals. The RESLIK plot request has the following options:
INDEX
plots as a function of observation number.
XBETA
plots as a function of linear predictor.
If you do not specify an option, likelihood residuals are plotted as a function of observation number.
RESRAW< (options) >
plots raw residuals. The RESRAW plot request has the following options:
INDEX
plots as a function of observation number.
XBETA
plots as a function of linear predictor.
If you do not specify an option, raw residuals are plotted as a function of observation
number.
STDRESCHI< (options) >
plots standardized Pearson residuals. The STDRESCHI plot request has the following
options:
INDEX
plots as a function of observation number.
XBETA
plots as a function of linear predictor.
If you do not specify an option, standardized Pearson residuals are plotted as a function
of observation number.
STDRESDEV< (options) >
plots standardized deviance residuals. The STDRESDEV plot request has the following
options:
INDEX
plots as a function of observation number.
XBETA
plots as a function of linear predictor.
If you do not specify an option, standardized deviance residuals are plotted as a function
of observation number.
PROC GENMOD Statement F 2461
If you fit a model by using generalized estimating equations (GEEs), the following additional
plot requests are available:
CLEVERAGE
plots the cluster leverage as a function of ordered cluster.
CLUSTERCOOKSD
DCLS
plots the cluster Cook’s distance statistic as a function of ordered cluster.
CLUSTERDFIT
MCLS
plots the studentized cluster Cook’s distance statistic as a function of ordered cluster.
DFBETAC
plots the cluster deletion statistic as a function of ordered cluster for each regression
parameter in the model.
DFBETACS
plots the standardized cluster deletion statistic as a function of ordered cluster for each
regression parameter in the model.
RORDER=keyword
specifies the sorting order for the levels of the response variable. This order determines which intercept parameter in the model corresponds to each level in the data. If
RORDER=FORMATTED for numeric variables for which you have supplied no explicit
format, the levels are ordered by their internal values. Note that this represents a change
from previous releases for how class levels are ordered. Before SAS 8, numeric class levels
with no explicit format were ordered by their BEST12. formatted values, and to revert to the
previous order you can specify this format explicitly for the response variable. The change
was implemented because the former default behavior for RORDER=FORMATTED often
resulted in levels not being ordered numerically and usually required the user to intervene with
an explicit format or RORDER=INTERNAL to get the more natural ordering. The following
table displays the valid keywords and describes how PROC GENMOD interprets them.
RORDER=keyword
Levels Sorted by
DATA
FORMATTED
Order of appearance in the input data set
External formatted value, except for numeric
variables with no explicit format, which are
sorted by their unformatted (internal) value
Descending frequency count; levels with the
most observations come first in the order
Unformatted value
FREQ
INTERNAL
By default, RORDER=FORMATTED. For RORDER=FORMATTED and RORDER=INTERNAL,
the sort order is machine dependent. The DESCENDING option in the PROC GENMOD
statement causes the response variable to be sorted in the reverse of the order displayed in the
2462 F Chapter 37: The GENMOD Procedure
previous table. For more information about sorting order, refer to the chapter on the SORT
procedure in the Base SAS Procedures Guide.
The NOPRINT option, which suppresses displayed output in other SAS procedures, is not
available in the PROC GENMOD statement. However, you can use the Output Delivery
System (ODS) to suppress all displayed output, store all output on disk for further analysis,
or create SAS data sets from selected output. You can suppress all displayed output with
the statement ODS SELECT NONE; and turn displayed output back on with the statement ODS
SELECT ALL;. See Table 37.8 and Table 37.9 for the names of output tables available from
PROC GENMOD. For more information about ODS, see Chapter 20, “Using the Output
Delivery System.”
ASSESS Statement
ASSESS VAR=(effect) | LINK < / options > ;
ASSESSMENT VAR=(effect) | LINK < / options > ;
The ASSESS statement computes and plots, using ODS Graphics, model-checking statistics based on
aggregates of residuals. See the section “Assessment of Models Based on Aggregates of Residuals”
on page 2541 for details about the model assessment methods available in GENMOD.
The types of aggregates available are cumulative residuals, moving sums of residuals, and loess
smoothed residuals. If you do not specify which aggregate to use, the assessments are based
on cumulative sums. PROC GENMOD uses ODS Graphics for graphical displays. For specific
information about the graphics available in PROC GENMOD, see the section “ODS Graphics” on
page 2572.
You must specify either LINK or VAR= in order to create an analysis.
LINK requests the assessment of the link function by performing the analysis with respect to the
linear predictor.
VAR=(effect) specifies that the functional form of a covariate be checked by performing the analysis
with respect to the variable identified by the effect. The effect must be specified in the MODEL
statement and must contain only continuous variables (variables not listed in a CLASS statement).
You can specify the following options after the slash (/).
CRPANEL
requests that a plot with four panels showing just a few of the paths from the default aggregate
plot to make it easier to compare simulated and observed paths. The plot in each panel contains
aggregates of the observed residuals and two simulated curves (fewer if NPATHS= is less than
8).
LOESS< (number ) >
LOWESS< (number ) >
requests model assessment based on loess smoothed residuals with optional number the
BAYES Statement F 2463
fraction of data used; number must be between zero and one. If number is not specified, the
default value one-third is used.
NPATHS=number
NPATH=number
PATHS=number
PATH=number
specifies the number of simulated paths to plot in the default aggregate residuals plot. The
default value of number is twenty.
RESAMPLE< =number >
RESAMPLES< =number >
specifies that a p-value be computed based on 1,000 simulated paths, or number paths, if
number is specified.
SEED=number
specifies a seed for the normal random number generator used in creating simulated realizations
of aggregates of residuals for plots and estimating p-values. Specifying a seed enables you to
produce identical graphs and p-values from one run of the procedure to the next run. If a seed
is not specified, or if number is negative or zero, a random number seed is derived from the
time of day.
WINDOW< (number ) >
requests assessment based on a moving sum window of width number. If number is not
specified, a value of one-half of the range of the x-coordinate is used.
BAYES Statement
BAYES < options > ;
The BAYES statement requests a Bayesian analysis of the regression model by using Gibbs sampling.
The Bayesian posterior samples (also known as the chain) for the regression parameters are not
tabulated. The Bayesian posterior samples (also known as the chain) for the regression parameters
can be output to a SAS data set. Table 37.1 summarizes the options available in the BAYES statement.
Table 37.1 BAYES Statement Options
Option
Monte Carlo Options
INITIAL=
INITIALMLE
METROPOLIS=
NBI=
NMC=
Description
Specifies the initial values of the chain
Specifies that maximum likelihood estimates be used as
initial values of the chain
Specifies the use of a Metropolis step in the ARMS algorithm
Specifies the number of burn-in iterations
Specifies the number of iterations after burn-in
2464 F Chapter 37: The GENMOD Procedure
Table 37.1 (continued)
Option
Description
SAMPLING=
Specifies the algorithm used to sample the posterior distribution
Specifies the random number generator seed
Controls the thinning of the Markov chain
SEED=
THINNING=
Model and Prior Options
COEFFPRIOR=
Specifies the prior of the regression coefficients
DISPERSIONPRIOR= Specifies the prior of the dispersion parameter
PRECISIONPRIOR=
Specifies the prior of the precision parameter
SCALEPRIOR=
Specifies the prior of the scale parameter
Summary Statistics and Convergence Diagnostics
DIAGNOSTICS=
Displays convergence diagnostics
PLOTS=
Displays diagnostic plots
STATISTICS=
Displays summary statistics of the posterior samples
Posterior Samples
OUTPOST=
Names a SAS data set for the posterior samples
The following list describes these options and their suboptions.
COEFFPRIOR=JEFFREYS< (option) > | NORMAL< (options) > | UNIFORM
COEFF=JEFFREYS< (options) > | NORMAL< (options) > | UNIFORM
CPRIOR=JEFFREYS< (options) > | NORMAL< (options) > | UNIFORM
specifies the prior distribution for the regression coefficients. The default is COEFFPRIOR=UNIFORM, which specifies the noninformative and improper prior of a constant.
Jeffreys’ prior is specified by COEFFPRIOR=JEFFREYS, which can be followed by the
1
following option in parentheses. Jeffreys’ prior is proportional to jI.ˇ/j 2 , where I.ˇ/ is the
Fisher information matrix. See the section “Jeffreys’ Prior” on page 2551 and Ibrahim and
Laud (1991) for more details.
CONDITIONAL
specifies that the Jeffreys’ prior, conditional on the current Markov chain value of the
1
generalized linear model precision parameter , is proportional to j I.ˇ/j 2 .
The normal prior is specified by COEFFPRIOR=NORMAL, which can be followed by one of
the following options enclosed in parentheses. However, if you do not specify an option, the
normal prior N.0; 106 I/, where I is the identity matrix, is used. See the section “Normal Prior”
on page 2551 for more details.
CONDITIONAL
specifies that the normal prior, conditional on the current Markov chain value of the
generalized linear model precision parameter , is N.; 1 †/, where and † are the
mean and covariance of the normal prior specified by other normal options.
BAYES Statement F 2465
INPUT=SAS-data-set
specifies a SAS data set containing the mean and covariance information of the normal
prior. The data set must have a _TYPE_ variable to represent the type of each observation and a variable for each regression coefficient. If the data set also contains a
_NAME_ variable, the values of this variable are used to identify the covariances for the
_TYPE_=’COV’ observations; otherwise, the _TYPE_=’COV’ observations are assumed
to be in the same order as the explanatory variables in the MODEL statement. PROC
GENMOD reads the mean vector from the observation with _TYPE_=’MEAN’ and
reads the covariance matrix from observations with _TYPE_=’COV’. For an independent
normal prior, the variances can be specified with _TYPE_=’VAR’; alternatively, the
precisions (inverse of the variances) can be specified with _TYPE_=’PRECISION’.
RELVAR< =c >
specifies the normal prior N.0; cJ/, where J is a diagonal matrix with diagonal elements
equal to the variances of the corresponding ML estimator. By default, c D 106 .
VAR< =c >
specifies the normal prior N.0; cI/, where I is the identity matrix.
DIAGNOSTICS=ALL | NONE | (keyword-list)
DIAG=ALL | NONE | (keyword-list)
controls the number of diagnostics produced. You can request all the following diagnostics
by specifying DIAGNOSTICS=ALL. If you do not want any of these diagnostics, specify
DIAGNOSTICS=NONE. If you want some but not all of the diagnostics, or if you want to
change certain settings of these diagnostics, specify a subset of the following keywords. The
default is DIAGNOSTICS=(AUTOCORR ESS GEWEKE).
AUTOCORR < (LAGS= numeric-list) >
computes the autocorrelations of lags given by LAGS= list for each parameter. Elements
in the list are truncated to integers and repeated values are removed. If the LAGS=
option is not specified, autocorrelations of lags 1, 5, 10, and 50 are computed for each
variable. See the section “Autocorrelations” on page 168 for details.
ESS
computes Carlin’s estimate of the effective sample size, the correlation time, and the
efficiency of the chain for each parameter. See the section “Effective Sample Size” on
page 168 for details.
GELMAN < (gelman-options) >
computes the Gelman and Rubin convergence diagnostics. You can specify one or more
of the following gelman-options:
NCHAIN | N=number
specifies the number of parallel chains used to compute the diagnostic, and must
be 2 or larger. The default is NCHAIN=3. If an INITIAL= data set is used,
NCHAIN defaults to the number of rows in the INITIAL= data set. If any number
other than this is specified with the NCHAIN= option, the NCHAIN= value is
ignored.
2466 F Chapter 37: The GENMOD Procedure
ALPHA=value
specifies the significance level for the upper bound. The default is ALPHA=0.05,
resulting in a 97.5% bound.
See the section “Gelman and Rubin Diagnostics” on page 160 for details.
GEWEKE < (geweke-options) >
computes the Geweke spectral density diagnostics, which are essentially a two-sample
t test between the first f1 portion and the last f2 portion of the chain. The default is
f1 D 0:1 and f2 D 0:5, but you can choose other fractions by using the following
geweke-options:
FRAC1=value
specifies the fraction f1 for the first window.
FRAC2=value
specifies the fraction f2 for the second window.
See the section “Geweke Diagnostics” on page 162 for details.
HEIDELBERGER < (heidel-options) >
computes the Heidelberger and Welch diagnostic for each variable, which consists of a
stationarity test of the null hypothesis that the sample values form a stationary process.
If the stationarity test is not rejected, a halfwidth test is then carried out. Optionally, you
can specify one or more of the following heidel-options:
SALPHA=value
specifies the ˛ level .0 < ˛ < 1/ for the stationarity test.
HALPHA=value
specifies the ˛ level .0 < ˛ < 1/ for the halfwidth test.
EPS=value
specifies a positive number such that if the halfwidth is less than times the
sample mean of the retained iterates, the halfwidth test is passed.
See the section “Heidelberger and Welch Diagnostics” on page 164 for details.
MCSE
MCERROR
computes the Monte Carlo standard error for each parameter. The Monte Caro standard
error, which measures the simulation accuracy, is the standard error of the posterior
mean estimate and is calculated as the posterior standard deviation divided by the square
root of the effective sample size. See the section “Standard Error of the Mean Estimate”
on page 169 for details.
RAFTERY< (raftery-options) >
computes the Raftery and Lewis diagnostics that evaluate the accuracy of the estimated
quantile (OQ for a given Q 2 .0; 1/) of a chain. OQ can achieve any degree of accuracy
when the chain is allowed to run for a long time. A stopping criterion is when the
BAYES Statement F 2467
estimated probability POQ D Pr. OQ / reaches within ˙R of the value Q with
probability S ; that is, Pr.Q R POQ Q C R/ D S. The following raftery-options
enable you to specify Q; R; S , and a precision level for the test:
QUANTILE | Q=value
specifies the order (a value between 0 and 1) of the quantile of interest. The
default is 0.025.
ACCURACY | R=value
specifies a small positive number as the margin of error for measuring the accuracy
of estimation of the quantile. The default is 0.005.
PROBABILITY | S=value
specifies the probability of attaining the accuracy of the estimation of the quantile.
The default is 0.95.
EPSILON | EPS=value
specifies the tolerance level (a small positive number) for the stationary test. The
default is 0.001.
See the section “Raftery and Lewis Diagnostics” on page 165 for details.
DISPERSIONPRIOR=GAMMA< (options) > | IGAMMA< (options) > | IMPROPER
DPRIOR=GAMMA< (options) > | IGAMMA< (options) > | IMPROPER
specifies that Gibbs sampling be performed on the generalized linear model dispersion parameter and the prior distribution for the dispersion parameter, if there is a dispersion parameter in
the model. For models that do not have a dispersion parameter (the Poisson and binomial), this
option is ignored. Note that you can specify Gibbs sampling on either the dispersion parameter
1
, the scale parameter D 2 , or the precision parameter D 1 , with the DPRIOR=,
SPRIOR=, and PPRIOR= options, respectively. These three parameters are transformations of
one another, and you should specify Gibbs sampling for only one of them.
a 1
bt
A gamma prior G.a; b/ with density f .t / D b.bt /€.a/e
is specified by DISPERSIONPRIOR=GAMMA, which can be followed by one of the following gamma-options enclosed in
parentheses. The hyperparameters a and b are the shape and inverse-scale parameters of the
gamma distribution, respectively. See the section “Gamma Prior” on page 2550 for details.
The default is G.10 4 ; 10 4 /.
RELSHAPE< =c >
O c/ distribution, where O is the MLE of the dispersion
specifies independent G.c ;
parameter. With this choice of hyperparameters, the mean of the prior distribution is O
O
and the variance is c . By default, c =10
4.
SHAPE=a
ISCALE=b
when both specified, results in a G.a; b/ prior.
SHAPE=c
when specified alone, results in a G.c; c/ prior.
2468 F Chapter 37: The GENMOD Procedure
ISCALE=c
when specified alone, results in a G.c; c/ prior.
a
b
An inverse gamma prior IG.a; b/ with density f .t / D €.a/
t .aC1/ e b=t is specified by
DISPERSIONPRIOR=IGAMMA, which can be followed by one of the following inverse
gamma-options enclosed in parentheses. The hyperparameters a and b are the shape and scale
parameters of the inverse gamma distribution, respectively. See the section “Inverse Gamma
Prior” on page 2550 for details. The default is IG.2:001; 0:001/.
RELSHAPE< =c >
O
specifies independent IG. cCO ; c/ distribution, where O is the MLE of the dispersion
O
parameter. With this choice of hyperparameters, the mean of the prior distribution is .
4
By default, c =10 .
SHAPE=a
SCALE=b
when both specified, results in a IG.a; b/ prior.
SHAPE=c
when specified alone, results in an IG.c; c/ prior.
SCALE=c
when specified alone, results in an IG.c; c/ prior.
An improper prior with density f .t / proportional to t
PRIOR=IMPROPER.
1
is specified with DISPERSION-
INITIAL=SAS-data-set
specifies the SAS data set that contains the initial values of the Markov chains. The INITIAL=
data set must contain all the variables of the model. You can specify multiple rows as the
initial values of the parallel chains for the Gelman-Rubin statistics, but posterior summaries,
diagnostics, and plots are computed only for the first chain. If the data set also contains the
variable _SEED_, the value of the _SEED_ variable is used as the seed of the random number
generator for the corresponding chain.
INITIALMLE
specifies that maximum likelihood estimates of the model parameters be used as initial values
of the Markov chain. If this option is not specified, estimates of the mode of the posterior
distribution obtained by optimization are used as initial values.
METROPOLIS=YES
METROPOLIS=NO
specifies the use of a Metropolis step to generate Gibbs samples for posterior distributions that
are not log concave. The default value is METROPOLIS=YES.
NBI=number
specifies the number of burn-in iterations before the chains are saved. The default is 2000.
NMC=number
specifies the number of iterations after the burn-in. The default is 10000.
BAYES Statement F 2469
OUTPOST=SAS-data-set
OUT=SAS-data-set
names the SAS data set that contains the posterior samples. See the section “OUTPOST=
Output Data Set” on page 2553 for more information. Alternatively, you can create the output
data set by specifying an ODS OUTPUT statement as follows:
ODS output posteriorsample = SAS-data-set ;
PRECISIONPRIOR=GAMMA< (options) > | IMPROPER
PPRIOR=GAMMA< (options) > | IMPROPER
specifies that Gibbs sampling be performed on the generalized linear model precision parameter
and the prior distribution for the precision parameter, if there is a precision parameter in the
model. For models that do not have a precision parameter (the Poisson and binomial), this
option is ignored. Note that you can specify Gibbs sampling on either the dispersion parameter
1
, the scale parameter D 2 , or the precision parameter D 1 , with the DPRIOR=,
SPRIOR=, and PPRIOR= options, respectively. These three parameters are transformations of
one another, and you should specify Gibbs sampling for only one of them.
a 1
bt
A gamma prior G.a; b/ with density f .t / D b.bt /€.a/e
is specified by PRECISIONPRIOR=GAMMA, which can be followed by one of the following gamma-options enclosed in
parentheses. The hyperparameters a and b are the shape and inverse-scale parameters of the
gamma distribution, respectively. See the section “Gamma Prior” on page 2550 for details.
The default is G.10 4 ; 10 4 /.
RELSHAPE< =c >
specifies independent G.c O ; c/ distribution, where O is the MLE of the dispersion
parameter. With this choice of hyperparameters, the mean of the prior distribution is O
and the variance is cO . By default, c D 10 4 .
SHAPE=a
ISCALE=b
when both specified, results in a G.a; b/ prior.
SHAPE=c
when specified alone, results in an G.c; c/ prior.
ISCALE=c
when specified alone, results in an G.c; c/ prior.
An improper prior with density f .t / proportional to t
PRIOR=IMPROPER.
1
is specified with PRECISION-
PLOTS< (global-plot-options) >= plot-request
PLOTS< (global-plot-options) >= (plot-request < . . . plot-request>)
controls the display of diagnostic plots. Three types of plots can be requested: trace plots,
autocorrelation function plots, and kernel density plots. By default, the plots are displayed
in panels unless the global plot option UNPACK is specified. Also, when you are specifying
more than one type of plots, the plots are displayed by parameters unless the global plot option
GROUPBY is specified. When you specify only one plot request, you can omit the parentheses
around the plot request. For example:
2470 F Chapter 37: The GENMOD Procedure
plots=none
plots(unpack)=trace
plots=(trace autocorr)
You must enable ODS Graphics before requesting plots. For example, the following SAS
statements enable ODS Graphics:
ods graphics on;
proc genmod;
model y=x;
bayes plots=trace;
run;
end;
ods graphics off;
The global plot options are as follows:
FRINGE
creates a fringe plot on the X axis of the density plot.
GROUPBY=PARAMETER
GROUPBY=TYPE
specifies how the plots are grouped when there is more than one type of plot.
GROUPBY=TYPE
specifies that the plots be grouped by type.
GROUPBY=PARAMETER
specifies that the plots be grouped by parameter.
GROUPBY=PARAMETER is the default.
LAGS=n
specifies that autocorrelations be plotted up to lag n. If this option is not specified,
autocorrelations are plotted up to lag 50.
SMOOTH
displays a fitted penalized B-spline curve for each trace plot.
UNPACKPANEL
UNPACK
specifies that all paneled plots be unpacked, meaning that each plot in a panel is displayed
separately.
The plot requests include the following:
ALL
specifies all types of plots. PLOTS=ALL is equivalent to specifying PLOTS=(TRACE
AUTOCORR DENSITY).
BAYES Statement F 2471
AUTOCORR
displays the autocorrelation function plots for the parameters.
DENSITY
displays the kernel density plots for the parameters.
NONE
suppresses all diagnostic plots.
TRACE
displays the trace plots for the parameters. See the section “Visual Analysis via Trace
Plots” on page 155 for details.
SAMPLING=option
specifies an algorithm used to sample the posterior distribution. The fololowing options are
available:
ARMS
GIBBS
use the ARMS algorithm. This is the default method except for the normal distribution
with a conjugate prior. In this case a closed form for the posterior distribution is available,
and samples are obtained directly from the posterior distribution.
GAMERMAN
GAM
use the Gamerman algorithm.
IM
Use the independent Metropolis algorithm.
SCALEPRIOR=GAMMA< (options) > | IMPROPER
SPRIOR=GAMMA< (options) > | IMPROPER
specifies that Gibbs sampling be performed on the generalized linear model scale parameter
and the prior distribution for the scale parameter, if there is a scale parameter in the model. For
models that do not have a scale parameter (the Poisson and binomial), this option is ignored.
Note that you can specify Gibbs sampling on either the dispersion parameter , the scale
1
parameter D 2 , or the precision parameter D 1 , with the DPRIOR=, SPRIOR=, and
PPRIOR= options, respectively. These three parameters are transformations of one another,
and you should specify Gibbs sampling for only one of them.
a 1
bt
b.bt /
e
A gamma prior G.a; b/ with density f .t / D
is specified by
€.a/
SCALEPRIOR=GAMMA, which can be followed by one of the following gamma-options
enclosed in parentheses. The hyperparameters a and b are the shape and inverse-scale
parameters of the gamma distribution, respectively. See the section “Gamma Prior” on
page 2550 for details. The default is G.10 4 ; 10 4 /.
RELSHAPE< =c >
specifies independent G.c O ; c/ distribution, where O is the MLE of the dispersion
parameter. With this choice of hyperparameters, the mean of the prior distribution is O
and the variance is cO . By default, c D 10 4 .
2472 F Chapter 37: The GENMOD Procedure
SHAPE=a
ISCALE=b
when both specified, results in a G.a; b/ prior.
SHAPE=c
when specified alone, results in an G.c; c/ prior.
ISCALE=c
when specified alone, results in an G.c; c/ prior.
An improper prior with density f .t / proportional to t
SCALEPRIOR=IMPROPER.
1
is specified with
SEED=number
specifies an integer seed in the range 1 to 231 1 for the random number generator in the
simulation. Specifying a seed enables you to reproduce identical Markov chains for the same
specification. If the SEED= option is not specified, or if you specify a nonpositive seed, a
random seed is derived from the time of day.
STATISTICS < (global-options) > = ALL | NONE | keyword | (keyword-list)
STATS < (global-options) > = ALL | NONE | keyword | (keyword-list)
controls the number of posterior statistics produced. Specifying STATISTICS=ALL is equivalent to specifying STATISTICS= (SUMMARY INTERVAL COV CORR). If you do not
want any posterior statistics, you specify STATISTICS=NONE. The default is STATISTICS=(SUMMARY INTERVAL). See the section “Summary Statistics” on page 169 for
details. The global-options include the following:
ALPHA=numeric-list
controls the probabilities of the credible intervals. The ALPHA= values must be between
0 and 1. Each ALPHA= value produces a pair of 100(1–ALPHA)% equal-tail and HPD
intervals for each parameters. The default is the value of the ALPHA= option in the
MODEL statement, or 0.05 if that option is not specified (yielding the 95% credible
intervals for each parameter).
PERCENT=numeric-list
requests the percentile points of the posterior samples. The PERCENT= values must be
between 0 and 100. The default is PERCENT=25, 50, 75, which yield the 25th, 50th,
and 75th percentile points, respectively, for each parameter.
The list of keywords includes the following:
CORR
produces the posterior correlation matrix.
COV
produces the posterior covariance matrix.
SUMMARY
produces the means, standard deviations, and percentile points for the posterior samples.
The default is to produce the 25th, 50th, and 75th percentile points, but you can use the
global PERCENT= option to request specific percentile points.
BY Statement F 2473
INTERVAL
produces equal-tail credible intervals and HPD intervals. The default is to produce the
95% equal-tail credible intervals and 95% HPD intervals, but you can use the global
ALPHA= option to request intervals of any probabilities.
THINNING=number
THIN=number
controls the thinning of the Markov chain. Only one in every k samples is used when
THINNING=k, and if NBI=n0 and NMC=n, the number of samples kept is
n0
n0 C n
k
k
where [a] represents the integer part of the number a. The default is THINNING=1.
BY Statement
BY variables ;
You can specify a BY statement with PROC GENMOD to obtain separate analyses on 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 GENMOD
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).
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.
2474 F Chapter 37: The GENMOD Procedure
CLASS Statement
CLASS variable < (options) >: : :< variable < (options) > > < / options > ;
The CLASS statement names the classification variables to be used in the analysis. 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 the options 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. The following options are available:
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 sorting order of the classification variable. If both the DESCENDING and
ORDER= options are specified, PROC GENMOD 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.
ORDER=DATA | FORMATTED | FREQ | INTERNAL
specifies the sorting order for the levels of classification variables. This ordering determines
which parameters in the model correspond to each level in the data, so the ORDER= option
can be useful when you use the CONTRAST statement. By default, ORDER=FORMATTED.
For ORDER=FORMATTED and ORDER=INTERNAL, the sort order is machine-dependent.
When ORDER=FORMATTED is in effect for numeric variables for which you have supplied
no explicit format, the levels are ordered by their internal values.
The following table shows how PROC GENMOD interprets values of the ORDER= option.
CLASS Statement F 2475
Value of ORDER=
Levels Sorted By
DATA
FORMATTED
Order of appearance in the input data set
External formatted values, except for numeric
variables with no explicit format, which are sorted
by their unformatted (internal) values
Descending frequency count; levels with more
observations come earlier in the order
Unformatted value
FREQ
INTERNAL
For more information about sorting 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; Design matrix columns are created
from CLASS variables according to the corresponding coding schemes:
Value of PARAM=
Coding
EFFECT
Effect coding
GLM
Less-than-full-rank reference cell coding (this
keyword can be used only in a global option)
ORDINAL
THERMOMETER
Cumulative parameterization for an ordinal
CLASS variable
POLYNOMIAL
POLY
Polynomial coding
REFERENCE
REF
Reference cell coding
ORTHEFFECT
Orthogonalizes PARAM=EFFECT coding
ORTHORDINAL
ORTHOTHERM
Orthogonalizes PARAM=ORDINAL coding
ORTHPOLY
Orthogonalizes PARAM=POLYNOMIAL coding
ORTHREF
Orthogonalizes PARAM=REFERENCE coding
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.
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 414 of Chapter 19, “Shared
Concepts and Topics,” for further details.
2476 F Chapter 37: The GENMOD Procedure
REF=’level’ | keyword
specifies the reference level for PARAM=EFFECT, PARAM=REFERENCE, and their orthogonalizations. For an individual (but not a global) variable 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.
FIRST
designates the first ordered level as reference.
LAST
designates the last ordered level as reference.
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.
Class Variable Naming Convention
Parameter names for a CLASS predictor variable are constructed by concatenating the CLASS
variable name with the CLASS levels. However, for the POLYNOMIAL and orthogonal parameterizations, parameter names are formed by concatenating the CLASS variable name and keywords that
reflect the parameterization. See the section “Other Parameterizations” on page 414 in Chapter 19,
“Shared Concepts and Topics,” for examples and further details.
Class Variable Parameterization with Unbalanced Designs
PROC GENMOD initially parameterizes the CLASS variables by looking at the levels of the variables
across the complete data set. If you have an unbalanced replication of levels across variables or BY
groups, then the design matrix and the parameter interpretation might be different from what you
expect. For instance, suppose you have a model with one CLASS variable A with three levels (1, 2,
and 3), and another CLASS variable B with two levels (1 and 2). If the third level of A occurs only
with the first level of B, if you use the EFFECT parameterization, and if your model contains the
effect A(B) and an intercept, then the design for A within the second level of B is not a differential
effect. In particular, the design looks like the following:
B
A
1
1
1
2
2
1
2
3
1
2
Design Matrix
A(B=1)
A(B=2)
A1 A2
A1 A2
1
0
1
0
0
0
1
1
0
0
0
0
0
1
0
0
0
0
0
1
CONTRAST Statement F 2477
PROC GENMOD detects linear dependency among the last two design variables and sets the
parameter for A2(BD2) 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.
CONTRAST Statement
CONTRAST ’label’ contrast-specification < / options > ;
The CONTRAST statement provides a means of obtaining a test of a specified hypothesis concerning
the model parameters. This is accomplished by specifying a matrix L for testing the hypothesis
L0 ˇ D 0. You must be familiar with the details of the model parameterization that PROC GENMOD
uses. For more information, see the section “Parameterization Used in PROC GENMOD” on
page 2523 and the section “CLASS Statement” on page 2474. Computed statistics are based on the
asymptotic chi-square distribution of the likelihood ratio statistic, or the generalized score statistic
for GEE models, with degrees of freedom determined by the number of linearly independent rows in
the L0 matrix. You can request Wald chi-square statistics with the Wald option in the CONTRAST
statement.
There is no limit to the number of CONTRAST statements that you can specify, but they must
appear after the MODEL statement and after the ZEROMODEL statement for zero-inflated models.
Statistics for multiple CONTRAST statements are displayed in a single table.
The elements of the CONTRAST statement are as follows:
label
identifies the contrast on the output. A label is required for every contrast specified. Labels
can be up to 20 characters and must be enclosed in single quotes.
contrast-specification identifies the effects and their coefficients from which the L matrix is formed.
The contrast-specification can be specified in two different ways. The first method applies
to all models except the zero-inflated (ZI) distributions (zero-inflated Poisson and zeroinflated negative binomial), and the syntax is:
effect values < ,. . . effect values >
The second method of specifying a contrast applies only to ZI models, and the syntax is:
effect values < ,. . . effect values > @zero effect values < ,. . . effect values >
Specification of sets of effect values before the @zero separator results in a row of the L0
matrix with coefficients for effects in the regression part of the model set to values and
with the coefficients for the zero-inflation part of the model set to zero. Specification of
sets of effect values after the @zero separator results in a row of the L matrix with the
coefficients for the regression part of the model set to zero and with the coefficients of
effects in the zero-inflation part of the model set to values.
For example, the statements
CLASS A;
MODEL y=A;
CONTRAST 'Label1' A 1 -1;
2478 F Chapter 37: The GENMOD Procedure
specify an L0 matrix with one row with coefficients 1 for the first level of A and –1 for the
second level of A.
The statements
CLASS A B;
MODEL y=A / Dist=ZIP;
ZEROMODEL B;
CONTRAST 'Label2' A 1 -1 @ZERO B 1 -1;
specify an L0 matrix with two rows: the first row has coefficients 1 for the first level of A,
–1 for the second level of A, and zeros for all levels of B; the second row has coefficients 0
for all levels of A, 1 for the first level of B, and –1 for the second level of B.
effect
identifies an effect that appears in the MODEL statement. The value INTERCEPT or
intercept can be used as an effect when an intercept is 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 vector associated with the effect.
The rows of L0 are specified in order and are separated by commas.
If you use the default less-than-full-rank PROC GLM CLASS variable parameterization, each row of
the L0 matrix is checked for estimability. If PROC GENMOD finds a contrast to be nonestimable, it
displays missing values in corresponding rows in the results. See Searle (1971) for a discussion of
estimable functions. If the elements of L0 are not specified for an effect that contains a specified effect,
then the elements of the specified effect are distributed over the levels of the higher-order effect just
as the GLM procedure does for its CONTRAST and ESTIMATE statements. For example, suppose
that the model contains effects A and B and their interaction A*B. If you specify a CONTRAST
statement involving A alone, the L0 matrix contains nonzero terms for both A and A*B, since A*B
contains A.
When you use any of the full-rank PARAM= CLASS variable options, all parameters are directly
estimable, and rows of L0 are not checked for estimability.
If an effect is not specified in the CONTRAST statement, all of its coefficients in the L0 matrix are
set to 0. 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.
PROC GENMOD handles missing level combinations of classification variables in the same manner
as the GLM and MIXED procedures. Parameters corresponding to missing level combinations are
not included in the model. This convention can affect the way in which you specify the L matrix in
your CONTRAST statement.
If you specify the WALD option, the test of hypothesis is based on a Wald chi-square statistic. If
you omit the WALD option, the test statistic computed depends on whether an ordinary generalized
linear model or a GEE-type model is specified.
For an ordinary generalized linear model, the CONTRAST statement computes the likelihood
ratio statistic. This is defined to be twice the difference between the log likelihood of the model
unconstrained by the contrast and the log likelihood with the model fitted under the constraint that
the linear function of the parameters defined by the contrast is equal to 0. A p-value is computed
based on the asymptotic chi-square distribution of the chi-square statistic.
DEVIANCE Statement F 2479
If you specify a GEE model with the REPEATED statement, the test is based on a score statistic.
The GEE model is fit under the constraint that the linear function of the parameters defined by the
contrast is equal to 0. The score chi-square statistic is computed based on the generalized score
function. See the section “Generalized Score Statistics” on page 2540 for more information.
The degrees of freedom is the number of linearly independent constraints implied by the CONTRAST
statement—that is, the rank of L.
You can specify the following options after a slash (/).
E
requests that the L matrix be displayed.
SINGULAR=number
EPSILON=number
tunes the estimability checking. If v is a vector, define ABS(v) to be the absolute value of
the element of v with the largest absolute value. Let K0 be any row in the contrast matrix
L. Define C to be equal to ABS.K0 / if ABS.K0 / is greater than 0; otherwise, C equals 1. If
ABS.K0 K0 T/ is greater than Cnumber, then K is declared nonestimable. T is the Hermite
form matrix .X0 X/ .X0 X/, and .X0 X/ represents a generalized inverse of the matrix X0 X.
The value for number must be between 0 and 1; the default value is 1E 4. The SINGULAR=
option in the MODEL statement affects the computation of the generalized inverse of the
matrix X0 X. It might also be necessary to adjust this value for some data.
WALD
requests that a Wald chi-square statistic be computed for the contrast rather than the default
likelihood ratio or score statistic. The Wald statistic for testing L0 ˇ D 0 is defined by
O 0 .L0 †L/ .L0 ˇ/
O
S D .L0 ˇ/
where ˇO is the maximum likelihood estimate and † is its estimated covariance matrix. The
asymptotic distribution of S is 2r , where r is the rank of L. Computed p-values are based on
this distribution.
If you specify a GEE model with the REPEATED statement, † is the empirical covariance
matrix estimate.
DEVIANCE Statement
DEVIANCE variable = expression ;
You can specify a probability distribution other than those available in PROC GENMOD by using
the DEVIANCE and VARIANCE statements. You do not need to specify the DEVIANCE or
VARIANCE statement if you use the DIST= MODEL statement option to specify a probability
distribution. The variable identifies the deviance contribution from a single observation to the
procedure, and it must be a valid SAS variable name that does not appear in the input data set. The
expression can be any arithmetic expression supported by the DATA step language, and it is used
2480 F Chapter 37: The GENMOD Procedure
to define the functional dependence of the deviance on the mean and the response. You use the
automatic variables _MEAN_ and _RESP_ to represent the mean and response in the expression.
Alternatively, the deviance function can be defined using programming statements (see the section
“Programming Statements” on page 2502) and assigned to a variable, which is then listed as the
expression. This form is convenient for using complex statements such as IF-THEN/ELSE clauses.
The DEVIANCE statement is ignored unless the VARIANCE statement is also specified.
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 37.2 describes the available plot-types and their plotdefinition-options.
Table 37.2 Plot-Types and Plot-Definition-Options
Description
Plot-Definition-Options
BOX plot-type
Displays a box plot of continuous response data at each
level of a CLASS effect, with predicted values
superimposed and connected by a line. This is an
alternative to the INTERACTION plot-type.
PLOTBY= variable or CLASS effect
X= CLASS variable or effect
CONTOUR plot-type
Displays a contour plot of predicted values against two
continuous covariates.
PLOTBY= variable or CLASS effect
X= continuous variable
Y= continuous variable
FIT plot-type
Displays a curve of predicted values versus a
continuous variable.
PLOTBY= variable or CLASS effect
X= continuous variable
INTERACTION plot-type
Displays a plot of predicted values (possibly with error
bars) versus the levels of a CLASS effect. The
predicted values are connected with lines and can be
grouped by the levels of another CLASS effect.
PLOTBY= variable or CLASS effect
SLICEBY= variable or CLASS effect
X= CLASS variable or effect
SLICEFIT plot-type
Displays a curve of predicted values versus a
continuous variable grouped by the levels of a
CLASS effect.
PLOTBY= variable or CLASS effect
SLICEBY= variable or CLASS effect
X= continuous variable
For full details about the syntax and options of the EFFECTPLOT statement, see the section
“EFFECTPLOT Statement” on page 436 of Chapter 19, “Shared Concepts and Topics.”
ESTIMATE Statement F 2481
ESTIMATE Statement
ESTIMATE ’label’ effect values < ,. . . effect values > < /options > ;
The ESTIMATE statement is similar to a CONTRAST statement, except only one-row L0 matrices
are permitted.
In the case of zero-inflated (ZI) models, the statement syntax is:
ESTIMATE ’label’ effect values < ,. . . effect values > @zero effect values < ,. . . effect values >
< /options > ;
where sets of effects values before the @zero separator correspond to the regression part of the model,
and effects values after the @zero separator correspond to the zero-inflation part of the model. In
the case of ZI models, a one-row L0 matrix is created for the regression part of the model, another
one-row L0 matrix is created for the zero-inflation part of the model, and separate estimates for the
two L matrices are computed and displayed.
If you use the default less-than-full-rank GLM CLASS variable parameterization, each row is
checked for estimability. If PROC GENMOD finds a contrast to be nonestimable, it displays missing
values in corresponding rows in the results. See Searle (1971) for a discussion of estimable functions.
The actual estimate, L0 , its approximate standard error, and its confidence limits are displayed. A
Wald chi-square test that L0 ˇ = 0 is also displayed.
O where †
O
The approximate standard error of the estimate is computed as the square root of L0 †L,
is the estimated covariance matrix of the parameter estimates. If you specify a GEE model in the
O is the empirical covariance matrix estimate.
REPEATED statement, †
If you specify the EXP option, then exp.L0 ˇ/, its standard error, and its confidence limits are also
displayed.
The construction of the L vector and the checking for estimability for an ESTIMATE statement
follow the same rules as listed under the CONTRAST statement.
You can specify the following options in the ESTIMATE statement after a slash (/).
ALPHA=number
requests that a confidence interval be constructed with confidence level 1
of number must be between 0 and 1; the default value is 0.05.
number. The value
E
requests that the L matrix coefficients be displayed.
EXP
requests that exp.L0 ˇ/, its standard error, and its confidence limits be computed. If you specify
the EXP option, standard errors and confidence intervals are computed using the delta method.
SINGULAR=number
EPSILON=number
tunes the estimability checking as described for the CONTRAST statement.
2482 F Chapter 37: The GENMOD Procedure
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 or a stratified exact
Poisson 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).
The ASSESSMENT, BAYES, CONTRAST, EFFECTPLOT, ESTIMATE, LSMEANS, LSMESTIMATE, OUTPUT, SLICE, and STORE statements are not available with an exact analysis. Exact analyses are not performed when you specify a WEIGHT statement, or a model other than
LINK=LOGIT with DIST=BIN or LINK=LOG with DIST=POISSON. 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 MODEL 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 Logistic and Poisson Regression” on page 2553 for details.
ESTIMATE < =keyword >
estimates the individual parameters (conditioned on all other parameters) for the effects
specified in the EXACT statement. For each parameter, a point estimate, a standard error, a
confidence interval, and a p-value for a two-sided test that the parameter is zero are displayed.
Note that the two-sided p-value is twice the one-sided p-value. You can optionally specify one
of the following keywords:
PARM
specifies that the parameters be estimated. This is the default.
EXACT Statement F 2483
ODDS
specifies that the odds ratios be estimated. If you have classification variables, then
you must also specify the PARAM=REF option in the CLASS statement.
BOTH
specifies that both the parameters and odds ratios be estimated.
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 parameters 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 j .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 Logistic and
Poisson Regression” on page 2553 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 Logistic and Poisson Regression” on page 2553 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 2554
for more information.
EXACT Statement Examples
In the following example, two exact tests are computed: one for x1 and the other for x2. The test for
x1 is based on the exact conditional distribution of the sufficient statistic for the x1 parameter given
the observed 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 genmod;
2484 F Chapter 37: The GENMOD Procedure
model y= x1 x2 x3/d=b;
exact x1 x2;
run;
PROC GENMOD 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:
exact 'One' x1 / estimate=parm;
exact 'Two' x1 / estimate=parm onesided;
For each EXACT statement, individual tests for the parameters of the specified effects are computed
unless the JOINTONLY option is specified. Consider the following EXACT statements:
exact
exact
exact
exact
'E12'
'E1'
'E2'
'J12'
x1 x2 / estimate;
x1
/ estimate;
x2
/ estimate;
x1 x2 / joint;
In the E12 statement, the parameters for x1 and x2 are estimated and tested separately. Specifying the
E12 statement is equivalent to specifying both the E1 and E2 statements. In the J12 statement, the
joint test for the parameters of x1 and x2 is computed in addition to the individual tests for x1 and x2.
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;
optimizations 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 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.
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
EXACTOPTIONS Statement F 2485
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
controls how the partial sums
default, value=1E–8.
Pj
i D1 yi xi
are compared. value must be between 0 and 1; by
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 .
By default, FCONV=1E–8. You can also specify the ABSFCONV= and XCONV= criteria; if
more than one criterion is specified, then optimizations are terminated as soon as one criterion
is satisfied.
MAXTIME=seconds
specifies the maximum clock time (in seconds) that PROC GENMOD can use to calculate the
exact distributions. If the limit is exceeded, the procedure halts all computations and prints a
note to the 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.
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.
NETWORKMC invokes the hybrid network and Monte Carlo algorithm of Mehta, Patel,
and Senchaudhuri (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.
2486 F Chapter 37: The GENMOD Procedure
N=n
specifies the number of Monte Carlo samples to take when the METHOD=NETWORKMC
option is specified. By default, nD 10; 000. If the procedure cannot obtain n samples due to
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 X 2 be continuous variables, denote their joint distribution by f .X1; X 2/,
and let f .X1jX 2 D x2/ denote the marginal distribution of X1 conditioned on the observed
value of X 2. If you request the JOINT test of X1 and X 2, then n samples are used to
generate the estimate fO.X1; X 2/ of f .X1; X 2/, from which the test is computed. However,
the parameter estimate for X1 is computed from the subset of fO.X1; X 2/ that has X 2 D 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:
Remove the JOINT option from the EXACT statement.
Create dummy variables in a DATA step to represent the levels of a CLASS variable, and
specify them as independent variables in the MODEL statement.
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.
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 the METHOD=NETWORKMC option is specified. 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
zero, then PROC GENMOD 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 the
METHOD=NETWORKMC option is specified. The number of samples 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.
FREQ Statement F 2487
STATUSTIME=seconds
specifies the time interval (in seconds) for printing a status line in the LOG. You can use this
status line to track the progress of the computation of the exact conditional distributions. The
time interval 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
.i /
ıj
D
8 .i /
< ˇj
.i/
ˇj
:
.i 1/
ˇj
.i 1/
ˇj
.i 1/
ˇj
.i 1/
jˇj
j < 0:01
otherwise
.i /
and ˇj is the estimate of the j th parameter at iteration i .
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 ;
FREQUENCY variable ;
The variable in the FREQ statement identifies a variable in the input data set containing the frequency
of occurrence of each observation. PROC GENMOD 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 it is less than 1 or missing, the observation is not used. In
the case of models fit with generalized estimating equations (GEEs), the frequencies apply to the
subject/cluster and therefore must be the same for all observations within each subject.
FWDLINK Statement
FWDLINK variable = expression ;
You can define a link function other than a built-in link function by using the FWDLINK statement.
If you use the MODEL statement option LINK= to specify a link function, you do not need to use
the FWDLINK statement. The variable identifies the link function to the procedure. The expression
2488 F Chapter 37: The GENMOD Procedure
can be any arithmetic expression supported by the DATA step language, and it is used to define the
functional dependence on the mean.
Alternatively, the link function can be defined by using programming statements (see the section
“Programming Statements” on page 2502) and assigned to a variable, which is then listed as the
expression. The second form is convenient for using complex statements such as IF-THEN/ELSE
clauses. The GENMOD procedure automatically computes derivatives of the link function required
for iterative fitting. You must specify the inverse of the link function in the INVLINK statement when
you specify the FWDLINK statement to define the link function. You use the automatic variable
_MEAN_ to represent the mean in the preceding expression.
INVLINK Statement
INVLINK variable = expression ;
If you define a link function in the FWDLINK statement, then you must define the inverse link
function by using the INVLINK statement. If you use the MODEL statement option LINK= to
specify a link function, you do not need to use the INVLINK statement. The variable identifies the
inverse link function to the procedure. The expression can be any arithmetic expression supported by
the DATA step language, and it is used to define the functional dependence on the linear predictor.
Alternatively, the inverse link function can be defined using programming statements (see the section
“Programming Statements” on page 2502) and assigned to a variable, which is then listed as the
expression. The second form is convenient for using complex statements such as IF-THEN/ELSE
clauses. The automatic variable _XBETA_ represents the linear predictor in the preceding expression.
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 37.3 summarizes important options in the LSMEANS statement. If you specify the BAYES
statement, the ADJUST=, STEPDOWN, and LINES options are ignored. The PLOTS= option is not
available for a maximum likelihood analysis; it is available only for a Bayesian analysis.
If you specify a zero-inflated model (that is, a model for either the zero-inflated Poisson or the
zero-inflated negative binomial distribution), then the least squares means are computed only for
effects in the model for the distribution mean, and not for effects in the zero-inflation probability part
of the model.
LSMESTIMATE Statement F 2489
Table 37.3 Important LSMEANS Statement Options
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
Degrees of Freedom and p-values
ADJUST=
Determines the method for multiple comparison adjustment of LSmeans differences
ALPHA=˛
Determines the confidence level (1 ˛)
STEPDOWN
Adjusts multiple comparison p-values further in a step-down
fashion
Statistical Output
CL
CORR
COV
E
LINES
MEANS
PLOTS=
SEED=
Constructs confidence limits for means and mean differences
Displays the correlation matrix of LS-means
Displays the covariance matrix of LS-means
Prints the L matrix
Produces a “Lines” display for pairwise LS-means differences
Prints the LS-means
Requests ODS statistical graphics of means and mean comparisons
Specifies the seed for computations that depend on random numbers
Generalized Linear Modeling
Exponentiates and displays estimates of LS-means or LS-means
EXP
differences
Computes and displays estimates and standard errors of LS-means
ILINK
(but not differences) on the inverse linked scale
ODDSRATIO
Reports (simple) differences of least squares means in terms of
odds ratios if permitted by the link function
For details about the syntax of the LSMEANS statement, see the section “LSMEANS Statement” on
page 479 of Chapter 19, “Shared Concepts and Topics.”
LSMESTIMATE Statement
LSMESTIMATE model-effect < ‘label’ > values < divisor =n >
< , . . . < ‘label’ > values < divisor =n > >
< / options > ;
2490 F Chapter 37: The GENMOD Procedure
The LSMESTIMATE statement provides a mechanism for obtaining custom hypothesis tests among
least squares means.
Table 37.4 summarizes important options in the LSMESTIMATE statement.
Table 37.4
Important LSMESTIMATE Statement Options
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
Degrees of Freedom and p-values
ADJUST=
Determines the method for multiple comparison adjustment of LSmeans differences
ALPHA=˛
Determines the confidence level (1 ˛)
LOWER
Performs one-sided, lower-tailed inference
STEPDOWN
Adjusts multiple comparison p-values further in a step-down fashion
TESTVALUE=
Specifies values under the null hypothesis for tests
UPPER
Performs one-sided, upper-tailed inference
Statistical Output
CL
CORR
COV
E
ELSM
JOINT
PLOTS=
SEED=
Constructs confidence limits for means and mean differences
Displays the correlation matrix of LS-means
Displays the covariance matrix of LS-means
Prints the L matrix
Prints the K matrix
Produces a joint F or chi-square test for the LS-means and LSmeans differences
Requests ODS statistical graphics of means and mean comparisons
Specifies the seed for computations that depend on random numbers
Generalized Linear Modeling
CATEGORY=
Specifies how to construct estimable functions with multinomial
data
EXP
Exponentiates and displays LS-means estimates
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 496 of Chapter 19, “Shared Concepts and Topics.”
MODEL Statement F 2491
MODEL Statement
MODEL response = < effects > < /options > ;
MODEL events/trials = < effects > < /options > ;
The MODEL statement specifies the response, or dependent variable, and the effects, or explanatory
variables. If you omit the explanatory variables, the procedure fits an intercept-only model. An
intercept term is included in the model by default. The intercept can be removed with the NOINT
option.
You can specify the response in the form of a single variable or in the form of a ratio of two variables
denoted events/trials. The first form is applicable to all responses. The second form is applicable
only to summarized binomial response data. When each observation in the input data set contains
the number of events (for example, successes) and the number of trials from a set of binomial trials,
use the events/trials syntax.
In the events/trials model syntax, you specify two variables that contain the event and trial counts.
These two variables are separated by a slash (/). The values of both events and (trials events) must
be nonnegative, and the value of the trials variable must be greater than 0 for an observation to be
valid. The variable events or trials can take noninteger values.
When each observation in the input data set contains a single trial from a binomial or multinomial
experiment, use the first form of the preceding MODEL statements. The response variable can be
numeric or character. The ordering of response levels is critical in these models. You can use the
RORDER= option in the PROC GENMOD statement to specify the response level ordering.
Responses for the Poisson distribution must be all nonnegative, but they can be noninteger values.
The effects in the MODEL statement consist of an explanatory variable or combination of variables.
Explanatory variables can be continuous or classification variables. Classification variables can
be character or numeric. Explanatory variables representing nominal, or classification, data must
be declared in a CLASS statement. Interactions between variables can also be included as effects.
Columns of the design matrix are automatically generated for classification variables and interactions.
The syntax for specification of effects is the same as for the GLM procedure. See the section
“Specification of Effects” on page 2522 for more information. Also refer to Chapter 39, “The GLM
Procedure.”
You can specify the following options in the MODEL statement after a slash (/).
AGGREGATE= (variable-list)
AGGREGATE= variable
AGGREGATE
specifies the subpopulations on which the Pearson chi-square and the deviance are calculated.
This option applies only to the multinomial distribution or the binomial distribution with
binary (single trial syntax) response. It is ignored if specified for other cases. Observations
with common values in the given list of variables are regarded as coming from the same
subpopulation. This affects the computation of the deviance and Pearson chi-square statistics.
Variables in the list can be any variables in the input data set. Specifying the AGGREGATE
2492 F Chapter 37: The GENMOD Procedure
option is equivalent to specifying the AGGREGATE= option with a variable list that includes
all explanatory variables in the MODEL statement. Pearson chi-square and deviance statistics
are not computed for multinomial models unless this option is specified.
ALPHA=number
ALPH=number
A=number
sets the confidence coefficient for parameter confidence intervals to 1 number. The value of
number must be between 0 and 1. The default value of number is 0.05.
CICONV=number
sets the convergence criterion for profile likelihood confidence intervals. See the section
“Confidence Intervals for Parameters” on page 2525 for the definition of convergence. The
value of number must be between 0 and 1. By default, CICONV=1E 4.
CL
requests that confidence limits for predicted values be displayed (see the OBSTATS option).
CODING=EFFECT
CODING=FULLRANK
specifies that effect coding be used for all classification variables in the model. This is the
same as specifying PARAM=EFFECT as a CLASS statement option.
CONVERGE=number
sets the convergence criterion. The value of number must be between 0 and 1. The iterations
are considered to have converged when the maximum change in the parameter estimates
between iteration steps is less than the value specified. The change is a relative change if the
parameter is greater than 0.01 in absolute value; otherwise, it is an absolute change. By default,
CONVERGE=1E 4. This convergence criterion is used in parameter estimation for a single
model fit, Type 1 statistics, and likelihood ratio statistics for Type 3 analyses and CONTRAST
statements.
CONVH=number
sets the relative Hessian convergence criterion. The value of number must be between 0 and
1. After convergence is determined with the change in parameter criterion specified with
0
1
the CONVERGE= option, the quantity t c D g Hjf j g is computed and compared to number,
where g is the gradient vector, H is the Hessian matrix for the model parameters, and f is
the log-likelihood function. If t c is greater than number, a warning that the relative Hessian
convergence criterion has been exceeded is printed. This criterion detects the occasional
case where the change in parameter convergence criterion is satisfied, but a maximum in the
log-likelihood function has not been attained. By default, CONVH=1E 4.
CORRB
requests that the parameter estimate correlation matrix be displayed.
COVB
requests that the parameter estimate covariance matrix be displayed.
MODEL Statement F 2493
DIAGNOSTICS
INFLUENCE
requests that case deletion diagnostic statistics be displayed (see the OBSTATS option).
DIST=keyword
D=keyword
ERROR=keyword
ERR=keyword
specifies the built-in probability distribution to use in the model. If you specify the DIST=
option and you omit a user-defined link function, a default link function is chosen as displayed
in the following table. If you specify no distribution and no link function, then the GENMOD
procedure defaults to the normal distribution with the identity link function.
DIST=
Distribution
Default Link Function
BINOMIAL | BIN | B
GAMMA | GAM | G
GEOMETRIC | GEOM
IGAUSSIAN | IG
MULTINOMIAL | MULT
NEGBIN | NB
NORMAL | NOR | N
POISSON | POI | P
ZIP
ZINB
Binomial
Gamma
Geometric
Inverse Gaussian
Multinomial
Negative binomial
Normal
Poisson
Zero-inflated Poisson
Zero-inflated negative binomial
Logit
Inverse ( power( 1) )
Log
Inverse squared ( power( 2) )
Cumulative logit
Log
Identity
Log
Log/logit
Log/logit
EXPECTED
requests that the expected Fisher information matrix be used to compute parameter estimate
covariances and the associated statistics. The default action is to use the observed Fisher
information matrix. This option does not affect the model fitting, only the way in which the
covariance matrix is computed (see the SCORING= option.)
ID=variable
causes the values of variable in the input data set to be displayed in the OBSTATS table. If
an explicit format for variable has been defined, the formatted values are displayed. If the
OBSTATS option is not specified, this option has no effect.
INITIAL=numbers
sets initial values for parameter estimates in the model. The default initial parameter values
are weighted least squares estimates based on using the response data as the initial mean
estimate. This option can be useful in case of convergence difficulty. The intercept parameter
is initialized with the INTERCEPT= option and is not included here. The values are assigned to
the variables in the MODEL statement in the same order in which they appear in the MODEL
statement. The order of levels for CLASS variables is determined by the ORDER= option.
Note that some levels of classification variables can be aliased; that is, they correspond to
linearly dependent parameters that are not estimated by the procedure. Initial values must be
assigned to all levels of classification variables, regardless of whether they are aliased or not.
The procedure ignores initial values corresponding to parameters not being estimated. If you
2494 F Chapter 37: The GENMOD Procedure
specify a BY statement, all classification variables must take on the same number of levels
in each BY group. Otherwise, classification variables in some of the BY groups are assigned
incorrect initial values. Types of INITIAL= specifications are illustrated in the following table.
Type of List
Specification
List separated by blanks
List separated by commas
x to y
x to y by z
Combination of list types
INITIAL D 3 4 5
INITIAL D 3, 4, 5
INITIAL D 3 to 5
INITIAL D 3 to 5 by 1
INITIAL D 1, 3 to 5, 9
INTERCEPT=number
INTERCEPT=number-list
initializes the intercept term to number for parameter estimation. If you specify both the
INTERCEPT= and the NOINT options, the intercept term is not estimated, but an intercept
term of number is included in the model. If you specify a multinomial model for ordinal data,
you can specify a number-list for the multiple intercepts in the model.
ITPRINT
displays the iteration history for all iterative processes: parameter estimation, fitting constrained
models for contrasts and Type 3 analyses, and profile likelihood confidence intervals. The
last evaluation of the gradient and the negative of the Hessian (second derivative) matrix are
also displayed for parameter estimation. If you perform a Bayesian analysis by specifying the
BAYES statement, the iteration history for computing the mode of the posterior distribution is
also displayed.
This option might result in a large amount of displayed output, especially if some of the
optional iterative processes are selected.
LINK=keyword
specifies the link function to use in the model. The keywords and their associated built-in link
functions are as follows.
LINK=
CUMCLL
CCLL
CUMLOGIT
CLOGIT
CUMPROBIT
CPROBIT
CLOGLOG
CLL
IDENTITY
ID
LOG
LOGIT
PROBIT
POWER(number) | POW(number)
Link Function
Cumulative complementary log-log
Cumulative logit
Cumulative probit
Complementary log-log
Identity
Log
Logit
Probit
Power with = number
MODEL Statement F 2495
If no LINK= option is supplied and there is a user-defined link function, the user-defined link
function is used. If you specify neither the LINK= option nor a user-defined link function,
then the default canonical link function is used if you specify the DIST= option. Otherwise, if
you omit the DIST= option, the identity link function is used.
The cumulative link functions are appropriate only for the multinomial distribution.
LRCI
requests that two-sided confidence intervals for all model parameters be computed based on
the profile likelihood function. This is sometimes called the partially maximized likelihood
function. See the section “Confidence Intervals for Parameters” on page 2525 for more
information about the profile likelihood function. This computation is iterative and can
consume a relatively large amount of CPU time. The confidence coefficient can be selected
with the ALPHA=number option. The resulting confidence coefficient is 1 number. The
default confidence coefficient is 0.95.
MAXITER=number
MAXIT=number
sets the maximum allowable number of iterations for all iterative computation processes in
PROC GENMOD. By default, MAXITER=50.
NOINT
requests that no intercept term be included in the model. An intercept is included unless this
option is specified.
NOSCALE
holds the scale parameter fixed. Otherwise, for the normal, inverse Gaussian, and gamma
distributions, the scale parameter is estimated by maximum likelihood. If you omit the
SCALE= option, the scale parameter is fixed at the value 1.
OBSTATS
specifies that an additional table of statistics be displayed. Formulas for the statistics are
given in the section “Predicted Values of the Mean” on page 2527, the section “Residuals” on
page 2528, and the section “Case Deletion Diagnostic Statistics” on page 2544. Residuals and
fit diagnostics are not computed for multinomial models.
For each observation, the following items are displayed:
the value of the response variable (variables if the data are binomial), frequency, and
weight variables
the values of the regression variables
predicted mean, O D g 1 ./, where D x0i ˇO is the linear predictor and g is the link
O
function. If there is an offset, it is included in x0i ˇ.
O If there is an offset, it is included in x0 ˇ.
O
estimate of the linear predictor x0i ˇ.
i
standard error of the linear predictor x0i ˇO
the value of the Hessian weight at the final iteration
2496 F Chapter 37: The GENMOD Procedure
lower confidence limit of the predicted value of the mean. The confidence coefficient is
specified with the ALPHA= option. See the section “Confidence Intervals on Predicted
Values” on page 2528 for the computational method.
upper confidence limit of the predicted value of the mean
raw residual, defined as Y
Pearson, or chi residual, defined as the square root of the contribution for the observation
to the Pearson chi-square—that is,
Y p
V ./=w
where Y is the response, is the predicted mean, w is the value of the prior weight
variable specified in a WEIGHT statement, and V() is the variance function evaluated
at .
the standardized Pearson residual
deviance residual, defined as the square root of the deviance contribution for the observation, with sign equal to the sign of the raw residual
the standardized deviance residual
the likelihood residual
a Cook distance type statistic for assessing the influence of individual observations on
overall model fit
observation leverage
O where
DFBETA, defined as an approximation to ˇO ˇOŒi  for each parameter estimate ˇ,
ˇOŒi  is the parameter estimate with the i th observation deleted
standardized DFBETA, defined as DFBETA, normalized by its standard deviation
zero inflation probability for zero-inflated models
the mean of a zero-inflated response
The following additional cluster deletion diagnostic statistics are created and displayed for
each cluster if a REPEATED statement is specified:
a Cook distance type statistic for assessing the influence of entire clusters on overall
model fit
a studentized Cook distance for assessing influence of clusters
cluster leverage
cluster DFBETA for assessing the influence of entire clusters on individual parameter
estimates
cluster DFBETA normalized by its standard deviation
If you specify the multinomial distribution, only regression variable values, response values,
predicted values, confidence limits for the predicted values, and the linear predictor are
displayed in the table. Residuals and other diagnostic statistics are not available for the
multinomial distribution.
MODEL Statement F 2497
The RESIDUALS, DIAGNOSTICS | INFLUENCE, PREDICTED, XVARS, and CL options
cause only subgroups of the observation statistics to be displayed. You can specify more than
one of these options to include different subgroups of statistics.
The ID=variable option causes the values of variable in the input data set to be displayed in the
table. If an explicit format for variable has been defined, the formatted values are displayed.
If a REPEATED statement is present, a table is displayed for the GEE model specified in the
REPEATED statement. Regression variables, response values, predicted values, confidence
limits for the predicted values, linear predictor, raw residuals, Pearson residuals for each
observation in the input data set are available. Case deletion diagnostic statistics are available
for each observation and for each cluster.
OFFSET=variable
specifies a variable in the input data set to be used as an offset variable. This variable cannot be
a CLASS variable, and it cannot be the response variable or one of the explanatory variables.
An OFFSET= variable is required when you perform an exact Poisson regression. Let oi
be the offset for the i th observation. Then exp.oi / should be a nonnegative integer which is
greater than or equal to the response value. If exp.oi / is not an integer, then the integer-part is
used. See the section “Exact Logistic and Poisson Regression” on page 2553 for information
about the use of the offset in the exact Poisson model.
PREDICTED
PRED
P
requests that predicted values, the linear predictor, its standard error, and the Hessian weight
be displayed (see the OBSTATS option).
RESIDUALS
R
requests that residuals and standardized residuals be displayed. Residuals and other diagnostic
statistics are not available for the multinomial distribution (see the OBSTATS option).
SCALE=number
SCALE=PEARSON
SCALE=P
PSCALE
SCALE=DEVIANCE
SCALE=D
DSCALE
sets the value used for the scale parameter where the NOSCALE option is used. For the
binomial and Poisson distributions, which have no free scale parameter, this can be used
to specify an overdispersed model. In this case, the parameter covariance matrix and the
likelihood function are adjusted by the scale parameter. See the section “Dispersion Parameter”
on page 2520 and the section “Overdispersion” on page 2521 for more information. If the
NOSCALE option is not specified, then number is used as an initial estimate of the scale
parameter.
2498 F Chapter 37: The GENMOD Procedure
Specifying SCALE=PEARSON or SCALE=P is the same as specifying the PSCALE option.
This fixes the scale parameter at the value 1 in the estimation procedure. After the parameter
estimates are determined, the exponential family dispersion parameter is assumed to be given
by Pearson’s chi-square statistic divided by the degrees of freedom, and all statistics such as
standard errors and likelihood ratio statistics are adjusted appropriately.
Specifying SCALE=DEVIANCE or SCALE=D is the same as specifying the DSCALE option.
This fixes the scale parameter at a value of 1 in the estimation procedure.
After the parameter estimates are determined, the exponential family dispersion parameter is
assumed to be given by the deviance divided by the degrees of freedom. All statistics such as
standard errors and likelihood ratio statistics are adjusted appropriately.
SCORING=number
requests that on iterations up to number, the Hessian matrix be computed using the Fisher
scoring method. For further iterations, the full Hessian matrix is computed. The default value
is 1. A value of 0 causes all iterations to use the full Hessian matrix, and a value greater than
or equal to the value of the MAXITER option causes all iterations to use Fisher scoring. The
value of the SCORING= option must be 0 or a positive integer.
SINGULAR=number
sets the tolerance for testing singularity of the information matrix and the crossproducts matrix.
Roughly, the test requires that a pivot be at least this number times the original diagonal value.
By default, number is 107 times the machine epsilon. The default number is approximately
10 9 on most machines. This value also controls the check on estimability for ESTIMATE
and CONTRAST statements.
TYPE1
requests that a Type 1, or sequential, analysis be performed. This consists of sequentially fitting
models, beginning with the null (intercept term only) model and continuing up to the model
specified in the MODEL statement. The likelihood ratio statistic between each successive pair
of models is computed and displayed in a table.
A Type 1 analysis is not available for GEE models, since there is no associated likelihood.
TYPE3
requests that statistics for Type 3 contrasts be computed for each effect specified in the MODEL
statement. The default analysis is to compute likelihood ratio statistics for the contrasts or
score statistics for GEEs. Wald statistics are computed if the WALD option is also specified.
WALD
requests Wald statistics for Type 3 contrasts. You must also specify the TYPE3 option in order
to compute Type 3 Wald statistics.
WALDCI
requests that two-sided Wald confidence intervals for all model parameters be computed
based on the asymptotic normality of the parameter estimators. This computation is not as
time-consuming as the LRCI method, since it does not involve an iterative procedure. However,
it is thought to be less accurate, especially for small sample sizes. The confidence coefficient
can be selected with the ALPHA= option in the same way as for the LRCI option.
OUTPUT Statement F 2499
XVARS
requests that the regression variables be included in the OBSTATS table.
OUTPUT Statement
OUTPUT < OUT=SAS-data-set > < keyword=name . . . keyword=name > ;
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 (XBETA) and their standard error estimates, the
weights for the Hessian matrix, predicted values of the mean, confidence limits for predicted values,
residuals, and case deletion diagnostics. Residuals and diagnostic statistics are not computed for
multinomial models.
You can also request these statistics with the OBSTATS, PREDICTED, RESIDUALS, DIAGNOSTICS | INFLUENCE, CL, or XVARS option in the MODEL statement. You can then create a SAS
data set containing them with ODS OUTPUT commands. You might prefer to specify the OUTPUT
statement for requesting these statistics since the following are true:
The OUTPUT statement produces no tabular output.
The OUTPUT statement creates a SAS data set more efficiently than ODS. This can be an
advantage for large data sets.
You can specify the individual statistics to be included in the SAS data set.
If you use the multinomial distribution with one of the cumulative link functions for ordinal data, the
data set also contains variables named _ORDER_ and _LEVEL_ that indicate the levels of the ordinal
response variable and the values of the variable in the input data set corresponding to the sorted
levels. These variables indicate that the predicted value for a given observation is the probability that
the response variable is as large as the value of the _LEVEL_ variable. Residuals and other diagnostic
statistics are not available for the multinomial distribution.
The estimated linear predictor, its standard error estimate, and the predicted values and their confidence intervals are computed for all observations in which the explanatory variables are all nonmissing, 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.
The following list explains specifications in the OUTPUT statement.
OUT=SAS-data-set
specifies the output data set. If you omit the OUT=option, the output data set is created and
given a default name that uses the DATAn convention.
keyword=name
specifies the statistics to be included in the output data set and names the new variables that
contain the statistics. Specify a keyword for each desired statistic (see the following list of
2500 F Chapter 37: The GENMOD Procedure
keywords), an equal sign, and the name of the new variable or variables to contain the statistic.
You can list only one variable after the equal sign for all the statistics, except for the case
deletion diagnostics for individual parameter estimates, DFBETA, DFBETAS, DFBETAC,
and DFBETACS. You can list variables enclosed in parentheses to correspond to the variables
in the model, or you can specify the keyword _all_, without parentheses, to include deletion
diagnostics for all of the parameters in the model.
Although you can use the OUTPUT statement without any keyword=name specifications,
the output data set then contains only the original variables and, possibly, the variables Level
and Value (if you use the multinomial model with ordinal data). Note that the residuals and
deletion diagnostics are not available for the multinomial model with ordinal data. Some of
the case deletion diagnostic statistics apply only to models for correlated data specified with a
REPEATED statement. If you request these statistics for ordinary generalized linear models,
the values of the corresponding variables are set to missing in the output data set. Formulas for
the statistics are given in the section “Predicted Values of the Mean” on page 2527, the section
“Residuals” on page 2528, and the section “Case Deletion Diagnostic Statistics” on page 2544.
The keywords allowed and the statistics they represent are as follows:
DFBETA | DBETA represents the effect of deleting an observation on parameter estimates.
If you specify the keyword _all_ after the equal sign, variables named
DFBETA_ParameterName will be included in the output data set to contain
the values of the diagnostic statistic to measure the influence of deleting a
single observation on the individual parameter estimates. ParameterName
is the name of the regression model parameter formed from the input
variable names concatenated with the appropriate levels, if classification
variables are involved.
DFBETAS | DBETAS represents the effect of deleting an observation on standardized parameter estimates. If you specify the keyword _all_ after the equal sign,
variables named DFBETAS_ParameterName will be included in the output
data set to contain the values of the diagnostic statistic to measure the influence of deleting a single observation on the individual parameter estimates.
ParameterName is the name of the regression model parameter formed
from the input variable names concatenated with the appropriate levels, if
classification variables are involved.
DOBS | COOKD | COOKSD represents the Cook distance type statistic to measure the
influence of deleting a single observation on the overall model fit.
HESSWGT
represents the diagonal element of the weight matrix used in computing the
Hessian matrix.
H | LEVERAGE
represents the leverage of a single observation.
LOWER | L
represents the lower confidence limit for the predicted value of the mean,
or the lower confidence limit for the probability that the response is less
than or equal to the value of Level or Value. The confidence coefficient is
determined by the ALPHA=number option in the MODEL statement as
.1 number/ 100%. The default confidence coefficient is 95%.
PREDICTED | PRED | PROB | P represents the predicted value of the mean of the response
or the predicted probability that the response variable is less than or equal
OUTPUT Statement F 2501
to the value of _LEVEL_ if the multinomial model for ordinal data is used
(in other words, Pr.Y _LEVEL_/, where Y is the response variable).
PZERO
represents the zero-inflation probability for zero-inflated models.
RESCHI
represents the Pearson (chi) residual for identifying observations that are
poorly accounted for by the model.
RESDEV
represents the deviance residual for identifying poorly fitted observations.
RESLIK
represents the likelihood residual for identifying poorly fitted observations.
RESRAW
represents the raw residual for identifying poorly fitted observations.
STDRESCHI
represents the standardized Pearson (chi) residual for identifying observations that are poorly accounted for by the model.
STDRESDEV
represents the standardized deviance residual for identifying poorly fitted
observations.
STDXBETA
represents the standard error estimate of XBETA (see the XBETA keyword).
UPPER | U
represents the upper confidence limit for the predicted value of the mean,
or the upper confidence limit for the probability that the response is less
than or equal to the value of Level or Value. The confidence coefficient is
determined by the ALPHA=number option in the MODEL statement as
.1 number/ 100%. The default confidence coefficient is 95%.
XBETA
represents the estimate of the linear predictor x0i ˇ for observation i , or
˛j C x0i ˇ, where j is the corresponding ordered value of the response
variable for the multinomial model with ordinal data. If there is an offset, it
is included in x0i ˇ.
The keywords in the following list apply only to models specified with a REPEATED statement,
fit by generalized estimating equations (GEEs).
CH | CLUSTERH | CLEVERAGE
CLUSTER
represents the leverage of a cluster.
represents the numerical cluster index, in order of sorted clusters.
DCLS | CLUSTERCOOKD | CLUSTERCOOKSD represents the Cook distance type statistic to measure the influence of deleting an entire cluster on the overall
model fit.
DFBETAC | DBETAC represents the effect of deleting an entire cluster on parameter estimates. If you specify the keyword _all_ after the equal sign, variables
named DFBETAC_ParameterName will be included in the output data set
to contain the values of the diagnostic statistic to measure the influence
of deleting the cluster on the individual parameter estimates. ParameterName is the name of the regression model parameter formed from the input
variable names concatenated with the appropriate levels, if classification
variables are involved.
DFBETACS | DBETACS represents the effect of deleting an entire cluster on normalized
parameter estimates. If you specify the keyword _all_ after the equal
sign, variables named DFBETACS_ParameterName will be included in the
2502 F Chapter 37: The GENMOD Procedure
output data set to contain the values of the diagnostic statistic to measure
the influence of deleting the cluster on the individual parameter estimates,
normalized by their standard errors. ParameterName is the name of the
regression model parameter formed from the input variable names concatenated with the appropriate levels, if classification variables are involved.
MCLS | CLUSTERDFIT represents the studentized Cook distance type statistic to measure
the influence of deleting an entire cluster on the overall model fit.
Programming Statements
Although the most commonly used link and probability distributions are available as built-in functions,
the GENMOD procedure enables you to define your own link functions and response probability
distributions by using the FWDLINK, INVLINK, VARIANCE, and DEVIANCE statements. The
variables assigned in these statements can have values computed in programming statements. These
programming statements can occur anywhere between the PROC GENMOD statement and the RUN
statement. Variable names used in programming statements must be unique. Variables from the input
data set can be referenced in programming statements. The mean, linear predictor, and response
are represented by the automatic variables _MEAN_, _XBETA_, and _RESP_, respectively, which
can be referenced in your programming statements. Programming statements are used to define the
functional dependencies of the link function, the inverse link function, the variance function, and the
deviance function on the mean, linear predictor, and response variable.
The following statements illustrate the use of programming statements. Even though you usually
request the Poisson distribution by specifying DIST=POISSON as a MODEL statement option, you
can define the variance and deviance functions for the Poisson distribution by using the VARIANCE
and DEVIANCE statements. For example, the following statements perform the same analysis as the
Poisson regression example in the section “Getting Started: GENMOD Procedure” on page 2435.
The statements must be in logical order for computation, just as in a DATA step.
proc genmod ;
class car age;
a = _MEAN_;
y = _RESP_;
d = 2 * ( y * log( y / a ) - ( y - a ) );
variance var = a;
deviance dev = d;
model c = car age / link = log offset = ln;
run;
The variables var and dev are dummy variables used internally by the procedure to identify the
variance and deviance functions. Any valid SAS variable names can be used.
Similarly, the log link function and its inverse could be defined with the FWDLINK and INVLINK
statements, as follows:
fwdlink link = log(_MEAN_);
invlink ilink = exp(_XBETA_);
REPEATED Statement F 2503
These statements are for illustration, and they work well for most Poisson regression problems. If,
however, in the iterative fitting process, the mean parameter becomes too close to 0, or a 0 response
value occurs, an error condition occurs when the procedure attempts to evaluate the log function.
You can circumvent this kind of problem by using IF-THEN/ELSE clauses or other conditional
statements to check for possible error conditions and appropriately define the functions for these
cases.
Data set variables can be referenced in user definitions of the link function and response distributions
by using programming statements and the FWDLINK, INVLINK, DEVIANCE, and VARIANCE
statements.
See the DEVIANCE, VARIANCE, FWDLINK, and INVLINK statements for more information.
REPEATED Statement
REPEATED SUBJECT= subject-effect < / options > ;
The REPEATED statement specifies the covariance structure of multivariate responses for GEE
model fitting in the GENMOD procedure. In addition, the REPEATED statement controls the
iterative fitting algorithm used in GEEs and specifies optional output. Other GENMOD procedure
statements, such as the MODEL and CLASS statements, are used in the same way as they are for
ordinary generalized linear models to specify the regression model for the mean of the responses.
SUBJECT=subject-effect
identifies subjects in the input data set. The subject-effect can be a single variable, an interaction
effect, a nested effect, or a combination. Each distinct value, or level, of the effect identifies a
different subject, or cluster. Responses from different subjects are assumed to be statistically
independent, and responses within subjects are assumed to be correlated. A subject-effect must
be specified, and variables used in defining the subject-effect must be listed in the CLASS
statement. The input data set does not need to be sorted by subject (see the SORTED option).
The options control how the model is fit and what output is produced. You can specify the
following options after a slash (/).
ALPHAINIT=numbers
specifies initial values for log odds ratio regression parameters if the LOGOR= option is
specified for binary data. If this option is not specified, an initial value of 0.01 is used for all
the parameters.
CONVERGE=number
specifies the convergence criterion for GEE parameter estimation. If the maximum absolute
difference between regression parameter estimates is less than the value of number on two
successive iterations, convergence is declared. If the absolute value of a regression parameter
estimate is greater than 0.08, then the absolute difference normalized by the regression
parameter value is used instead of the absolute difference. The default value of number is
0.0001.
2504 F Chapter 37: The GENMOD Procedure
CORRW
displays the estimated working correlation matrix. If you specify an exchangeable working
correlation structure with the CORR=EXCH option, the CORRW option is not needed to view
the estimated correlation, since a table is printed by default that contains the single estimated
correlation.
CORRB
displays the estimated regression parameter correlation matrix. Both model-based and empirical correlations are displayed.
COVB
displays the estimated regression parameter covariance matrix. Both model-based and empirical covariances are displayed.
ECORRB
displays the estimated regression parameter empirical correlation matrix.
ECOVB
displays the estimated regression parameter empirical covariance matrix.
INTERCEPT=number
specifies either an initial or a fixed value of the intercept regression parameter in the GEE
model. If you specify the NOINT option in the MODEL statement, then the intercept is fixed
at the value of number.
INITIAL=numbers
specifies initial values of the regression parameters estimation, other than the intercept parameter, for GEE estimation. If this option is not specified, the estimated regression parameters
assuming independence for all responses are used for the initial values.
LOGOR=log-odds-ratio-structure-keyword
specifies the regression structure of the log odds ratio used to model the association of the
responses from subjects for binary data. The response syntax must be of the single variable
type, the distribution must be binomial, and the data must be binary. Table 37.5 displays the
log odds ratio structure keywords and the corresponding log odds ratio regression structures.
See the section “Alternating Logistic Regressions” on page 2536 for definitions of the log
odds ratio types and examples of specifying log odds ratio models. You should specify either
the LOGOR= or the TYPE= option, but not both.
REPEATED Statement F 2505
Table 37.5 Log Odds Ratio Regression Structures
Keyword
Log Odds Ratio Regression Structure
EXCH
FULLCLUST
LOGORVAR(variable)
NESTK
NEST1
ZFULL
ZREP
Exchangeable
Fully parameterized clusters
Indicator variable for specifying block effects
k-nested
1-nested
Fully specified z matrix specified in ZDATA= data set
Single cluster specification for replicated z matrix specified
in ZDATA= data set
Single cluster specification for replicated z matrix
ZREP(matrix)
MAXITER=number
MAXIT=number
specifies the maximum number of iterations allowed in the iterative GEE estimation process.
The default number is 50.
MCORRB
displays the estimated regression parameter model-based correlation matrix.
MCOVB
displays the estimated regression parameter model-based covariance matrix.
MODELSE
displays an analysis of parameter estimates table that uses model-based standard errors for
inference. By default, an “Analysis of Parameter Estimates” table based on empirical standard
errors is displayed.
PRINTMLE
displays an analysis of maximum likelihood parameter estimates table. The maximum likelihood estimates are not displayed unless this option is specified.
RUPDATE=number
specifies the number of iterations between updates of the working correlation matrix. For
example, RUPDATE=5 specifies that the working correlation is updated once for every five
regression parameter updates. The default value of number is 1; that is, the working correlation
is updated every time the regression parameters are updated.
SORTED
specifies that the input data are grouped by subject and sorted within subject. If this option is
not specified, then the procedure internally sorts by subject-effect and within subject-effect, if
a within subject-effect is specified.
SUBCLUSTER=variable
SUBCLUST=variable
specifies a variable defining subclusters for the 1-nested or k-nested log odds ratio association
modeling structures. This variable must be listed in the CLASS statement.
2506 F Chapter 37: The GENMOD Procedure
TYPE=correlation-structure keyword
CORR=correlation-structure keyword
specifies the structure of the working correlation matrix used to model the correlation of
the responses from subjects. Table 37.6 displays the correlation structure keywords and the
corresponding correlation structures. The default working correlation type is the independent
(CORR=IND). See the section “Details: GENMOD Procedure” on page 2510 for definitions
of the correlation matrix types. You should specify LOGOR= or TYPE= but not both.
Table 37.6
Correlation Structure Types
Keyword
Correlation Matrix Type
AR
AR(1)
EXCH
CS
IND
MDEP(number)
UNSTR
UN
USER
FIXED (matrix)
Autoregressive(1)
Exchangeable
Independent
m-dependent with m=number
Unstructured
Fixed, user-specified correlation matrix
For example, you can specify a fixed 4 4 correlation matrix with the following option:
TYPE=USER( 1.0
0.9
0.8
0.6
0.9
1.0
0.9
0.8
0.8
0.9
1.0
0.9
0.6
0.8
0.9
1.0 )
V6CORR
specifies that the SAS ‘Version 6’ method of computing the normalized Pearson chi-square be
used for working correlation estimation and for model-based covariance matrix scale factor.
WITHINSUBJECT | WITHIN=within subject-effect
defines an effect specifying the order of measurements within subjects. Each distinct level of
the within subject-effect defines a different response from the same subject. If the data are in
proper order within each subject, you do not need to specify this option.
If some measurements do not appear in the data for some subjects, this option properly orders
the existing measurements and treats the omitted measurements as missing values. If the
WITHINSUBJECT= option is not used in this situation, measurements might be improperly
ordered and missing values assumed for the last measurements in a cluster.
Variables used in defining the within subject-effect must be listed in the CLASS statement.
YPAIR=variable-list
specifies the variables in the ZDATA= data set corresponding to pairs of responses for log odds
ratio association modeling.
SLICE Statement F 2507
ZDATA=SAS-data-set
specifies a SAS data set containing either the full z matrix for log odds ratio association
modeling or the z matrix for a single complete cluster to be replicated for all clusters.
ZROW=variable-list
specifies the variables in the ZDATA= data set corresponding to rows of the z matrix for log
odds ratio association modeling.
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
Table 19.19. For details about the syntax of the SLICE statement, see the section “SLICE Statement”
on page 526 of Chapter 19, “Shared Concepts and Topics.”
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 is 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 529 of Chapter 19, “Shared Concepts and Topics.”
STRATA Statement
STRATA variable < (option) > : : : < variable < (option) > > < / options > ;
The STRATA statement names the variables that define strata or matched sets to use in stratified
exact logistic regression of binary response data, or a stratified exact Poisson regression of count
data. An EXACT statement must also be specified.
Observations that have the same variable values are in the same matched set. For a stratified logistic
model, you can analyze 1W 1, 1W n, mW n, and general mi W ni matched sets where the number of
cases and controls varies across strata. For a stratified Poisson model, you can have any number of
2508 F Chapter 37: The GENMOD Procedure
observations in each stratum. 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
0
logit.hi / D ˛h C xhi
ˇ
where hi is the event probability for the i th 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 1W5 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. For an exact
Poisson regression, the “Strata Summary” table displays the number of strata that contain a specific
number of observations, which enables you to check whether every stratum in the analysis has the
same number of observations.
The ASSESSMENT, BAYES, CONTRAST, EFFECTPLOT, ESTIMATE, LSMEANS,
LSMESTIMATE, OUTPUT, SLICE, and STORE statements are not available with a STRATA
statement. Exact analyses are not performed when you specify a WEIGHT statement, a model other
than LINK=LOGIT with DIST=BIN or LINK=LOG with DIST=POISSON, or an offset variable.
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.
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.
COVARIATES
VARIANCE Statement F 2509
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, and the total frequency for each stratum. Since the
number of strata can be very large, this table is displayed only by request.
VARIANCE Statement
VARIANCE variable = expression ;
You can specify a probability distribution other than the built-in distributions by using the VARIANCE
and DEVIANCE statements. The variable name variable identifies the variance function to the
procedure. The expression is used to define the functional dependence on the mean, and it can be
any arithmetic expression supported by the DATA step language. You use the automatic variable
_MEAN_ to represent the mean in the expression.
Alternatively, you can define the variance function with programming statements, as detailed in
the section “Programming Statements” on page 2502. This form is convenient for using complex
statements such as IF-THEN/ELSE clauses. Derivatives of the variance function for use during
optimization are computed automatically. The DEVIANCE statement must also appear when the
VARIANCE statement is used to define the variance function.
WEIGHT Statement
WEIGHT | SCWGT variable ;
The WEIGHT statement identifies a variable in the input data set to be used as the exponential family
dispersion parameter weight for each observation. The exponential family dispersion parameter is
divided by the WEIGHT variable value for each observation. This is true regardless of whether the
parameter is estimated by the procedure or specified in the MODEL statement with the SCALE=
option. It is also true for distributions such as the Poisson and binomial that are not usually defined to
have a dispersion parameter. For these distributions, a WEIGHT variable weights the overdispersion
parameter, which has the default value of 1.
The WEIGHT variable does not have to be an integer; if it is less than or equal to 0 or if it is missing,
the corresponding observation is not used.
2510 F Chapter 37: The GENMOD Procedure
ZEROMODEL Statement
ZEROMODEL effects < /options > ;
The ZEROMODEL statement enables you to perform zero-inflated Poisson regression or zeroinflated negative binomial regression when those respective distributions are specified by the DIST=
option in the MODEL statement. The effects in the ZEROMODEL statement consist of explanatory
variables or combinations of variables for the zero-inflation probability regression model in a zeroinflated model. The same effects can be used in both the ZEROMODEL statement and the MODEL
statement, or effects can be used in one statement or the other separately. Explanatory variables
can be continuous or classification variables. Classification variables can be character or numeric.
Explanatory variables representing nominal, or classification, data must be declared in a CLASS
statement. Interactions between variables can also be included as effects. Columns of the design
matrix are automatically generated for classification variables and interactions. The syntax for
specification of effects is the same as for the GLM procedure. See the section “Specification of
Effects” on page 2522 for more information. Also refer to Chapter 39, “The GLM Procedure.”
You can specify the following option in the ZEROMODEL statement after a slash (/).
LINK=keyword
specifies the link function to use in the model. The keywords and their associated link functions
are as follows.
LINK=
Link Function
CLOGLOG
CLL
LOGIT
PROBIT
Complementary log-log
Logit
Probit
If no LINK= option is supplied, the LOGIT link is used. User-defined link functions are not allowed.
Details: GENMOD Procedure
Generalized Linear Models Theory
This is a brief introduction to the theory of generalized linear models.
Response Probability Distributions
In generalized linear models, the response is assumed to possess a probability distribution of the
exponential form. That is, the probability density of the response Y for continuous response variables,
Generalized Linear Models Theory F 2511
or the probability function for discrete responses, can be expressed as
y b. /
C c.y; /
f .y/ D exp
a./
for some functions a, b, and c that determine the specific distribution. For fixed , this is a oneparameter exponential family of distributions. The functions a and c are such that a./ D =w and
c D c.y; =w/, where w is a known weight for each observation. A variable representing w in the
input data set can be specified in the WEIGHT statement. If no WEIGHT statement is specified,
wi D 1 for all observations.
Standard theory for this type of distribution gives expressions for the mean and variance of Y :
E.Y / D b 0 . /
b 00 . /
Var.Y / D
w
where the primes denote derivatives with respect to . If represents the mean of Y; then the
variance expressed as a function of the mean is
Var.Y / D
V ./
w
where V is the variance function.
Probability distributions of the response Y in generalized linear models are usually parameterized in
terms of the mean and dispersion parameter instead of the natural parameter . The probability
distributions that are available in the GENMOD procedure are shown in the following list. The
zero-inflated Poisson and zero-inflated negative binomial distributions are not generalized linear
models. However, the zero-inflated distributions are included in PROC GENMOD since they are
useful extensions of generalized linear models. See Long (1997) for a discussion of the zero-inflated
Poisson and zero-inflated negative binomial distributions. The PROC GENMOD scale parameter
and the variance of Y are also shown.
Normal:
f .y/ D
p
1
2
D 2
scale D Var.Y / D 2
exp
1 y 2
2
for
1<y<1
2512 F Chapter 37: The GENMOD Procedure
Inverse Gaussian:
f .y/ D
"
1
p
exp
2y 3 1
2y
y
2 #
for 0 < y < 1
D 2
scale D Var.Y / D 2 3
Gamma:
f .y/ D
1
€./y
D y
exp
y
for 0 < y < 1
1
scale D 2
Var.Y / D
Geometric: This is a special case of the negative binomial with k D 1.
./y
for y D 0; 1; 2; : : :
.1 C /yC1
D 1
f .y/ D
Var.Y / D .1 C /
Negative binomial:
€.y C 1=k/
.k/y
for y D 0; 1; 2; : : :
€.y C 1/€.1=k/ .1 C k/yC1=k
dispersion D k
f .y/ D
Var.Y / D C k2
Poisson:
y e
yŠ
D 1
f .y/ D
Var.Y / D for y D 0; 1; 2; : : :
Generalized Linear Models Theory F 2513
Binomial:
f .y/ D
r .1
n
r
D 1
.1
Var.Y / D
/n
r
for y D
r
; r D 0; 1; 2; : : : ; n
n
/
n
Multinomial:
mŠ
y
y
y
p 1 p 2 pk k
y1 Šy2 Š yk Š 1 2
D 1
f .y1 ; y2 ; ; yk / D
Zero-inflated Poisson:
(
f .y/ D
! C .1 !/e for y D 0
ye for y D 1; 2; : : :
.1 !/ yŠ
D 1
D E.Y / D .1
!/
Var.Y / D .1
!/.1 C !/
!
D C
2
1 !
Zero-inflated negative binomial:
(
! C .1 !/.1 C k/ for y D 0
f .y/ D
.k/y
€.yC1=k/
for y D 1; 2; : : :
.1 !/ €.yC1/€.1=k/
.1Ck/yC1=k
D 1
dispersion D k
D E.Y / D .1
!/
Var.Y / D .1
!/.1 C ! C k/
k
!
D C
C
2
1 !
1 !
The negative binomial and the zero-inflated negative binomial distributions contain a parameter k,
called the negative binomial dispersion parameter. This is not the same as the generalized linear
model dispersion , but it is an additional distribution parameter that must be estimated or set to a
fixed value.
For the binomial distribution, the response is the binomial proportion Y D events=trials. The
variance function is V ./ D .1 /, and the binomial trials parameter n is regarded as a weight
w.
2514 F Chapter 37: The GENMOD Procedure
If a weight variable is present, is replaced with =w, where w is the weight variable.
PROC GENMOD works with a scale parameter that is related to the exponential family dispersion
parameter instead of working with itself. The scale parameters are related to the dispersion
parameter as shown previously with the probability distribution definitions. Thus, the scale parameter
output in the “Analysis of Parameter Estimates” table is related to the exponential family dispersion
parameter. If you specify a constant scale parameter with the SCALE= option in the MODEL
statement, it is also related to the exponential family dispersion parameter in the same way.
Link Function
For distributions other than the zero-inflated Poisson or zero-inflated negative binomial, the mean i
of the response in the ith observation is related to a linear predictor through a monotonic differentiable
link function g.
g.i / D x0i ˇ
Here, xi is a fixed known vector of explanatory variables, and ˇ is a vector of unknown parameters.
There are two link functions and linear predictors associated with zero-inflated distributions: one for
the zero inflation probability !, and another for the mean parameter . See the section “Zero-Inflated
Models” on page 2530 for more details about zero-inflated distributions.
Log-Likelihood Functions
Log-likelihood functions for the distributions that are available in the procedure are parameterized in
terms of the means i and the dispersion parameter . Zero-inflated log likelihoods are parameterized
in terms two parameters, and !. The parameter ! is the zero-inflation probability, and is a
function of the distribution mean. The relationship between the mean of the zero-inflated Poisson and
zero-inflated negative binomial distributions and the parameter is defined in the section “Response
Probability Distributions” on page 2510. The term yi represents the response for the i th observation,
and wi represents the known dispersion weight. The log-likelihood functions are of the form
X
L.y; ; / D
log .f .yi ; i ; //
i
where the sum is over the observations. The forms of the individual contributions
li D log .f .yi ; i ; //
are shown in the following list; the parameterizations are expressed in terms of the mean and
dispersion parameters.
For the discrete distributions (binomial, multinomial, negative binomial, and Poisson), the functions
computed as the sum of the li terms are not proper log-likelihood functions, since terms involving
binomial coefficients or factorials of the observed counts are dropped from the computation of the
log likelihood, and a dispersion parameter is included in the computation. Deletion of factorial
terms and inclusion of a dispersion parameter do not affect parameter estimates or their estimated
covariances for these distributions, and this is the function used in maximum likelihood estimation.
Generalized Linear Models Theory F 2515
The value of used in computing the reported log-likelihood function is either the final estimated
value, or the fixed value, if the dispersion parameter is fixed. Even though it is not a proper loglikelihood function in all cases, the function computed as the sum of the li terms is reported in the
output as the log likelihood. The proper log-likelihood function is also computed as the sum of the
l li terms in the following list, and it is reported as the full log likelihood in the output.
Normal:
1 wi .yi i /2
C log
C log.2/
2
wi
l li D li D
Inverse Gaussian:
1
2
l li D li D
"
yi3
wi .yi i /2
C
log
yi 2 wi
!
#
C log.2/
Gamma:
wi
wi yi
l li D li D
log
i
Negative binomial:
k
li D yi log
wi
k
l li D yi log
wi
wi yi
i
log.yi /
k
.yi C wi =k/ log 1 C
wi
wi
log €
k
.yi C wi =k/ log 1 C
wi
€.yi C wi =k/
C log
€.wi =k/
€.yi C wi =k/
C log
€.yi C 1/€.wi =k/
Poisson:
wi
Œyi log.i /
i 
l li D wi Œyi log.i /
i
log.yi Š/
wi
Œri log.pi / C .ni
ri / log.1
li D
Binomial:
li D
l li D wi Œlog
ni
ri
pi /
C ri log.pi / C .ni
ri / log.1
pi /
2516 F Chapter 37: The GENMOD Procedure
Multinomial (k categories):
li D
k
wi X
yij log.ij /
j D1
l li D wi Œlog.mi Š/ C
k
X
.yij log.ij /
log.yij Š//
j D1
Zero-inflated Poisson:
8
< wi logŒ!i C .1 !i / exp. i /
li D l li D
:
wi Œlog.1 !i / C yi log.i / i
yi D 0
log.yi Š/ yi > 0
Zero-inflated negative binomial:
8
logŒ!i C .1 !i /.1 C wki / yi D 0
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
< log.1 ! / C y log k
i
i
wi
li D l li D
ˆ .y C wi / log 1 C k
ˆ
i
ˆ
ˆ
wi
k
ˆ
w
ˆ
ˆ
€.yi C ki /
ˆ
: C log
yi > 0
wi
€.yi C1/€.
k
/
Maximum Likelihood Fitting
The GENMOD procedure uses a ridge-stabilized Newton-Raphson algorithm to maximize the loglikelihood function L.y; ; / with respect to the regression parameters. By default, the procedure
also produces maximum likelihood estimates of the scale parameter as defined in the section
“Response Probability Distributions” on page 2510 for the normal, inverse Gaussian, negative
binomial, and gamma distributions.
On the rth iteration, the algorithm updates the parameter vector ˇr with
ˇrC1 D ˇr
H
1
s
where H is the Hessian (second derivative) matrix, and s is the gradient (first derivative) vector of the
log-likelihood function, both evaluated at the current value of the parameter vector. That is,
@L
s D Œsj  D
@ˇj
and
@2 L
H D Œhij  D
@ˇi @ˇj
Generalized Linear Models Theory F 2517
In some cases, the scale parameter is estimated by maximum likelihood. In these cases, elements
corresponding to the scale parameter are computed and included in s and H.
If i D x0i ˇ is the linear predictor for observation i and g is the link function, then i D g.i /, so
that i D g 1 .x0i ˇ/ is an estimate of the mean of the ith observation, obtained from an estimate of
the parameter vector ˇ.
The gradient vector and Hessian matrix for the regression parameters are given by
s D
X wi .yi
i /xi
0
V .i /g .i /
i
X0 Wo X
H D
where X is the design matrix, xi is the transpose of the i th row of X, and V is the variance function.
The matrix Wo is diagonal with its i th diagonal element
woi D wei C wi .yi
i /
V .i /g 00 .i / C V 0 .i /g 0 .i /
.V .i //2 .g 0 .i //3 where
wei D
wi
V .i /.g 0 .i //2
The primes denote derivatives of g and V with respect to . The negative of H is called the observed
information matrix. The expected value of Wo is a diagonal matrix We with diagonal values wei . If
you replace Wo with We , then the negative of H is called the expected information matrix. We is the
weight matrix for the Fisher scoring method of fitting. Either Wo or We can be used in the update
equation. The GENMOD procedure uses Fisher scoring for iterations up to the number specified
by the SCORING option in the MODEL statement, and it uses the observed information matrix on
additional iterations.
Covariance and Correlation Matrix
The estimated covariance matrix of the parameter estimator is given by
†D
H
1
where H is the Hessian matrix evaluated using the parameter estimates on the last iteration. Note that
the dispersion parameter, whether estimated or specified, is incorporated into H. Rows and columns
corresponding to aliased parameters are not included in †.
The correlation matrix is the normalized covariance matrix. That is, if ij is an element of †, then
p
the corresponding element of the correlation matrix is ij =i j , where i D i i .
Goodness of Fit
Two statistics that are helpful in assessing the goodness of fit of a given generalized linear model are
the scaled deviance and Pearson’s chi-square statistic. For a fixed value of the dispersion parameter
2518 F Chapter 37: The GENMOD Procedure
, the scaled deviance is defined to be twice the difference between the maximum achievable log
likelihood and the log likelihood at the maximum likelihood estimates of the regression parameters.
Note that these statistics are not valid for GEE models.
If l.y; / is the log-likelihood function expressed as a function of the predicted mean values and
the vector y of response values, then the scaled deviance is defined by
D .y; / D 2.l.y; y/
l.y; //
For specific distributions, this can be expressed as
D .y; / D
D.y; /
where D is the deviance. The following table displays the deviance for each of the probability
distributions available in PROC GENMOD. The deviance cannot be directly calculated for zeroinflated models. Twice the negative of the log likelihood is reported instead of the proper deviance
for the zero-inflated Poisson and zero-inflated negative binomial.
Distribution
Inverse Gaussian
Deviance
P
i /2
i wi .yi
h
i
P
yi
2 i wi yi log .y
/
i
i
i
h
P
yi
1
2 i wi mi yi log C
.1
y
/
log
i
1
i
h
i
P
yi i
yi
C
2 i wi
log i
i
P wi .yi i /2
Multinomial
P P
Negative binomial
P h
2 i y log.y=/
Normal
Poisson
Binomial
Gamma
i
i
yi
i
i
yCwi =k
Cwi =k
i
2i yi
j
wi yij log
yij
pij mi
.y C wi =k/ log
Zero-inflated Poisson
8
ˆ
ˆ wi logŒ!i C .1 !i / exp. i / yi D 0
P <
2 i
w Œlog.1 !i / C yi log.i /
ˆ
ˆ
: i
i log.yi Š/
yi > 0
Zero-inflated negative binomial
8
logŒ!i C .1 !i /.1 C wki / yi D 0
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
P < log.1 !i / C yi log k
w
i
2 i
w
ˆ
k
i
ˆ
.y
C
/
log
1
C
C
i
ˆ
ˆ
k
wi
ˆ
wi
ˆ
ˆ
€.yi C k /
ˆ
: log
yi > 0
wi
€.yi C1/€.
k
/
In the binomial case, yi D ri =mi , where ri is a binomial count and mi is the binomial number of
trials parameter.
Generalized Linear Models Theory F 2519
In the multinomial case, yij refers to the observed number of occurrences of the j th category for
the i th subpopulation defined by the AGGREGATE= variable, mi is the total number in the ith
subpopulation, and pij is the category probability.
Pearson’s chi-square statistic is defined as
X2 D
i /2
V .i /
X wi .yi
i
and the scaled Pearson’s chi-square is X 2 =.
The scaled version of both of these statistics, under certain regularity conditions, has a limiting
chi-square distribution, with degrees of freedom equal to the number of observations minus the
number of parameters estimated. The scaled version can be used as an approximate guide to the
goodness of fit of a given model. Use caution before applying these statistics to ensure that all
the conditions for the asymptotic distributions hold. McCullagh and Nelder (1989) advise that
differences in deviances for nested models can be better approximated by chi-square distributions
than the deviances can themselves.
In cases where the dispersion parameter is not known, an estimate can be used to obtain an approximation to the scaled deviance and Pearson’s chi-square statistic. One strategy is to fit a model that
contains a sufficient number of parameters so that all systematic variation is removed, estimate from this model, and then use this estimate in computing the scaled deviance of submodels. The
deviance or Pearson’s chi-square divided by its degrees of freedom is sometimes used as an estimate
of the dispersion parameter . For example, since the limiting chi-square distribution of the scaled
deviance D D D= has n p degrees of freedom, where n is the number of observations and
p is the number of parameters, equating D to its mean and solving for yields O D D=.n p/.
Similarly, an estimate of based on Pearson’s chi-square X 2 is O D X 2 =.n p/. Alternatively, a
maximum likelihood estimate of can be computed by the procedure, if desired. See the discussion
in the section “Type 1 Analysis” on page 2523 for more about the estimation of the dispersion
parameter.
Other Fit Statistics
The Akaike information criterion (AIC) is a measure of goodness of model fit that balances model fit
against model simplicity. AIC has the form
AIC D
2LL C 2p
where p is the number of parameters estimated in the model, and LL is the log likelihood evaluated
at the value of the estimated parameters. An alternative form is the corrected AIC given by
AICC D
2LL C 2p
n
n
p
1
where n is the total number of observations used.
The Bayesian information criterion (BIC) is a similar measure. BIC is defined by
BIC D
2LL C p log.n/
2520 F Chapter 37: The GENMOD Procedure
See Akaike (1981, 1979) for details of AIC and BIC. See Simonoff (2003) for a discussion of using
AIC, AICC, and BIC with generalized linear models. These criteria are useful in selecting among
regression models, with smaller values representing better model fit. PROC GENMOD uses the
full log likelihoods defined in the section “Log-Likelihood Functions” on page 2514, with all terms
included, for computing all of the criteria.
Dispersion Parameter
There are several options available in PROC GENMOD for handling the exponential distribution
dispersion parameter. The NOSCALE and SCALE options in the MODEL statement affect the
way in which the dispersion parameter is treated. If you specify the SCALE=DEVIANCE option,
the dispersion parameter is estimated by the deviance divided by its degrees of freedom. If you
specify the SCALE=PEARSON option, the dispersion parameter is estimated by Pearson’s chi-square
statistic divided by its degrees of freedom.
Otherwise, values of the SCALE and NOSCALE options and the resultant actions are displayed in
the following table.
NOSCALE
SCALE=value
Action
Present
Present
Not present
Not present
Present
Not present
Not present
Present
Present (negative binomial)
Not present
Scale fixed at value
Scale fixed at 1
Scale estimated by ML
Scale estimated by ML,
starting point at value
k fixed at 0
The meaning of the scale parameter displayed in the “Analysis Of Parameter Estimates” table
is different for the gamma distribution than for the other distributions. The relation of the scale
parameter as used by PROC GENMOD to the exponential family dispersion parameter is displayed
in the following table. For the binomial and Poisson distributions, is the overdispersion parameter,
as defined in the “Overdispersion” section, which follows.
Distribution
Normal
Inverse Gaussian
Gamma
Binomial
Poisson
Scale
p
p
1=
p
p
In the case of the negative binomial distribution, PROC GENMOD reports the “dispersion” parameter
estimated by maximum likelihood. This is the negative binomial parameter k defined in the section
“Response Probability Distributions” on page 2510.
Generalized Linear Models Theory F 2521
Overdispersion
Overdispersion is a phenomenon that sometimes occurs in data that are modeled with the binomial
or Poisson distributions. If the estimate of dispersion after fitting, as measured by the deviance
or Pearson’s chi-square, divided by the degrees of freedom, is not near 1, then the data might be
overdispersed if the dispersion estimate is greater than 1 or underdispersed if the dispersion estimate
is less than 1. A simple way to model this situation is to allow the variance functions of these
distributions to have a multiplicative overdispersion factor :
Binomial: V ./ D .1
/
Poisson: V ./ D An alternative method to allow for overdispersion in the Poisson distribution is to fit a negative
binomial distribution, where V ./ D C k2 , instead of the Poisson. The parameter k can be
estimated by maximum likelihood, thus allowing for overdispersion of a specific form. This is
different from the multiplicative overdispersion factor , which can accommodate many forms of
overdispersion.
The models are fit in the usual way, and the parameter estimates are not affected by the value of . The
covariance matrix, however, is multiplied by , and the scaled deviance and log likelihoods used in
likelihood ratio tests are divided by . The profile likelihood function used in computing confidence
intervals is also divided by . If you specify a WEIGHT statement, is divided by the value of the
WEIGHT variable for each observation. This has the effect of multiplying the contributions of the
log-likelihood function, the gradient, and the Hessian by the value of the WEIGHT variable for each
observation.
p
The SCALE= option in the MODEL statement enables you to specify a value of D for the
binomial and Poisson distributions. If you specify the SCALE=DEVIANCE option in the MODEL
statement, the procedure uses the deviance divided by degrees of freedom as an estimate of , and
all statistics are adjusted appropriately. You can use Pearson’s chi-square instead of the deviance by
specifying the SCALE=PEARSON option.
The function obtained by dividing a log-likelihood function for the binomial or Poisson distribution by a dispersion parameter is not a legitimate log-likelihood function. It is an example of
a quasi-likelihood function. Most of the asymptotic theory for log likelihoods also applies to
quasi-likelihoods, which justifies computing standard errors and likelihood ratio statistics by using
quasi-likelihoods instead of proper log likelihoods. See McCullagh and Nelder (1989, Chapter 9),
McCullagh (1983), and Hardin and Hilbe (2003) for details on quasi-likelihood functions.
Although the estimate of the dispersion parameter is often used to indicate overdispersion or underdispersion, this estimate might also indicate other problems such as an incorrectly specified model or
outliers in the data. You should carefully assess whether this type of model is appropriate for your
data.
2522 F Chapter 37: The GENMOD Procedure
Specification of Effects
Each term in a model is called an effect. Effects are specified in the MODEL statement. You
specify effects with a special notation that uses variable names and operators. There are two types
of variables, classification (or CLASS) variables and continuous variables. There are two primary
types of operators, crossing and nesting. A third type, the bar operator, is used to simplify effect
specification. Crossing is the type of operator most commonly used in generalized linear models.
Variables that identify classification levels are called CLASS variables in SAS and are identified in a
CLASS statement. These might also be called categorical, qualitative, discrete, or nominal variables.
CLASS variables can be either character or numeric. The values of CLASS variables are called
levels. For example, the CLASS variable Sex could have the levels ‘male’ and ‘female’.
In a model, an explanatory variable that is not declared in a CLASS statement is assumed to be
continuous. Continuous variables must be numeric. For example, the heights and weights of subjects
in an experiment are continuous variables.
The types of effects most useful in generalized linear models are shown in the following list. Assume
that A, B, and C are classification variables and that X1 and X2 are continuous variables.
Regressor effects are specified by writing continuous variables by themselves: X1, X2.
Polynomial effects are specified by joining two or more continuous variables with asterisks:
X1*X2.
Main effects are specified by writing classification variables by themselves: A, B, C.
Crossed effects (interactions) are specified by joining two or more classification variables with
asterisks: A*B, B*C, A*B*C.
Nested effects are specified by following a main effect or crossed effect with a classification
variable or list of classification variables enclosed in parentheses: B(A), C(B A), A*B(C). In the
preceding example, B(A) is “B nested within A.”
Combinations of continuous and classification variables can be specified in the same way by
using the crossing and nesting operators.
The bar operator consists of two effects joined with a vertical bar (|). It is shorthand notation for
including the left-hand side, the right-hand side, and the cross between them as effects in the model.
For example, A | B is equivalent to A B A*B. The effects in the bar operator can be classification
variables, continuous variables, or combinations of effects defined using operators. Multiple bars are
permitted. For example, A | B | C means A B C A*B A*C B*C A*B*C.
You can specify the maximum number of variables in any effect that results from bar evaluation by
specifying the maximum number, preceded by an @ sign. For example, A | B | [email protected] results in effects
that involve two or fewer variables: A B C A*B A*C B*C.
Parameterization Used in PROC GENMOD F 2523
Parameterization Used in PROC GENMOD
Design Matrix
The linear predictor part of a generalized linear model is
D Xˇ
where ˇ is an unknown parameter vector and X is a known design matrix. By default, all models
automatically contain an intercept term; that is, the first column of X contains all 1s. Additional
columns of X are generated for classification variables, regression variables, and any interaction terms
included in the model. It is important to understand the ordering of classification variable parameters
when you use the ESTIMATE or CONTRAST statement. The ordering of these parameters is
displayed in the “CLASS Level Information” table and in tables displaying the parameter estimates
of the fitted model.
When you specify an overparameterized model with the PARAM=GLM option in the CLASS
statement, some columns of X can be linearly dependent on other columns. For example, when you
specify a model consisting of an intercept term and a classification variable, the column corresponding
to any one of the levels of the classification variable is linearly dependent on the other columns of
X. The columns of X0 X are checked in the order in which the model is specified for dependence on
preceding columns. If a dependency is found, the parameter corresponding to the dependent column
is set to 0 along with its standard error to indicate that it is not estimated. The order in which the
levels of a classification variable are checked for dependencies can be set by the ORDER= option in
the PROC GENMOD statement or by the ORDER= option in the CLASS statement. For full-rank
parameterizations, the columns of the X matrix are designed to be linearly independent.
You can exclude the intercept term from the model by specifying the NOINT option in the MODEL
statement.
Missing Level Combinations
All levels of interaction terms involving classification variables might not be represented in the data.
In that case, PROC GENMOD does not include parameters in the model for the missing levels.
Type 1 Analysis
A Type 1 analysis consists of fitting a sequence of models, beginning with a simple model with only
an intercept term, and continuing through a model of specified complexity, fitting one additional
effect on each step. Likelihood ratio statistics—that is, twice the difference of the log likelihoods—
are computed between successive models. This type of analysis is sometimes called an analysis of
deviance since, if the dispersion parameter is held fixed for all models, it is equivalent to computing
differences of scaled deviances. The asymptotic distribution of the likelihood ratio statistics, under
the hypothesis that the additional parameters included in the model are equal to 0, is a chi-square with
2524 F Chapter 37: The GENMOD Procedure
degrees of freedom equal to the difference in the number of parameters estimated in the successive
models. Thus, these statistics can be used in a test of hypothesis of the significance of each additional
term fit.
This type of analysis is not available for GEE models, since the deviance is not computed for this
type of model.
If the dispersion parameter is known, it can be included in the models; if it is unknown, there
are two strategies allowed by PROC GENMOD. The dispersion parameter can be estimated from a
maximal model by the deviance or Pearson’s chi-square divided by degrees of freedom, as discussed
in the section “Goodness of Fit” on page 2517, and this value can be used in all models. An alternative
is to consider the dispersion to be an additional unknown parameter for each model and estimate it
by maximum likelihood on each step. By default, PROC GENMOD estimates scale by maximum
likelihood at each step.
A table of likelihood ratio statistics is produced, along with associated p-values based on the
asymptotic chi-square distributions.
If you specify either the SCALE=DEVIANCE or the SCALE=PEARSON option in the MODEL
statement, the dispersion parameter is estimated using the deviance or Pearson’s chi-square statistic,
and F statistics are computed in addition to the chi-square statistics for assessing the significance
of each additional term in the Type 1 analysis. See the section “F Statistics” on page 2526 for a
definition of F statistics.
This Type 1 analysis has the general property that the results depend on the order in which the terms
of the model are fitted. The terms are fitted in the order in which they are specified in the MODEL
statement.
Type 3 Analysis
A Type 3 analysis is similar to the Type III sums of squares used in PROC GLM, except that
likelihood ratios are used instead of sums of squares. First, a Type III estimable function is defined
for an effect of interest in exactly the same way as in PROC GLM. Then maximum likelihood
estimates are computed under the constraint that the Type III function of the parameters is equal to 0,
by using constrained optimization. Let the resulting constrained parameter estimates be ˇQ and the
Q Then the likelihood ratio statistic
log likelihood be l.ˇ/.
O
S D 2.l.ˇ/
Q
l.ˇ//
where ˇO is the unconstrained estimate, has an asymptotic chi-square distribution under the hypothesis
that the Type III contrast is equal to 0, with degrees of freedom equal to the number of parameters
associated with the effect.
When a Type 3 analysis is requested, PROC GENMOD produces a table that contains the likelihood
ratio statistics, degrees of freedom, and p-values based on the limiting chi-square distributions for
each effect in the model. If you specify either the DSCALE or PSCALE option in the MODEL
statement, F statistics are also computed for each effect.
Confidence Intervals for Parameters F 2525
Options for handling the dispersion parameter are the same as for a Type 1 analysis. The dispersion
parameter can be specified to be a known value, estimated from the deviance or Pearson’s chisquare divided by degrees of freedom, or estimated by maximum likelihood individually for the
unconstrained and constrained models. By default, PROC GENMOD estimates scale by maximum
likelihood for each model fit.
The results of this type of analysis do not depend on the order in which the terms are specified in the
MODEL statement.
A Type 3 analysis can consume considerable computation time since a constrained model is fitted
for each effect. Wald statistics for Type 3 contrasts are computed if you specify the WALD option.
Wald statistics for contrasts use less computation time than likelihood ratio statistics but might be
less accurate indicators of the significance of the effect of interest. The Wald statistic for testing
L0 ˇ D 0, where L is the contrast matrix, is defined by
O 0 .L0 †L/
O
O
S D .L0 ˇ/
.L0 ˇ/
where ˇ is the maximum likelihood estimate and † is its estimated covariance matrix. The asymptotic
distribution of S is chi-square with r degrees of freedom, where r is the rank of L.
See Chapter 39, “The GLM Procedure,”and Chapter 15, “The Four Types of Estimable Functions,”
for more information about Type III estimable functions. Also refer to Littell, Freund, and Spector
(1991).
Generalized score tests for Type III contrasts are computed for GEE models if you specify the
TYPE3 option in the MODEL statement when a REPEATED statement is also used. See the section
“Generalized Score Statistics” on page 2540 for more information about generalized score statistics.
Wald tests are also available with the Wald option in the CONTRAST statement. In this case, the
robust covariance matrix estimate is used for † in the Wald statistic.
Confidence Intervals for Parameters
Likelihood Ratio-Based Confidence Intervals
PROC GENMOD produces likelihood ratio-based confidence intervals, also known as profile
likelihood confidence intervals, for parameter estimates for generalized linear models. These are
not computed for GEE models, since there is no likelihood for this type of model. Suppose that
the parameter vector is ˇ D Œˇ0 ; ˇ1 ; : : : ; ˇp 0 and that you want a confidence interval for ˇj . The
profile likelihood function for ˇj is defined as
l .ˇj / D max l.ˇ/
ˇQ
where ˇQ is the vector ˇ with the j th element fixed at ˇj and l is the log-likelihood function. If
O is the log likelihood evaluated at the maximum likelihood estimate ˇ,
O then 2.l l .ˇj //
l D l.ˇ/
has a limiting chi-square distribution with one degree of freedom if ˇj is the true parameter value. A
.1 ˛/100% confidence interval for ˇj is
˚
ˇj W l .ˇj / l0 D l 0:521 ˛;1
2526 F Chapter 37: The GENMOD Procedure
where 21 ˛;1 is the 100.1 ˛/th percentile of the chi-square distribution with one degree of freedom.
The endpoints of the confidence interval can be found by solving numerically for values of ˇj that
satisfy equality in the preceding relation. PROC GENMOD solves this by starting at the maximum
likelihood estimate of ˇ. The log-likelihood function is approximated with a quadratic surface,
for which an exact solution is possible. The process is iterated until convergence to an endpoint is
attained. The process is repeated for the other endpoint.
Convergence is controlled by the CICONV= option in the MODEL statement. Suppose is the
number specified in the CICONV= option. The default value of is 10 4 . Let the parameter
of interest be ˇj , and define r D uj , the unit vector with a 1 in position j and 0s elsewhere.
Convergence is declared on the current iteration if the following two conditions are satisfied:
jl .ˇj /
0
.s C r/ H
1
l0 j .s C r/ where l .ˇj /, s, and H are the log likelihood, the gradient, and the Hessian evaluated at the current
parameter vector and is a constant computed by the procedure. The first condition for convergence
means that the log-likelihood function must be within of the correct value, and the second condition
means that the gradient vector must be proportional to the restriction vector r.
When you specify the LRCI option in the MODEL statement, PROC GENMOD computes profile
likelihood confidence intervals for all parameters in the model, including the scale parameter, if
there is one. The interval endpoints are displayed in a table as well as the values of the remaining
parameters at the solution.
Wald Confidence Intervals
You can request that PROC GENMOD produce Wald confidence intervals for the parameters. The
(1 ˛)100% Wald confidence interval for a parameter ˇ is defined as
ˇO ˙ z1
O
˛=2 where zp is the 100pth percentile of the standard normal distribution, ˇO is the parameter estimate,
and O is the estimate of its standard error.
F Statistics
Suppose that D0 is the deviance resulting from fitting a generalized linear model and that D1 is the
deviance from fitting a submodel. Then, under appropriate regularity conditions, the asymptotic
distribution of .D1 D0 /= is chi-square with r degrees of freedom, where r is the difference
in the number of parameters between the two models and is the dispersion parameter. If is
unknown, and O is an estimate of based on the deviance or Pearson’s chi-square divided by degrees
O has an asymptotic chi-square distribution
of freedom, then, under regularity conditions, .n p/=
with n p degrees of freedom. Here, n is the number of observations and p is the number of
Lagrange Multiplier Statistics F 2527
parameters in the model that is used to estimate . Thus, the asymptotic distribution of
F D
D1
D0
r O
is the F distribution with r and n p degrees of freedom, assuming that .D1
O are approximately independent.
.n p/=
D0 /= and
This F statistic is computed for the Type 1 analysis, Type 3 analysis, and hypothesis tests specified
in CONTRAST statements when the dispersion parameter is estimated by either the deviance or
Pearson’s chi-square divided by degrees of freedom, as specified by the DSCALE or PSCALE option
in the MODEL statement. In the case of a Type 1 analysis, model 0 is the higher-order model
obtained by including one additional effect in model 1. For a Type 3 analysis and hypothesis tests,
model 0 is the full specified model and model 1 is the submodel obtained from constraining the Type
III contrast or the user-specified contrast to be 0.
Lagrange Multiplier Statistics
When you select the NOINT or NOSCALE option, restrictions are placed on the intercept or scale
parameters. Lagrange multiplier, or score, statistics are computed in these cases. These statistics
assess the validity of the restrictions, and they are computed as
2 D
s2
V
where s is the component of the score vector evaluated at the restricted maximum corresponding to
the restricted parameter and V D I11 I12 I221 I21 . The matrix I is the information matrix, 1 refers
to the restricted parameter, and 2 refers to the rest of the parameters.
Under regularity conditions, this statistic has an asymptotic chi-square distribution with one degree
of freedom, and p-values are computed based on this limiting distribution.
If you set k D 0 in a negative binomial model, s is the score statistic of Cameron and Trivedi (1998)
for testing for overdispersion in a Poisson model against alternatives of the form V ./ D C k2 .
See Rao (1973, p. 417) for more details.
Predicted Values of the Mean
Predicted Values
A predicted value, or fitted value, of the mean i corresponding to the vector of covariates xi is
given by
Oi D g
1
O
.x0i ˇ/
2528 F Chapter 37: The GENMOD Procedure
where g is the link function, regardless of whether xi corresponds to an observation or not. That is,
the response variable can be missing and the predicted value is still computed for valid xi . In the
case where xi does not correspond to a valid observation, xi is not checked for estimability. You
should check the estimability of xi in this case in order to ensure the uniqueness of the predicted
value of the mean. If there is an offset, it is included in the predicted value computation.
Confidence Intervals on Predicted Values
Approximate confidence intervals for predicted values of the mean can be computed as follows. The
variance of the linear predictor i D x0i ˇO is estimated by
x2 D x0i †xi
O The robust estimate of the covariance is used for † in the
where † is the estimated covariance of ˇ.
case of models fit with GEEs.
Approximate 100.1 ˛/% confidence intervals are computed as
g 1 x0i ˇO ˙ z1 ˛=2 x
where zp is the 100pth percentile of the standard normal distribution and g is the link function. If
either endpoint in the argument is outside the valid range of arguments for the inverse link function,
the corresponding confidence interval endpoint is set to missing.
Residuals
The GENMOD procedure computes three kinds of residuals. Residuals are available for all generalized linear models except multinomial models for ordinal response data, for which residuals
are not available. Raw residuals and Pearson residuals are available for models fit with generalized
estimating equations (GEEs).
The raw residual is defined as
ri D yi
i
where yi is the i th response and i is the corresponding predicted mean. You can request raw
residuals in an output data set with the keyword RESRAW in the OUTPUT statement.
The Pearson residual is the square root of the i th contribution to the Pearson’s chi-square:
r
wi
rP i D .yi i /
V .i /
You can request Pearson residuals in an output data set with the keyword RESCHI in the OUTPUT
statement.
Finally, the deviance residual is defined as the square root of the contribution of the i th observation
to the deviance, with the sign of the raw residual:
p
rDi D di .sign.yi i //
Multinomial Models F 2529
You can request deviance residuals in an output data set with the keyword RESDEV in the OUTPUT
statement.
The adjusted Pearson, deviance, and likelihood residuals are defined by Agresti (2002), Williams
(1987), and Davison and Snell (1991). These residuals are useful for outlier detection and for
assessing the influence of single observations on the fitted model.
For the generalized linear model, the variance of the i th individual observation is given by
vi D
V .i /
wi
where is the dispersion parameter, wi is a user-specified prior weight (if not specified, wi D 1),
i is the mean, and V .i / is the variance function. Let
wei D vi 1 .g 0 .i //
2
for the i th observation, where g 0 .i / is the derivative of the link function, evaluated at i . Let We
be the diagonal matrix with wei denoting the i th diagonal element. The weight matrix We is used in
computing the expected information matrix.
Define hi as the i th diagonal element of the matrix
1
We2 X.X0 We X/
1
1
X0 We2
The Pearson residuals, standardized to have unit asymptotic variance, are given by
rP i D p
yi
vi .1
i
hi /
You can request standardized Pearson residuals in an output data set with the keyword STDRESCHI
in the OUTPUT statement. The deviance residuals, standardized to have unit asymptotic variance,
are given by
p
sign.yi i / di
rDi D
p
.1 hi /
where di is the contribution to the total deviance from observation i , and sign.yi i / is 1 if yi i
is positive and 1 if yi i is negative. You can request standardized deviance residuals in an
output data set with the keyword STDRESDEV in the OUTPUT statement. The likelihood residuals
are defined by
q
2
rGi D sign.yi i / .1 hi /rDi
C hi rP2 i
You can request likelihood residuals in an output data set with the keyword RESLIK in the OUTPUT
statement.
Multinomial Models
This type of model applies to cases where an observation can fall into one of k categories. Binary
data occur in the special case where k D 2. If there are mi observations in a subpopulation i, then
2530 F Chapter 37: The GENMOD Procedure
the probability distribution of the number falling into the k categories yi D .yi1 ; yi 2 ; yi k / can be
modeled by the multinomial
distribution, defined in the section “Response Probability Distributions”
P
on page 2510, with j yij D mi . The multinomial model is an ordinal model if the categories have
a natural order.
Residuals are not available in the OBSTATS table or the output data set for multinomial models.
By default, and consistently with binomial models, the GENMOD procedure orders the response
categories for ordinal multinomial models from lowest to highest and models the probabilities of
the lower response levels. You can change the way PROC GENMOD orders the response levels
with the RORDER= option in the PROC GENMOD statement. The order that PROC GENMOD
uses is shown in the “Response Profiles” output table described in the section “Response Profile” on
page 2556.
The GENMOD procedure supports only the ordinal multinomial model. If .pi1 ; pi 2 ; pi k / are
the category probabilities, the cumulative category
P probabilities are modeled with the same link
functions used for binomial data. Let Pi r D rj D1 pij , r D 1; 2; ; k 1, be the cumulative
category probabilities (note that Pi k D 1). The ordinal model is
g.Pi r / D r C x0 ˇ for r D 1; 2; k
1
where 1 ; 2 ; k 1 are intercept terms that depend only on the categories and xi is a vector of
covariates that does not include an intercept term. The logit, probit, and complementary log-log
link functions g are available. These are obtained by specifying the MODEL statement options
DIST=MULTINOMIAL and LINK=CUMLOGIT (cumulative logit), LINK=CUMPROBIT (cumulative probit), or LINK=CUMCLL (cumulative complementary log-log). Alternatively,
Pi r D F.r C x0 ˇ/ for r D 1; 2; k
where F D g
distribution.
1
1
is a cumulative distribution function for the logistic, normal, or extreme-value
PROC GENMOD estimates the intercept parameters 1 ; 2 ; k
by maximum likelihood.
1
and regression parameters ˇ
The subpopulations i are defined by constant values of the AGGREGATE= variable. This has no
effect on the parameter estimates, but it does affect the deviance and Pearson chi-square statistics; it also affects parameter estimate standard errors if you specify the SCALE=DEVIANCE or
SCALE=PEARSON option.
Zero-Inflated Models
Count data that have an incidence of zeros greater than expected for the underlying probability
distribution of counts can be modeled with a zero-inflated distribution. In GENMOD, the underlying
distribution can be either Poisson or negative binomial. See Lambert (1992), Long (1997) and
Cameron and Trivedi (1998) for more information about zero-inflated models. The population is
considered to consist of two types of individuals. The first type gives Poisson or negative binomial
distributed counts, which might contain zeros. The second type always gives a zero count. Let be
Zero-Inflated Models F 2531
the underlying distribution mean and ! be the probability of an individual being of the second type.
The parameter ! is called here the zero-inflation probability, and is the probability of zero counts
in excess of the frequency predicted by the underlying distribution. You can request that the zero
inflation probability be displayed in an output data set with the PZERO keyword. The probability
distribution of a zero-inflated Poisson random variable Y is given by
(
! C .1 !/e for y D 0
Pr.Y D y/ D
ye for y D 1; 2; : : :
.1 !/ yŠ
and the probability distribution of a zero-inflated negative binomial random variable Y is given by
(
! C .1 !/.1 C k/
for y D 0
Pr.Y D y/ D
€.yC1=k/
.k/y
.1 !/ €.yC1/€.1=k/
for
y D 1; 2; : : :
.1Ck/yC1=k
where k is the negative binomial dispersion parameter.
You can model the parameters ! and in GENMOD with the regression models:
h.!i / D z0i g.i / D x0i ˇ
where h is one of the binary link functions: logit, probit, or complementary log-log. The link
function h is the logit link by default, or the link function option specified in the ZEROMODEL
statement. The link function g is the log link function by default, or the link function specified in the
MODEL statement, for both the Poisson and the negative binomial. The covariates zi for observation
i are determined by the model specified in the ZEROMODEL statement, and the covariates xi are
determined by the model specified in the MODEL statement. The regression parameters and ˇ are
estimated by maximum likelihood.
The mean and variance of Y for the zero-inflated Poisson are given by
E.Y / D D .1 !/
!
Var.Y / D C
2
1 !
and for the zero-inflated negative binomial by
E.Y / D D .1 !/
!
k
Var.Y / D C
C
2
1 !
1 !
You can request that the mean of Y be displayed for each observation in an output data set with the
PRED keyword.
2532 F Chapter 37: The GENMOD Procedure
Generalized Estimating Equations
Let yij , j D 1; : : : ; ni , i D 1; : : : ; K, represent the j th measurement on the i th subject. There are
P
ni measurements on subject i and K
i D1 ni total measurements.
Correlated data are modeled using the same link function and linear predictor setup (systematic
component) as the independence case. The random component is described by the same variance
functions as in the independence case, but the covariance structure of the correlated measurements
must also be modeled. Let the vector of measurements on the ith subject be Yi D Œyi1 ; : : : ; yi ni 0
with corresponding vector of means i D Œi1 ; : : : ; i ni 0 , and let Vi be the covariance matrix of
Yi . Let the vector of independent, or explanatory, variables for the j th measurement on the i th
subject be
xij D Œxij1 ; : : : ; xijp 0
The generalized estimating equation of Liang and Zeger (1986) for estimating the p 1 vector of
regression parameters ˇ is an extension of the independence estimating equation to correlated data
and is given by
S.ˇ/ D
K
X
D0i Vi 1 .Yi
i .ˇ// D 0
i D1
where
Di D
@i
@ˇ
Since
g.ij / D xij 0 ˇ
where g is the link function, the p ni matrix of partial derivatives of the mean with respect to the
regression parameters for the i th subject is given by
2 x
xi ni 1 3
i11
:
:
:
6 g 0 .i1 /
g 0 .i ni / 7
6
7
@0i
:
::
0
6
7
::
Di D
D6
:
7
@ˇ
4 xi1p
xi ni p 5
:::
g 0 .i1 /
g 0 .i ni /
Working Correlation Matrix
Let Ri .˛/ be an ni ni “working” correlation matrix that is fully specified by the vector of parameters
˛. The covariance matrix of Yi is modeled as
1
1
1
1
Vi D Ai2 Wi 2 R.˛/Wi 2 Ai2
Generalized Estimating Equations F 2533
where Ai is an ni ni diagonal matrix with v.ij / as the j th diagonal element and Wi is an ni ni
diagonal matrix with wij as the j th diagonal, where wij is a weight specified with the WEIGHT
statement. If there is no WEIGHT statement, wij D 1 for all i and j . If Ri .˛/ is the true correlation
matrix of Yi , then Vi is the true covariance matrix of Yi .
The working correlation matrix is usually unknown and must be estimated. It is estimated in the
iterative fitting process by using the current value of the parameter vector ˇ to compute appropriate
functions of the Pearson residual
yij ij
eij D p
v.ij /=wij
If you specify the working correlation as R0 D I, which is the identity matrix, the GEE reduces to
the independence estimating equation.
Following are the structures of the working correlation supported by the GENMOD procedure and
the estimators used to estimate the working correlations.
Working Correlation Structure
Estimator
Fixed
Corr.Yij ; Yi k / D rj k
The working correlation is not estiwhere rj k is the j kth element of a constant, mated in this case.
user-specified correlation matrix R0 .
Independent
Corr.Yij ; Yi k / D
1 j Dk
0 j ¤k
The working correlation is not estimated in this case.
m-dependent
8
< 1
˛t
Corr.Yij ; Yi;j Ct / D
:
0
t D0
t D 1; 2; : : : ; m
t >m
˛O t
1
.Kt p/
Kt D
D
PK P
PK
i D1
j ni t
i D1 .ni
eij ei;j Ct
t/
Exchangeable
Corr.Yij ; Yi k / D
1 j Dk
˛ j ¤k
˛O D
1
.N p/
N D 0:5
PK P
i D1
PK
i D1 ni .ni
j <k eij ei k
1/
Unstructured
Corr.Yij ; Yi k / D
1
j Dk
˛j k j ¤ k
Autoregressive
AR(1)
Corr.Yij ; Yi;j Ct / D ˛ t
for t D 0; 1; 2; : : : ; ni j
˛O j k D
˛O
1
.K p/
1
.K1 p/
K1 D
PK
i D1 eij ei k
D
PK P
PK
i D1
i D1 .ni
j ni 1 eij ei;j C1
1/
2534 F Chapter 37: The GENMOD Procedure
Dispersion Parameter
The dispersion parameter is estimated by
O D
ni
K X
X
1
N
where N D
parameters.
p
2
eij
i D1 j D1
PK
iD1 ni
is the total number of measurements and p is the number of regression
The square root of O is reported by PROC GENMOD as the scale parameter in the “Analysis of GEE
Parameter Estimates Model-Based Standard Error Estimates” output table. If a fixed scale parameter
is specified with the NOSCALE option in the MODEL statement, then the fixed value is used in
estimating the model-based covariance matrix and standard errors.
Fitting Algorithm
The following is an algorithm for fitting the specified model by using GEEs. Note that this is not
in general a likelihood-based method of estimation, so that inferences based on likelihoods are not
possible for GEE methods.
1. Compute an initial estimate of ˇ with an ordinary generalized linear model assuming independence.
2. Compute the working correlations R based on the standardized residuals, the current ˇ, and
the assumed structure of R.
3. Compute an estimate of the covariance:
1
1
1
1
2
2
O
Vi D Ai2 Wi 2 R.˛/W
i Ai
4. Update ˇ:
"
ˇrC1 D ˇr C
K
X
@i 0
i D1
@ˇ
Vi
1 @i
@ˇ
#
1" K
X
i D1
@i 0 1
V .Yi
@ˇ i
#
i /
5. Repeat steps 2-4 until convergence.
Missing Data
See Diggle, Liang, and Zeger (1994, Chapter 11) for a discussion of missing values in longitudinal
data. Suppose that you intend to take measurements Yi1 ; : : : ; Yi n for the i th unit. Missing values for
which Yij are missing whenever Yi k is missing for all j k are called dropouts. Otherwise, missing
values that occur intermixed with nonmissing values are intermittent missing values. The GENMOD
procedure can estimate the working correlation from data containing both types of missing values by
using the all available pairs method, in which all nonmissing pairs of data are used in the moment
Generalized Estimating Equations F 2535
estimators of the working correlation parameters defined previously. The resulting covariances and
standard errors are valid under the missing completely at random (MCAR) assumption.
For example, for the unstructured working correlation model,
˛O j k D
1
.K 0
X
p/
eij ei k
where the sum is over the units that have nonmissing measurements at times j and k, and K 0 is the
number of units with nonmissing measurements at j and k. Estimates of the parameters for other
working correlation types are computed in a similar manner, using available nonmissing pairs in the
appropriate moment estimators.
The contribution of the i th unit to the parameter update equation is computed by omitting the
0
elements of .Yi i /, the columns of D0i D @
, and the rows and columns of Vi corresponding to
@ˇ
missing measurements.
Parameter Estimate Covariances
O is given by
The model-based estimator of Cov.ˇ/
O D I0 1
†m .ˇ/
where
I0 D
K
X
@i 0
i D1
@ˇ
Vi
1 @i
@ˇ
This is the GEE equivalent of the inverse of the Fisher information matrix that is often used in
generalized linear models as an estimator of the covariance estimate of the maximum likelihood
estimator of ˇ. It is a consistent estimator of the covariance matrix of ˇO if the mean model and the
working correlation matrix are correctly specified.
The estimator
†e D I0 1 I1 I0 1
O where
is called the empirical, or robust, estimator of the covariance matrix of ˇ,
I1 D
K
X
@i 0
i D1
@ˇ
Vi 1 Cov.Yi /Vi
1 @i
@ˇ
O even if the working
It has the property of being a consistent estimator of the covariance matrix of ˇ,
correlation matrix is misspecified—that is, if Cov.Yi / ¤ Vi . See Zeger, Liang, and Albert (1988)),
Royall (1986), and White (1982) for further information about the robust variance estimate. In
computing †e , ˇ and are replaced by estimates, and Cov.Yi / is replaced by the estimate
.Yi
O
i .ˇ//.Y
i
O 0
i .ˇ//
2536 F Chapter 37: The GENMOD Procedure
Multinomial GEEs
Lipsitz, Kim, and Zhao (1994) and Miller, Davis, and Landis (1993) describe how to extend GEEs to
multinomial data. Currently, only the independent working correlation is available for multinomial
models in PROC GENMOD.
Alternating Logistic Regressions
If the responses are binary (that is, they take only two values), then there is an alternative method to
account for the association among the measurements. The alternating logistic regressions (ALR)
algorithm of Carey, Zeger, and Diggle (1993) models the association between pairs of responses with
log odds ratios, instead of with correlations, as ordinary GEEs do.
For binary data, the correlation between the jth and kth response is, by definition,
Pr.Yij D 1; Yi k D 1/ ij i k
Corr.Yij ; Yi k / D p
ij .1 ij /i k .1 i k /
The joint probability in the numerator satisfies the following bounds, by elementary properties of
probability, since ij D Pr.Yij D 1/:
max.0; ij C i k
1/ Pr.Yij D 1; Yi k D 1/ min.ij ; i k /
The correlation, therefore, is constrained to be within limits that depend in a complicated way on the
means of the data.
The odds ratio, defined as
OR.Yij ; Yi k / D
Pr.Yij D 1; Yi k D 1/ Pr.Yij D 0; Yi k D 0/
Pr.Yij D 1; Yi k D 0/ Pr.Yij D 0; Yi k D 1/
is not constrained by the means and is preferred, in some cases, to correlations for binary data.
The ALR algorithm seeks to model the logarithm of the odds ratio, ij k D log.OR.Yij ; Yi k //, as
ij k D z0ij k ˛
where ˛ is a q 1 vector of regression parameters and zij k is a fixed, specified vector of coefficients.
The parameter ij k can take any value in . 1; 1/ with ij k D 0 corresponding to no association.
The log odds ratio, when modeled in this way with a regression model, can take different values in
subgroups defined by zij k . For example, zij k can define subgroups within clusters, or it can define
“block effects” between clusters.
You specify a GEE model for binary data that uses log odds ratios by specifying a model for the
mean, as in ordinary GEEs, and a model for the log odds ratios. You can use any of the link functions
appropriate for binary data in the model for the mean, such as logistic, probit, or complementary
log-log. The ALR algorithm alternates between a GEE step to update the model for the mean and a
logistic regression step to update the log odds ratio model. Upon convergence, the ALR algorithm
provides estimates of the regression parameters for the mean, ˇ, the regression parameters for the
log odds ratios, ˛, their standard errors, and their covariances.
Generalized Estimating Equations F 2537
Specifying Log Odds Ratio Models
Specifying a regression model for the log odds ratio requires you to specify rows of the z matrix zij k
for each cluster i and each unique within-cluster pair .j; k/. The GENMOD procedure provides
several methods of specifying zij k . These are controlled by the LOGOR=keyword and associated
options in the REPEATED statement. The supported keywords and the resulting log odds ratio
models are described as follows.
EXCH
specifies exchangeable log odds ratios. In this model, the log odds ratio
is a constant for all clusters i and pairs .j; k/. The parameter ˛ is the
common log odds ratio.
zij k D 1 for all i; j; k
FULLCLUST
specifies fully parameterized clusters. Each cluster is parameterized in the
same way, and there is a parameter for each unique pair within clusters.
If a complete cluster is of size n, then there are n.n2 1/ parameters in
the vector ˛. For example, if a full cluster is of size 4, then there are
43
2 D 6 parameters, and the z matrix is of the form
2
6
6
6
ZD6
6
6
4
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
3
7
7
7
7
7
7
5
The elements of ˛ correspond to log odds ratios for cluster pairs in the
following order:
Pair
Parameter
(1,2)
(1,3)
(1,4)
(2.3)
(2,4)
(3,4)
Alpha1
Alpha2
Alpha3
Alpha4
Alpha5
Alpha6
LOGORVAR(variable)
specifies log odds ratios by cluster. The argument variable is a variable
name that defines the “block effects” between clusters. The log odds
ratios are constant within clusters, but they take a different value for
each different value of the variable. For example, if Center is a variable
in the input data set taking a different value for k treatment centers,
then specifying LOGOR=LOGORVAR(Center) requests a model with
different log odds ratios for each of the k centers, constant within center.
NESTK
specifies k-nested log odds ratios. You must also specify the SUBCLUST=variable option to define subclusters within clusters. Within
each cluster, PROC GENMOD computes a log odds ratio parameter for
2538 F Chapter 37: The GENMOD Procedure
pairs having the same value of variable for both members of the pair and
one log odds ratio parameter for each unique combination of different
values of variable.
NEST1
specifies 1-nested log odds ratios. You must also specify the SUBCLUST=variable option to define subclusters within clusters. There are
two log odds ratio parameters for this model. Pairs having the same value
of variable correspond to one parameter; pairs having different values
of variable correspond to the other parameter. For example, if clusters
are hospitals and subclusters are wards within hospitals, then patients
within the same ward have one log odds ratio parameter, and patients
from different wards have the other parameter.
ZFULL
specifies the full z matrix. You must also specify a SAS data set containing the z matrix with the ZDATA=data-set-name option. Each observation in the data set corresponds to one row of the z matrix. You
must specify the ZDATA data set as if all clusters are complete—that
is, as if all clusters are the same size and there are no missing observations. The ZDATA data set has KŒnmax .nmax 1/=2 observations, where K is the number of clusters and nmax is the maximum
cluster size. If the members of cluster i are ordered as 1; 2; ; n,
then the rows of the z matrix must be specified for pairs in the order
.1; 2/; .1; 3/; ; .1; n/; .2; 3/; ; .2; n/; ; .n 1; n/. The variables
specified in the REPEATED statement for the SUBJECT effect must
also be present in the ZDATA= data set to identify clusters. You must
specify variables in the data set that define the columns of the z matrix
by the ZROW=variable-list option. If there are q columns (q variables
in variable-list), then there are q log odds ratio parameters. You can
optionally specify variables indicating the cluster pairs corresponding to
each row of the z matrix with the YPAIR=(variable1, variable2) option.
If you specify this option, the data from the ZDATA data set are sorted
within each cluster by variable1 and variable2. See Example 37.6 for an
example of specifying a full z matrix.
ZREP
specifies a replicated z matrix. You specify z matrix data exactly as you
do for the ZFULL case, except that you specify only one complete cluster.
The z matrix for the one cluster is replicated for each cluster. The number
of observations in the ZDATA data set is nmax .n2max 1/ , where nmax is
the size of a complete cluster (a cluster with no missing observations).
ZREP(matrix)
specifies direct input of the replicated z matrix.
You specify
the z matrix for one cluster with the syntax LOGOR=ZREP
( .y1 y2 /z1 z2 zq ; ), where y1 and y2 are numbers representing a pair of observations and the values z1 ; z2 ; ; zq make up the
corresponding row of the z matrix. The number of rows specified is
nmax .nmax 1/
, where nmax is the size of a complete cluster (a cluster
2
with no missing observations). For example,
Generalized Estimating Equations F 2539
LOGOR =
ZREP((1
(1
(1
(2
(2
(3
2)
3)
4)
3)
4)
4)
1
1
1
1
1
1
0,
0,
0,
1,
1,
1)
specifies the 43
2 D 6 rows of the z matrix for a cluster of size 4 with
q D 2 log odds ratio parameters. The log odds ratio for the pairs (1 2),
(1 3), (1 4) is ˛1 , and the log odds ratio for the pairs (2 3), (2 4), (3 4) is
˛1 C ˛2 .
Quasi-likelihood Information Criterion
The quasi-likelihood information criterion (QIC) was developed by Pan (2001) as a modification of
the Akaike information criterion (AIC) to apply to models fit by GEEs.
Define the quasi-likelihood under the independence working correlation assumption, evaluated with
the parameter estimates under the working correlation of interest as
O
Q.ˇ.R/;
/ D
ni
K X
X
O
Q.ˇ.R/;
I .Yij ; Xij //
i D1 j D1
where the quasi-likelihood contribution of the j th observation in the i th cluster is defined in the
O
section “Quasi-likelihood Functions” on page 2539 and ˇ.R/
are the parameter estimates obtained
from GEEs with the working correlation of interest R.
QIC is defined as
QIC.R/ D
O
O I VOR /
2Q.ˇ.R/;
/ C 2trace.
O I is the inverse of the model-based covariance
where VOR is the robust covariance estimate and 
O
estimate under the independent working correlation assumption, evaluated at ˇ.R/,
the parameter
estimates obtained from GEEs with the working correlation of interest R.
PROC GENMOD also computes an approximation to QIC.R/ defined by Pan (2001) as
QICu .R/ D
O
2Q.ˇ.R/;
/ C 2p
where p is the number of regression parameters.
Pan (2001) notes that QIC is appropriate for selecting regression models and working correlations,
whereas QICu is appropriate only for selecting regression models.
Quasi-likelihood Functions
See McCullagh and Nelder (1989) and Hardin and Hilbe (2003) for discussions of quasi-likelihood
functions. The contribution of observation j in cluster i to the quasi-likelihood function evaluated
2540 F Chapter 37: The GENMOD Procedure
Q
at the regression parameters ˇ is given by Q.ˇ; I .Yij ; Xij // D ij , where Qij is defined in the
following list. These are used in the computation of the quasi-likelihood information criteria (QIC)
for goodness of fit of models fit with GEEs. The wij are prior weights, if any, specified with the
WEIGHT or FREQ statements. Note that the definition of the quasi-likelihood for the negative
binomial differs from that given in McCullagh and Nelder (1989). The definition used here allows
the negative binomial quasi-likelihood to approach the Poisson as k ! 0.
Normal:
Qij D
1
wij .yij
2
ij /2
Inverse Gaussian:
wij .ij :5yij /
Qij D
2ij
Gamma:
Qij D
wij
yij
C log.ij /
ij
Negative binomial:
1
Qij D wij log € yij C
k
kij
1
1
1
log €
C yij log
C log
k
1 C kij
k
1 C kij
Poisson:
Qij D wij .yij log.ij /
ij /
Binomial:
Qij D wij Œrij log.pij / C .nij
rij / log.1
pij /
Multinomial (s categories):
Qij D wij
s
X
yij k log.ij k /
kD1
Generalized Score Statistics
Boos (1992) and Rotnitzky and Jewell (1990) describe score tests applicable to testing L0 ˇ D 0 in
GEEs, where L0 is a user-specified r p contrast matrix or a contrast for a Type 3 test of hypothesis.
Let ˇQ be the regression parameters resulting from solving the GEE under the restricted model
Q be the generalized estimating equation values at ˇ.
Q
L0 ˇ D 0, and let S.ˇ/
The generalized score statistic is
Q 0 †m L.L0 †e L/
T D S.ˇ/
1 0
Q
L †m S.ˇ/
where †m is the model-based covariance estimate and †e is the empirical covariance estimate. The
p-values for T are computed based on the chi-square distribution with r degrees of freedom.
Assessment of Models Based on Aggregates of Residuals F 2541
Assessment of Models Based on Aggregates of Residuals
Lin, Wei, and Ying (2002) present graphical and numerical methods for model assessment based on
the cumulative sums of residuals over certain coordinates (such as covariates or linear predictors)
or some related aggregates of residuals. The distributions of these stochastic processes under the
assumed model can be approximated by the distributions of certain zero-mean Gaussian processes
whose realizations can be generated by simulation. Each observed residual pattern can then be
compared, both graphically and numerically, with a number of realizations from the null distribution.
Such comparisons enable you to assess objectively whether the observed residual pattern reflects
anything beyond random fluctuation. These procedures are useful in determining appropriate
functional forms of covariates and link function. You use the ASSESS|ASSESSMENT statement to
perform this kind of model-checking with cumulative sums of residuals, moving sums of residuals, or
LOESS smoothed residuals. See Example 37.8 and Example 37.9 for examples of model assessment.
Let the model for the mean be
g.i / D x0i ˇ
where i is the mean of the response yi and xi is the vector of covariates for the i th observation.
Denote the raw residual resulting from fitting the model as
ei D y i
O i
and let xij be the value of the j th covariate in the model for observation i . Then to check the
functional form of the j th covariate, consider the cumulative sum of residuals with respect to xij ,
n
1 X
I.xij x/ei
Wj .x/ D p
n
i D1
where I./ is the indicator function. For any x, Wj .x/ is the sum of the residuals with values of xj
less than or equal to x.
Denote the score, or gradient vector, by
U.ˇ/ D
n
X
h.x0 ˇ/xi .yi
.x0 ˇ//
i D1
where .r/ D g
h.r/ D
1 .r/,
and
1
g0..r//V ..r//
Let J be the Fisher information matrix
J.ˇ/ D
@U.ˇ/
@ˇ 0
Define
n
1 X
O
ŒI.xij x/ C 0 .xI ˇ/J
WO j .x/ D p
n
i D1
1
O i h.x0 ˇ/e
O i Zi
.ˇ/x
2542 F Chapter 37: The GENMOD Procedure
where
.xI ˇ/ D
n
X
I.xij x/
i D1
@.x0i ˇ/
@ˇ
and Zi are independent N.0; 1/ random variables. Then the conditional distribution of WO j .x/, given
.yi ; xi /; i D 1; : : : ; n, under the null hypothesis H0 that the model for the mean is correct, is the
same asymptotically as n ! 1 as the unconditional distribution of Wj .x/ (Lin, Wei, and Ying
2002).
You can approximate realizations from the null hypothesis distribution of Wj .x/ by repeatedly
generating normal samples Zi ; i D 1; : : : ; n, while holding .yi ; xi /; i D 1; : : : ; n, at their observed
values and computing WO j .x/ for each sample.
You can assess the functional form of covariate j by plotting a few realizations of WO j .x/ on the
same plot as the observed Wj .x/ and visually comparing to see how typical the observed Wj .x/ is
of the null distribution samples.
You can supplement the graphical inspection method with a Kolmogorov-type supremum test. Let
sj be the observed value of Sj D supx jWj .x/j. The p-value PrŒSj sj  is approximated by
PrŒSOj sj , where SOj D supx jWO j .x/j. PrŒSOj sj  is estimated by generating realizations of
WO j .:/ (1,000 is the default number of realizations).
You can check the link function instead of the j th covariate by using values of the linear predictor
x0i ˇO in place of values of the j th covariate xij . The graphical and numerical methods described
previously are then sensitive to inadequacies in the link function.
An alternative aggregate of residuals is the moving sum statistic
n
1 X
Wj .x; b/ D p
I.x
n
b xij x/ei
i D1
If you specify the keyword WINDOW(b), then the moving sum statistic with window size b is used
instead of the cumulative sum of residuals, with I.x b xij x/ replacing I.xij x/ in the
earlier equation.
If you specify the keyword LOESS(f ), loess smoothed residuals are used in the preceding formulas,
where f is the fraction of the data to be used at a given point. If f is not specified, f D 13 is used.
For data .Yi ; Xi /; i D 1; : : : ; n, define r as the nearest integer to nf and h as the rth smallest among
jXi xj; i D 1; : : : ; n. Let
Xi x
Ki .x/ D K
h
where
K.t / D
Define
70
.1
81
jt j3 /3 I. 1 t 1/
Assessment of Models Based on Aggregates of Residuals F 2543
wi .x/ D Ki .x/ŒS2 .x/
.Xi
x/S1 .x/
where
S1 .x/ D
n
X
Ki .x/.Xi
x/
Ki .x/.Xi
x/2
i D1
S2 .x/ D
n
X
i D1
Then the loess estimate of Y at x is defined by
YO .x/ D
n
X
i D1
wi .x/
Pn
Yi
i D1 wi .x/
Loess smoothed residuals for checking the functional form of the j th covariate are defined by
replacing Yi with ei and Xi with xij . To implement the graphical and numerical assessment methods,
I.xij x/ is replaced with Pnwi .x/
in the formulas for Wj .x/ and WO j .x/.
w .x/
i D1
i
You can perform the model checking described earlier for marginal models for dependent responses
fit by generalized estimating equations (GEEs). Let yi k denote the kth measurement on the ith
cluster, i D 1; : : : ; K, k D 1; : : : ; ni , and let xi k denote the corresponding vector of covariates. The
marginal mean of the response i k D E.yi k / is assumed to depend on the covariate vector by
g.i k / D x0i k ˇ
where g is the link function.
Define the vector of residuals for the i th cluster as
ei D .ei1 ; : : : ; ei ni /0 D .yi1
O i1 ; : : : ; yi ni
O i ni /0
You use the following extension of Wj .x/ defined earlier to check the functional form of the j th
covariate:
K ni
1 XX
Wj .x/ D p
I.xi kj x/ei k
K i D1 kD1
where xi kj is the j th component of xi k .
The null distribution of Wj .x/ can be approximated by the conditional distribution of
(n
)
K
i
X
X
1
O 0V
O 1
O 0 1D
WO j .x/ D p
I.xi kj x/ei k C 0 .x; ˇ/I
i i ei Zi
K i D1 kD1
2544 F Chapter 37: The GENMOD Procedure
O i and V
O i are defined as in the section “Generalized Estimating Equations” on page 2532
where D
with the unknown parameters replaced by their estimated values,
.x; ˇ/ D
ni
K X
X
i D1 kD1
I0 D
K
X
I.xi kj x/
@i k
@ˇ
O 1O
O 0V
D
i i Di
i D1
and Zi ; i D 1; : : : ; K, are independent N.0; 1/ random variables. You replace xi kj with the linear
predictor x0i k ˇO in the preceding formulas to check the link function.
Case Deletion Diagnostic Statistics
For ordinary generalized linear models, regression diagnostic statistics developed by Williams (1987)
can be requested in an output data set or in the OBSTATS table by specifying the DIAGNOSTICS | INFLUENCE option in the MODEL statement. These diagnostics measure the influence of an
individual observation on model fit, and generalize the one-step diagnostics developed by Pregibon
(1981) for the logistic regression model for binary data.
Preisser and Qaqish (1996) further generalized regression diagnostics to apply to models for correlated data fit by generalized estimating equations (GEEs), where the influence of entire clusters
of correlated observations, or the influence of individual observations within a cluster, is measured.
These diagnostic statistics can be requested in an output data set or in the OBSTATS table if a model
for correlated data is specified with a REPEATED statement.
The next two sections use the following notation:
ˇO
is the maximum likelihood estimate of the regression parameters ˇ, or, in the case of correlated
data, the solution of the GEEs.
ˇOŒi 
is the corresponding estimate evaluated with the ith observation deleted, or, in the case of
correlated data, with the i th cluster deleted.
p
is the dimension of the regression parameter vector ˇ.
rpi
is the standardized Pearson residual pvyi.1ih / , where vi is the variance of the i th response
i
i
and hi is the leverage defined in the section “H | LEVERAGE” on page 2545.
vi
is the variance of response i , var.Yi / D V .i /, where V ./ is the variance function and is the dispersion parameter.
wi
is the prior weight of the i th observation specified with the WEIGHT statement. If there is no
WEIGHT statement, wi D 1 for all i .
All unknown quantities are replaced by their estimated values in the following two sections.
Case Deletion Diagnostic Statistics F 2545
Diagnostics for Ordinary Generalized Linear Models
The following statistics are available for generalized linear models.
DFBETA
The DFBETA statistic for measuring the influence of the i th observation is defined as the one-step
approximation to the difference in the MLE of the regression parameter vector and the MLE of the
regression parameter vector without the i th observation. This one-step approximation assumes a
Fisher scoring step, and is given by
ˇO
ˇOŒi  DFBETAi D .X 0 W X /
1
1
Xi0 Wi 2 .1
hi /
1
2
rpi
where hi is the leverage defined in the section “H | LEVERAGE” on page 2545.
DFBETAS
The standardized DFBETA statistic for assessing the influence of the i th observation on the j th
regression parameter is defined as the DFBETA statistic for the j th parameter divided by its estimated
standard deviation, where the standard deviation is estimated from all the data.
DFBETASij D DFBETAij =O .ˇj /
DOBS | COOKD | COOKSD
In normal linear regression, the influence of observation i can be measured by Cook’s distance (Cook
and Weisberg 1982). A measure of influence of observation i for generalized linear models that is
equivalent to Cook’s distance for normal linear regression is given by
DOBSi D p
1
hi .1
hi /
1 2
rpi
where hi is the leverage defined in the section “H | LEVERAGE” on page 2545. This measure is the
O
one-step approximation to 2p 1 ŒL.ˇ/
L.ˇOŒi  /, where L.ˇ/ is the log likelihood evaluated at ˇ.
H | LEVERAGE
wi
The Fisher scores, or expected, weight for observation i is wei D V . /.g
0 . //2 . Let W be the
i
i
diagonal matrix with wei as the i th diagonal. The leverage hi of the i th observation is defined as the
i th diagonal element of the hat matrix
1
H D W 2 .X 0 W X /
1
W
1
2
Diagnostics for Models Fit by Generalized Estimating Equations (GEEs)
The diagnostic statistics in this section were developed by Preisser and Qaqish (1996). See the
section “Generalized Estimating Equations” on page 2532 for further information and notation for
generalized estimating equations (GEEs). The following additional notation is used in this section.
2546 F Chapter 37: The GENMOD Procedure
0 0
Partition the design matrix X and response vector Y by cluster; that is, let X D .X10 ; : : : ; XK
/ , and
0
0 0
Y D .Y1 ; : : : ; YK / corresponding to the K clusters.
P
Let ni be the number of responses for cluster i , and denote by N D K
i D1 ni the total number of
observations. Denote by Ai the ni ni diagonal matrix with V .ij / as the j th diagonal element. If
there is a WEIGHT statement, the diagonal element of Ai is V .ij /=wij , where wij is the specified
weight of the j th observation in the i th cluster. Let B the N N diagonal matrix with g 0 .ij / as
diagonal elements, i D 1; : : : ; K, j D 1; : : : ; ni . Let Bi the ni ni diagonal matrix corresponding
to cluster i with g 0 .ij / as the j th diagonal element.
Let W be the N N block diagonal weight matrix whose i th block, corresponding to the i th cluster,
is the ni ni matrix
1
1
O i 2 Bi
Wei D Bi 1 Ai 2 Ri 1 .˛/A
1
where Ri is the working correlation matrix for cluster i .
Let
Qi D Xi .X 0 W X /
1
Xi0
where Xi is the ni p design matrix corresponding to cluster i .
Define the adjusted residual vector as
E D B.Y
and Ei D Bi .Yi
O
/
O i /, the estimated residual for the i th cluster.
Let the subscript Œi  denote estimates evaluated without the i th cluster, Œi t  estimates evaluated using
all the data except the t th observation of the i th cluster, and let i Œt  denote matrices corresponding to
the i th cluster without the t th observation.
The following statistics are available for generalized estimating equation models.
CH | CLUSTERH | CLEVERAGE
The leverage of cluster i is contained in the matrix Hi D Qi Wei , and is summarized by the trace of
Hi ,
chi D t r.Hi /
The leverage hi of the tth observation in the i th cluster is the t th diagonal element of Hi .
DFBETAC
The effect of deleting cluster i on the estimated parameter vector is given by the following one-step
approximation for ˇO ˇOŒi  :
DBETACi D .X 0 W X /
1
Xi0 .Wei 1
Qi /
1
Ei
Case Deletion Diagnostic Statistics F 2547
DFBETACS
The cluster deletion statistic DFBETAC can be standardized using the variances of ˇO based on the
complete data. The standardized one-step approximation for the change in ˇOj due to deletion of
cluster i is
DBETACSij D
DBETACij
1
0W X /
O
Œ.X
1 2
jj
DFBETAO
Partition the matrices Wei and Vi as
Wei D
Wei t
Wei Œt t
Vi D Wei 1 D
Wei t Œt 
Wei Œt 
Vi t
Vi Œt t
and let Ei t D Bi t .Yi t
Vi t Œt 
Vi Œt 
O i t / and Ei Œt  D Bi Œt  .Yi Œt 
O i Œt  /.
The effect of deleting the tth observation from the i th cluster is given by the following one-step
approximation to ˇO ˇOŒi t  :
DBETAOi t D .X 0 W X /
1
XQ i0t
EQ i t
Wei t1
QQ i t
where XQ i t D Xi t Vi t Œt  Vi Œt1 Xi Œt  , QQ i t D XQ i t .X 0 W X /
Note that Wei t , QQ i t , and EQ i t are scalars.
1 XQ 0 ,
it
and EQ i t D Ei t
Vi t Œt  Vi Œt1 Ei Œt  .
DFBETAOS
The observation deletion statistic DFBETAO can be standardized using the variances of ˇO based on
the complete data. The standardized one-step approximation for the change in ˇOj due to deletion of
observation t in cluster i is
DBETAOSi tj D
DBETAOi tj
0W X /
O
Œ.X
1
1 2
jj
DCLS | CLUSTERCOOKD | CLUSTERCOOKSD
A measure of the standardized influence of the subset m of observations on the overall fit is .ˇO
O For deletion of cluster i , this is approximated by
ˇOŒm /0 .X 0 W X /.ˇO ˇOŒm /=p .
DCLSi D Ei0 .Wei 1
Qi /
1
/Qi .Wei 1
Qi /
1
/Ei =p O
2548 F Chapter 37: The GENMOD Procedure
DOBS | COOKD | COOKSD
The measure of overall fit in the section “DCLS | CLUSTERCOOKD | CLUSTERCOOKSD” on
page 2547 for the deletion of the tth observation in the i th cluster is approximated by
DOBSi t D
EQ i2t QQ i t
1
O
p .W
ei t
QQ i t /2
where EQ i t , QQ i t , and Wei t are defined in the section “DFBETAO” on page 2547. In the case of the
independence working correlation, this is equal to the measure for ordinary generalized linear models
defined in the section “DOBS | COOKD | COOKSD” on page 2545.
MCLS | CLUSTERDFIT
A studentized distance measure of the type defined in the section “DCLS | CLUSTERCOOKD | CLUSTERCOOKSD” on page 2547 of the influence of the i th cluster is given by
M CLSi D Ei0 .Wei 1
Qi /
1
Hi Ei =p O
Bayesian Analysis
In generalized linear models, the response has a probability distribution from a family of distributions
of the exponential form. That is, the probability density of the response Y for continuous response
variables, or the probability function for discrete responses, can be expressed as
y b. /
f .y/ D exp
C c.y; /
a./
for some functions a, b, and c that determine the specific distribution. The canonical parameters
depend only on the means of the response i , which are related to the regression parameters ˇ
through the link function g.i / D x 0 ˇ. The additional parameter is the dispersion parameter.
1
The GENMOD procedure estimates the regression parameters and the scale parameter D 2 by
maximum likelihood. However, the GENMOD procedure can also provide Bayesian estimates of the
regression parameters and either the scale , the dispersion , or the precision D 1 by sampling
from the posterior distribution. Except where noted, the following discussion applies to either , ,
or , although is used to illustrate the formulas. Note that the Poisson and binomial distributions do
not have a dispersion parameter, and the dispersion is considered to be fixed at D 1. The ASSESS,
CONTRAST, ESTIMATE, OUTPUT, and REPEATED statements, if specified, are ignored. Also
ignored are the PLOTS= option in the PROC GENMOD statement and the following options in the
MODEL statement: ALPHA=, CORRB, COVB, TYPE1, TYPE3, SCALE=DEVIANCE (DSCALE),
SCALE=PEARSON (PSCALE), OBSTATS, RESIDUALS, XVARS, PREDICTED, DIAGNOSTICS,
and SCALE= for Poisson and binomial distributions. The multinomial and zero-inflated Poisson
distributions are not available for Bayesian analysis.
See the section “Assessing Markov Chain Convergence” on page 155 for information about assessing
the convergence of the chain of posterior samples.
Several algorithms, specified with the SAMPLING= option in the BAYES statement, are available in
GENMOD for drawing samples from the posterior distribution.
Bayesian Analysis F 2549
ARMS Algorithm for Gibbs Sampling
This section provides details for Bayesian analysis by Gibbs sampling in generalized linear models.
See the section “Gibbs Sampler” on page 151 for a general discussion of Gibbs sampling. See Gilks,
Richardson, and Spiegelhalter (1996) for a discussion of applications of Gibbs sampling to a number
of different models, including generalized linear models.
Let D .1 ; : : : ; k /0 be the parameter vector. For generalized linear models, the i s are the
regression coefficients ˇi s and the dispersion parameter . Let L.Dj/ be the likelihood function,
where D is the observed data. Let ./ be the prior distribution. The full conditional distribution of
Œi jj ; i ¤ j  is proportional to the joint distribution; that is,
.i jj ; i ¤ j; D/ / L.Dj/p./
For instance, the one-dimensional conditional distribution of 1 given j D j ; 2 j k, is
computed as
.1 jj D j ; 2 j k; D/ D L.Dj. D .1 ; 2 ; : : : ; k /0 /p. D .1 ; 2 ; : : : ; k /0 /
.0/
.0/
Suppose you have a set of arbitrary starting values f1 ; : : : ; k g. Using the ARMS (adaptive
rejection Metropolis sampling) algorithm of Gilks and Wild (1992) and Gilks, Best, and Tan (1995),
you can do the following:
.1/
.0/
.1/
.1/
.1/
.1/
.0/
draw 1 from Œ1 j2 ; : : : ; k 
.0/
.0/
draw 2 from Œ2 j1 ; 3 ; : : : ; k 
:::
.1/

1
draw k from Œk j1 ; : : : ; k
.1/
.1/
This completes one iteration of the Gibbs sampler. After one iteration, you have f1 ; : : : ; k g.
.n/
.n/
After n iterations, you have f1 ; : : : ; k g. PROC GENMOD implements the ARMS algorithm
provided by Gilks (2003) to draw a sample from a full conditional distribution. See the section
“Adaptive Rejection Sampling Algorithm” on page 152 for more information about the ARMS
algorithm. The ARMS algorithm is the default method used to sample from the posterior distribution,
except in the case of a normal distribution with a conjugate prior, in which case a closed form
is available for the posterior distribution. See any of the introductory references in Chapter 7,
“Introduction to Bayesian Analysis Procedures,” for a discussion of conjugate prior distributions for a
linear model with the normal distribution.
Gamerman Algorithm
The Gamerman algorithm, unlike a Gibbs sampling algorithm, samples parameters from their
multivariate posterior conditional distribution. The algorithm uses the structure of generalized linear
models to efficiently sample from the posterior distribution of the model parameters. For a detailed
description and explanation of the algorithm, see Gamerman (1997) and the section “Gamerman
Algorithm” on page 153.
2550 F Chapter 37: The GENMOD Procedure
Independence Metropolis Algorithm
The independence Metropolis algorithm is another sampling algorithm that draws multivariate
samples from the posterior distribution. See the section “Independence Sampler” on page 153 for
more details.
Posterior Samples Output Data Set
You can output posterior samples into a SAS data set through ODS. The following SAS statement
outputs the posterior samples into the SAS data set Post:
OUTPOST= Post ;
The data set also includes the variables LogPost and LogLike, which represent the log of the posterior
likelihood and the log of the likelihood, respectively.
Priors for Model Parameters
The model parameters are the regression coefficients and the dispersion parameter (or the precision or
scale), if the model has one. The priors for the dispersion parameter and the priors for the regression
coefficients are assumed to be independent, while you can have a joint multivariate normal prior for
the regression coefficients.
Dispersion, Precision, or Scale Parameter
The gamma distribution G.a; b/ has a probability density function
Gamma Prior
f .u/ D
b.bu/a 1 e
€.a/
bu
;
u>0
where a is the shape parameter and b is the inverse-scale parameter. The mean is
is ba2 .
1
;
Inverse Gamma Prior
f .u/ D
ba
u
€.a/
u>0
The inverse gamma distribution IG.a; b/ has a probability density function
.aC1/
e
b=u
;
u>0
where a is the shape parameter and b is the scale parameter. The mean is
variance is
and the variance
The joint prior density is given by
Improper Prior
p.u/ / u
a
b
b2
.a 1/2 .a 2/
if a > 2.
b
a 1
if a > 1, and the
Bayesian Analysis F 2551
Regression Coefficients
Let ˇ be the regression coefficients.
Jeffreys’ Prior
The joint prior density is given by
1
p.ˇ/ / jI.ˇ/j 2
where I.ˇ/ is the Fisher information matrix for the model. If the underlying model has a scale
parameter (for example, a normal linear regression model), then the Fisher information matrix is
computed with the scale parameter set to a fixed value of one.
If you specify the CONDITIONAL option, then Jeffreys’ prior, conditional on the current Markov
chain value of the generalized linear model precision parameter , is given by
1
jI.ˇ/j 2
where is the model precision parameter.
See Ibrahim and Laud (1991) for a full discussion, with examples, of Jeffreys’ prior for generalized
linear models.
Assume ˇ has a multivariate normal prior with mean vector ˇ0 and covariance
matrix †0 . The joint prior density is given by
Normal Prior
p.ˇ/ / e
1
2 .ˇ
ˇ0 /0 †0 1 .ˇ ˇ0 /
If you specify the CONDITIONAL option, then, conditional on the current Markov chain value of
the generalized linear model precision parameter , the joint prior density is given by
p.ˇ/ / e
Uniform Prior
1
2 .ˇ
ˇ0 /0 †0 1 .ˇ ˇ0 /
The joint prior density is given by
p.ˇ/ / 1
Deviance Information Criterion
Let i be the model parameters at iteration i of the Gibbs sampler and let LL(i ) be the corresponding
model log likelihood. PROC GENMOD computes the following fit statistics defined by Spiegelhalter
et al. (2002):
Effective number of parameters:
pD D LL. /
LL.N /
2552 F Chapter 37: The GENMOD Procedure
Deviance information criterion (DIC):
DIC D LL. / C pD
where
LL. / D
1
n
Pn
N
1
n
Pn
D
i D1 LL.i /
i D1 i
PROC GENMOD uses the full log likelihoods defined in the section “Log-Likelihood Functions” on
page 2514, with all terms included, for computing the DIC.
Posterior Distribution
Denote the observed data by D.
The posterior distribution is
.ˇjD/ / LP .Djˇ/p.ˇ/
where LP .Djˇ/ is the likelihood function with regression coefficients ˇ as parameters.
Starting Values of the Markov Chains
When the BAYES statement is specified, PROC GENMOD generates one Markov chain containing
the approximate posterior samples of the model parameters. Additional chains are produced when
the Gelman-Rubin diagnostics are requested. Starting values (or initial values) can be specified in the
INITIAL= data set in the BAYES statement. If INITIAL= option is not specified, PROC GENMOD
picks its own initial values for the chains.
Denote Œx as the integral value of x. Denote sO .X / as the estimated standard error of the estimator X.
Regression Coefficients
For the first chain that the summary statistics and regression diagnostics are based on, the default
initial values are estimates of the mode of the posterior distribution. If the INITIALMLE option is
specified, the initial values are the maximum likelihood estimates; that is,
.0/
ˇi
D ˇOi
Initial values for the rth chain (r 2) are given by
r
.0/
ˇi D ˇOi ˙ 2 C
sO .ˇOi /
2
with the plus sign for odd r and minus sign for even r.
Exact Logistic and Poisson Regression F 2553
Dispersion, Scale, or Precision Parameter Let be the generalized linear model parameter you choose to sample, either the dispersion, scale,
or precision parameter. Note that the Poisson and binomial distributions do not have this additional
parameter.
For the first chain that the summary statistics and regression diagnostics are based on, the default
initial values are estimates of the mode of the posterior distribution. If the INITIALMLE option is
specified, the initial values are the maximum likelihood estimates; that is,
.0/ D O
The initial values of the rth chain (r 2) are given by
O
.0/ D e
O
˙ Œ 2r C2 sO ./
with the plus sign for odd r and minus sign for even r.
OUTPOST= Output Data Set
The OUTPOST= data set contains the generated posterior samples. There are 3+n variables, where n
is the number of model parameters. The variable Iteration represents the iteration number, the variable
LogLike contains the log of the likelihood, and the variable LogPost contains the log of the posterior.
The other n variables represent the draws of the Markov chain for the model parameters.
Exact Logistic and Poisson Regression
The theory of exact logistic regression, also called exact conditional logistic regression, is described
in the section “Exact Conditional Logistic Regression” on page 3974 of Chapter 51, “The LOGISTIC
Procedure.” The following discussion of exact Poisson regression, also called exact conditional
Poisson regression, uses the notation given in that section.
Note that in exact logistic regression, the coefficients C.t/ are the counts of the number of possible
response vectors y that generate t: C.t/ D jjfy W y 0 X D t 0 gjj. However, when performing an exact
Poisson regression, this value is replaced by
C.t/ D
n
y
XY
Ni i
yi Š
 i D1
where  D fyW y 0 X D tg and Ni D exp.oi / is the exponential of the offset oi for observation i .
The probability density function (pdf) for T is created by summing over all binary sequences y that
generate an observable t
C.t/ exp.t 0 ˇ/
Pr.T D t/ D Qn
xi0 ˇ
/
i D1 exp.Ni e
2554 F Chapter 37: The GENMOD Procedure
However, the conditional likelihood of TI given TN D tN is the same as that for exact logistic
regression.
For details about hypothesis testing and estimation, see the sections “Hypothesis Tests” on page 3976
and “Inference for a Single Parameter” on page 3977 of Chapter 51, “The LOGISTIC Procedure.” See
the section “Computational Resources for Exact Logistic Regression” on page 3985 of Chapter 51,
“The LOGISTIC Procedure,” for some computational notes about exact analyses.
The offset variable, oi , is required for exact Poisson regression computationally to provide an
stopping point for the algorithm. Denote Ni D exp.oi /. In exact logistic binary regression, there
are a finite number, 2n , of possible y vectors to be considered. Since a Poisson-distributed response
variable can take an infinite number of values,
Qthere is an infinite number of y-vectors to be scanned.
The offset variable reduces this number to niD1 Ni response vectors. On a practical level, as Ni
gets large the probability of the Poisson random variable achieving this value drops to zero, so Ni
can be thought of as the point at which you believe the value does not matter. If you are modeling
rates, then Ni is the maximum possible value for each observation in the experiment; for example, if
you are counting the number of rats in a cage that acquire a disease, then Ni is the number of rats in
cage i . Finally, if you are conditioning out
Pnthe intercept, and denoting the observed response as y0 ,
every Ni has an effective maximum of i D1 y0i , which is the sufficient statistic for the intercept
term.
OUTDIST= Output Data Set
The OUTDIST= data set contains every exact conditional distribution necessary to process the
corresponding EXACT statement. For example, the following statements create one distribution for
the x1 parameter and another for the x2 parameters, and produce the data set dist shown in Table 37.7:
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
;
proc genmod data=test exactonly;
class x2 / param=ref;
model y=x1 x2 / d=b;
exact x1 x2/ outdist=dist;
proc print data=dist;
run;
Exact Logistic and Poisson Regression F 2555
Table 37.7 OUTDIST= Data Set
Obs
x1
x20
x21
1
2
3
4
5
6
7
8
9
.
.
.
.
.
.
.
.
.
0
0
0
1
1
1
2
2
3
0
1
2
0
1
2
0
1
0
10
11
12
13
14
2
3
4
5
6
.
.
.
.
.
.
.
.
.
.
Count
Score
Prob
3
15
9
15
18
6
19
2
3
5.81151
1.66031
3.12728
1.46523
0.21675
4.58644
1.61869
3.27293
6.27189
0.03333
0.16667
0.10000
0.16667
0.20000
0.06667
0.21111
0.02222
0.03333
6
12
11
18
3
3.03030
0.75758
0.00000
0.75758
3.03030
0.12000
0.24000
0.22000
0.36000
0.06000
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 y 0 x1
was equal to 2, and for which y 0 x20, y 0 x21, and y 0 1 were equal to their actual observed values
(displayed in the “Sufficient Statistics” table).
N OTE : If you are performing an exact Poisson analysis, then the Count variable is replaced by a
variable named Weight.
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:
exact
exact
exact
exact
'O1'
'OJ12'
'OA12'
'OE12'
x1
/
x1 x2 / jointonly
x1 x2 / joint
x1 x2 / estimate
outdist=o1;
outdist=oj12;
outdist=oa12;
outdist=oe12;
The O1 statement outputs a single exact conditional distribution. The OJ12 statement outputs only
the joint distribution for x1 and x2. The OA12 statement outputs three conditional distributions:
one for x1, one for x2, and one jointly for x1 and x2. The OE12 statement outputs two conditional
distributions: one for x1 and the other for x2. Data set oe12 contains both the x1 and x2 variables;
2556 F Chapter 37: The GENMOD Procedure
the distribution for x1 has missing values in the x2 column while the distribution for x2 has missing
values in the x1 column.
Missing Values
For generalized linear models, PROC GENMOD ignores any observation with a missing value for
any variable involved in the model. You can score an observation in an output data set by setting
only the response value to missing. For models fit with generalized estimating equations (GEEs),
observations with missing values within a cluster are not used, and all available pairs are used in
estimating the working correlation matrix. Clusters with fewer observations than the full cluster size
are treated as having missing observations occurring at the end of the cluster. You can specify the
order of missing observations with the WITHINSUBJECT= option. See the section “Missing Data”
on page 2534 for more information about missing values in GEEs.
Displayed Output for Classical Analysis
The following output is produced by the GENMOD procedure. Note that some of the tables are
optional and appear only in conjunction with the REPEATED statement and its options or with
options in the MODEL statement. For details, see the section “ODS Table Names” on page 2568.
Model Information
The “Model Information” table displays the two-level data set name, the response distribution, the
link function, the response variable name, the offset variable name, the frequency variable name, the
scale weight variable name, the number of observations used, the number of events if events/trials
format is used for response, the number of trials if events/trials format is used for response, the
sum of frequency weights, the number of missing values in data set, and the number of invalid
observations (for example, negative or 0 response values with gamma distribution or number of
observations with events greater than trials with binomial distribution).
Class Level Information
If you use classification variables in the model, PROC GENMOD displays the levels of classification
variables specified in the CLASS statement and in the MODEL statement. The levels are displayed
in the same sorted order used to generate columns in the design matrix.
Response Profile
If you specify an ordinal model for the multinomial distribution, a table titled “Response Profile” is
displayed containing the ordered values of the response variable and the number of occurrences of
the values used in the model.
Displayed Output for Classical Analysis F 2557
Iteration History for Parameter Estimates
If you specify the ITPRINT model option, PROC GENMOD displays a table containing the following
for each iteration in the Newton-Raphson procedure for model fitting: the iteration number, the ridge
value, the log likelihood, and values of all parameters in the model.
Criteria for Assessing Goodness of Fit
In the “Criteria for Assessing Goodness of Fit” table, PROC GENMOD displays the degrees of
freedom for deviance and Pearson’s chi-square, equal to the number of observations minus the number
of regression parameters estimated, the deviance, the deviance divided by degrees of freedom, the
scaled deviance, the scaled deviance divided by degrees of freedom, Pearson’s chi-square, Pearson’s
chi-square divided by degrees of freedom, the scaled Pearson’s chi-square, the scaled Pearson’s
chi-square divided by degrees of freedom, the log likelihood (excludes factorial terms) the full log
likelihood, the Akaike information criterion, the corrected Akaike information criterion, and the
Bayesian information criterion. The information in this table is valid only for maximum likelihood
model fitting, and the table is not printed if the REPEATED statement is specified.
Last Evaluation of the Gradient
If you specify the model option ITPRINT, the GENMOD procedure displays the last evaluation of
the gradient vector.
Last Evaluation of the Hessian
If you specify the model option ITPRINT, the GENMOD procedure displays the last evaluation of
the Hessian matrix.
Analysis of (Initial) Parameter Estimates
The “Analysis of (Initial) Parameter Estimates” table contains the results from fitting a generalized
linear model to the data. If you specify the REPEATED statement, these GLM parameter estimates
are used as initial values for the GEE solution, and are displayed only if the PRINTMLE option in
the REPEATED statement is specified. For each parameter in the model, PROC GENMOD displays
the parameter name, as follows:
the variable name for continuous regression variables
the variable name and level for classification variables and interactions involving classification
variables
SCALE for the scale variable related to the dispersion parameter
2558 F Chapter 37: The GENMOD Procedure
In addition, PROC GENMOD displays the degrees of freedom for the parameter, the estimate value,
the standard error, the Wald chi-square value, the p-value based on the chi-square distribution, and
the confidence limits (Wald or profile likelihood) for parameters.
Lagrange Multiplier Statistics
If you specify that either the model intercept or the scale parameter is fixed, for those distributions
that have a distribution scale parameter, the GENMOD procedure displays a table of Lagrange
multiplier, or score, statistics for testing the validity of the constrained parameter that contains the
test statistic, and the p-value.
Estimated Covariance Matrix
If you specify the model option COVB, the GENMOD procedure displays the estimated covariance
matrix, defined as the inverse of the information matrix at the final iteration. This is based on
the expected information matrix if the EXPECTED option is specified in the MODEL statement.
Otherwise, it is based on the Hessian matrix used at the final iteration. This is, by default, the
observed Hessian unless altered by the SCORING option in the MODEL statement.
Estimated Correlation Matrix
If you specify the CORRB model option, PROC GENMOD displays the estimated correlation
matrix. This is based on the expected information matrix if the EXPECTED option is specified in the
MODEL statement. Otherwise, it is based on the Hessian matrix used at the final iteration. This is,
by default, the observed Hessian unless altered by the SCORING option in the MODEL statement.
Iteration History for LR Confidence Intervals
If you specify the ITPRINT and LRCI model options, PROC GENMOD displays an iteration history
table for profile likelihood-based confidence intervals. For each parameter in the model, PROC
GENMOD displays the parameter identification number, the iteration number, the log-likelihood
value, parameter values.
Likelihood Ratio-Based Confidence Intervals for Parameters
If you specify the LRCI and the ITPRINT options in the MODEL statement, a table is displayed that
summarizes profile likelihood-based confidence intervals for all parameters. For each parameter in
the model, the table displays the confidence coefficient, the parameter identification number, lower
and upper endpoints of confidence intervals for the parameter, and values of all other parameters at
the solution.
Displayed Output for Classical Analysis F 2559
LR Statistics for Type 1 Analysis
If you specify the TYPE1 model option, a table is displayed that contains the name of the effect, the
deviance for the model including the effect and all previous effects, the degrees of freedom for the
effect, the likelihood ratio statistic for testing the significance of the effect, and the p-value computed
from the chi-square distribution with the effect’s degrees of freedom.
If you specify either the SCALE=DEVIANCE or SCALE=PEARSON option in the MODEL
statement, columns are displayed that contain the name of the effect, the deviance for the model
including the effect and all previous effects, the numerator degrees of freedom, the denominator
degrees of freedom, the chi-square statistic for testing the significance of the effect, the p-value
computed from the chi-square distribution with numerator degrees of freedom, the F statistic for
testing the significance of the effect, and the p-value based on the F distribution.
Iteration History for Type 3 Contrasts
If you specify the model options ITPRINT and TYPE3, an iteration history table is displayed for
fitting the model with Type 3 contrast constraints for each effect that contains the effect name, the
iteration number, the ridge value, the log likelihood, and values of all parameters.
LR Statistics for Type 3 Analysis
If you specify the TYPE3 model option, a table is displayed that contains, for each effect in the
model, the name of the effect, the likelihood ratio statistic for testing the significance of the effect,
the degrees of freedom for the effect, and the p-value computed from the chi-square distribution.
If you specify either the SCALE=DEVIANCE or SCALE=PEARSON option in the MODEL
statement, columns are displayed that contain the name of the effect, the likelihood ratio statistic
for testing the significance of the effect, the F statistic for testing the significance of the effect, the
numerator degrees of freedom, the denominator degrees of freedom, the p-value based on the F
distribution, and the p-value computed from the chi-square distribution with the numerator’s degrees
of freedom.
Wald Statistics for Type 3 Analysis
If you specify the TYPE3 and WALD model options, a table is displayed that contains the name of
the effect, the degrees of freedom of the effect, the Wald statistic for testing the significance of the
effect, and the p-value computed from the chi-square distribution.
Parameter Information
If you specify the ITPRINT, COVB, CORRB, WALDCI, or LRCI option in the MODEL statement,
or if you specify a CONTRAST statement, a table is displayed that identifies parameters with
numbers, rather than names, for use in tables and matrices where a compact identifier for parameters
2560 F Chapter 37: The GENMOD Procedure
is helpful. For each parameter, the table contains an index number that identifies the parameter, and
the parameter name, including level information for effects containing classification variables.
Observation Statistics
If you specify the OBSTATS option in the MODEL statement, PROC GENMOD displays a table
containing miscellaneous statistics. Residuals and case deletion diagnostic statistics are not available
for the multinomial distribution. Case deletion diagnostics are not available for zero-inflated models.
For each observation in the input data set, the following are displayed:
the value of the response variable
the predicted value of the mean
the value of the linear predictor The value of an OFFSET variable is added to the linear
predictor.
the estimated standard error of the linear predictor
the value of the negative of the weight in the Hessian matrix at the final iteration. This is the
expected weight if the EXPECTED option is specified in the MODEL statement. Otherwise, it
is the weight used in the final iteration. That is, it is the observed weight unless the SCORING=
option has been specified.
approximate lower and upper endpoints for a confidence interval for the predicted value of the
mean
raw residual
Pearson residual
deviance residual
standardized Pearson residual
standardized deviance residual
likelihood residual
leverage
Cook’s distance statistic
DFBETA statistic, for each parameter
standardized DFBETA statistic, for each parameter
zero-inflation probability for zero-inflated models
response mean for zero-inflated models
Displayed Output for Classical Analysis F 2561
ESTIMATE Statement Results
If you specify a REPEATED statement, the ESTIMATE statement results apply to the specified GEE
model. Otherwise, they apply to the specified generalized linear model.
For each ESTIMATE statement, the table contains the contrast label, the estimated value of the
contrast, the standard error of the estimate, the significance level ˛, .1 ˛/ 100% confidence
intervals for contrast, the Wald chi-square statistic for the contrast, and the p-value computed from
the chi-square distribution.
If you specify the EXP option, an additional row is displayed with statistics for the exponentiated
value of the contrast.
CONTRAST Coefficients
If you specify the CONTRAST or ESTIMATE statement and you specify the E option, a table titled
“Coefficients For Contrast label” is displayed, where label is the label specified in the CONTRAST
statement. The table contains the contrast label, and the rows of the contrast matrix.
Iteration History for Contrasts
If you specify the ITPRINT option, an iteration history table is displayed for fitting the model with
contrast constraints for each effect. The table contains the contrast label, the iteration number, the
ridge value, the log likelihood, and values of all parameters.
CONTRAST Statement Results
If you specify a REPEATED statement, the CONTRAST statement results apply to the specified
GEE model. Otherwise, they apply to the specified generalized linear model.
A table is displayed that contains the contrast label, the degrees of freedom for the contrast, and the
likelihood ratio, score, or Wald statistic for testing the significance of the contrast. Score statistics
are used in GEE models, likelihood ratio statistics are used in generalized linear models, and Wald
statistics are used in both. Also displayed are the p-value computed from the chi-square distribution,
and the type of statistic computed for this contrast: Wald, LR, or score.
If you specify either the SCALE=DEVIANCE or SCALE=PEARSON option for generalized linear
models, columns are displayed that contain the contrast label, the likelihood ratio statistic for
testing the significance of the contrast, the F statistic for testing the significance of the contrast, the
numerator degrees of freedom, the denominator degrees of freedom, the p-value based on the F
distribution, and the p-value computed from the chi-square distribution with numerator degrees of
freedom.
2562 F Chapter 37: The GENMOD Procedure
LSMEANS Coefficients
If you specify the LSMEANS statement and you specify the E option, the “Coefficients for effect
Least Squares Means” table is displayed, where effect is the effect specified in the LSMEANS
statement. The table contains the effect names and the rows of least squares means coefficients.
Least Squares Means
If you specify the LSMEANS statement, the “Least Squares Means” table is displayed. The table
contains for each effect the following: the effect name, and for each level of each effect the following:
the least squares mean estimate
standard error
chi-square value
p-value computed from the chi-square distribution
If you specify the DIFF option, a table titled “Differences of Least Squares Means” is displayed
containing corresponding statistics for the differences between the least squares means for the levels
of each effect.
GEE Model Information
If you specify the REPEATED statement, the “GEE Model Information” table displays the correlation
structure of the working correlation matrix or the log odds ratio structure, the within-subject effect,
the subject effect, the number of clusters, the correlation matrix dimension, and the minimum and
maximum cluster size.
Log Odds Ratio Parameter Information
If you specify the REPEATED statement and specify a log odds ratio model for binary data with the
LOGOR= option, then the “Log Odds Ratio Parameter Information” table is displayed showing the
correspondence between data pairs and log odds ratio model parameters.
Iteration History for GEE Parameter Estimates
If you specify the REPEATED statement and the MODEL statement option ITPRINT, the “Iteration History For GEE Parameter Estimates” table is displayed. The table contains the parameter
identification number, the iteration number, and values of all parameters.
Displayed Output for Classical Analysis F 2563
Last Evaluation of the Generalized Gradient and Hessian
If you specify the REPEATED statement and select ITPRINT as a model option, PROC GENMOD
displays the “Last Evaluation Of The Generalized Gradient And Hessian” table.
GEE Parameter Estimate Covariance Matrices
If you specify the REPEATED statement and the COVB option, PROC GENMOD displays the
“Covariance Matrix (Model-Based)” and “Covariance Matrix (Empirical)” tables.
GEE Parameter Estimate Correlation Matrices
If you specify the REPEATED statement and the CORRB option, PROC GENMOD displays the
“Correlation Matrix (Model-Based)” and “Correlation Matrix (Empirical)” tables.
GEE Working Correlation Matrix
If you specify the REPEATED statement and the CORRW option, PROC GENMOD displays the
“Working Correlation Matrix” table.
GEE Fit Criteria
If you specify the REPEATED statement, PROC GENMOD displays the quasi-likelihood information
criteria for model fit QIC and QICu in the “GEE Fit Criteria” table.
Analysis of GEE Parameter Estimates
If you specify the REPEATED statement, PROC GENMOD uses empirical standard error estimates to
compute and display the “Analysis Of GEE Parameter Estimates Empirical Standard Error Estimates”
table that contains the parameter names as follows:
the variable name for continuous regression variables
the variable name and level for classification variables and interactions involving classification
variables
“Scale” for the scale variable related to the dispersion parameter
In addition, the parameter estimate, the empirical standard error, a 95% confidence interval, and the
Z score and p-value are displayed for each parameter.
If you specify the MODELSE option in the REPEATED statement, the “Analysis Of GEE Parameter
Estimates Model-Based Standard Error Estimates” table based on model-based standard errors is
also produced.
2564 F Chapter 37: The GENMOD Procedure
GEE Observation Statistics
If you specify the OBSTATS option in the REPEATED statement, PROC GENMOD displays a table
containing miscellaneous statistics. For each observation in the input data set, the following are
displayed:
the value of the response variable and all other variables in the model, denoted by the variable
names
the predicted value of the mean
the value of the linear predictor
the standard error of the linear predictor
confidence limits for the predicted values
raw residual
Pearson residual
cluster number
leverage
cluster leverage
cluster Cook’s distance statistic
studentized cluster Cook’s distance statistic
individual observation Cook’s distance statistic
cluster DFBETA statistic for each parameter
cluster standardized DFBETA statistic for each parameter
individual observation DFBETA statistic for each parameter
individual observation standardized DFBETA statistic for each parameter
Displayed Output for Bayesian Analysis
If a Bayesian analysis is requested with a BAYES statement, the displayed output includes the
following.
Displayed Output for Bayesian Analysis F 2565
Model Information
The “Model Information” table displays the two-level data set name, the number of burn-in iterations,
the number of iterations after the burn-in, the number of thinning iterations, the response distribution,
the link function, the response variable name, the offset variable name, the frequency variable name,
the scale weight variable name, the number of observations used, the number of events if events/trials
format is used for response, the number of trials if events/trials format is used for response, the
sum of frequency weights, the number of missing values in data set, and the number of invalid
observations (for example, negative or 0 response values with gamma distribution or number of
observations with events greater than trials with binomial distribution).
Class Level Information
The “Class Level Information” table displays the levels of classification variables if you specify a
CLASS statement.
Maximum Likelihood Estimates
The “Analysis of Maximum Likelihood Parameter Estimates” table displays the maximum likelihood
estimate of each parameter, the estimated standard error of the parameter estimator, and confidence
limits for each parameter.
Coefficient Prior
The “Coefficient Prior” table displays the prior distribution of the regression coefficients.
Independent Prior Distributions for Model Parameters
The “Independent Prior Distributions for Model Parameters” table displays the prior distributions of
additional model parameters (scale, exponential scale, Weibull scale, Weibull shape, gamma shape).
Initial Values and Seeds
The “Initial Values and Seeds” table displays the initial values and random number generator seeds
for the Gibbs chains.
Fit Statistics
The “Fit Statistics” table displays the deviance information criterion (DIC) and the effective number
of parameters.
2566 F Chapter 37: The GENMOD Procedure
Descriptive Statistics of the Posterior Samples
The “Descriptive Statistics of the Posterior Sample” table contains the size of the sample, the mean,
the standard deviation, and the quartiles for each model parameter.
Interval Estimates for Posterior Sample
The “Interval Estimates for Posterior Sample” table contains the HPD intervals and the credible
intervals for each model parameter.
Correlation Matrix of the Posterior Samples
The “Correlation Matrix of the Posterior Samples” table is produced if you include the CORR
suboption in the SUMMARY= option in the BAYES statement. This table displays the sample
correlation of the posterior samples.
Covariance Matrix of the Posterior Samples
The “Covariance Matrix of the Posterior Samples” table is produced if you include the COV suboption
in the SUMMARY= option in the BAYES statement. This table displays the sample covariance of
the posterior samples.
Autocorrelations of the Posterior Samples
The “Autocorrelations of the Posterior Samples” table displays the lag1, lag5, lag10, and lag50
autocorrelations for each parameter.
Gelman and Rubin Diagnostics
The “Gelman and Rubin Diagnostics” table is produced if you include the GELMAN suboption in
the DIAGNOSTIC= option in the BAYES statement. This table displays the estimate of the potential
scale reduction factor and its 97.5% upper confidence limit for each parameter.
Geweke Diagnostics
The “Geweke Diagnostics” table displays the Geweke statistic and its p-value for each parameter.
Displayed Output for Exact Analysis F 2567
Raftery and Lewis Diagnostics
The “Raftery Diagnostics” tables is produced if you include the RAFTERY suboption in the DIAGNOSTIC= option in the BAYES statement. This table displays the Raftery and Lewis diagnostics for
each variable.
Heidelberger and Welch Diagnostics
The “Heidelberger and Welch Diagnostics” table is displayed if you include the HEIDELBERGER
suboption in the DIAGNOSTIC= option in the BAYES statement. This table shows the results of a
stationary test and a halfwidth test for each parameter.
Effective Sample Size
The “Effective Sample Size” table displays, for each parameter, the effective sample size, the
correlation time, and the efficiency.
Monte Carlo Standard Errors
The “Monte Carlo Standard Errors” table displays, for each parameter, the Monte Carlo standard
error, the posterior sample standard deviation, and the ratio of the two.
Displayed Output for Exact Analysis
If an exact analysis is requested with an EXACT statement, the displayed output includes the
following tables. If the METHOD=NETWORKMC option is specified, the test and
p estimate tables
are renamed “Monte Carlo” tables and a Monte Carlo standard error column ( p.1 p/=n) is
displayed.
Sufficient Statistics
Displays if you request an OUTDIST= data set in an EXACT statement. The table lists the parameters
and their observed sufficient statistics.
(Monte Carlo) Conditional Exact Tests
This table tests the hypotheses that the parameters of interest are insignificant. See the section “Exact
Logistic and Poisson Regression” on page 2553 for details.
2568 F Chapter 37: The GENMOD Procedure
(Monte Carlo) Exact Parameter Estimates
Displays if you specify the ESTIMATE option in the EXACT statement. This table gives individual
parameter estimates for each variable (conditional on the values of all the other parameters in the
model), confidence limits, and a two-sided p-value (twice the one-sided p-value) for testing that the
parameter is zero. See the section “Exact Logistic and Poisson Regression” on page 2553 for details.
(Monte Carlo) Exact Odds Ratios
Displays if you specify the ESTIMATE=ODDS or ESTIMATE=BOTH option in the EXACT
statement. See the section “Exact Logistic and Poisson Regression” on page 2553 for details.
Strata Summary
Displays if a STRATA statement is also specified. Shows the pattern of the number of events and
the number of nonevents, or of the number of observations, in a stratum. See the section “STRATA
Statement” on page 2507 for more information.
Strata Information
Displays if a STRATA statement is specified with the INFO option.
ODS Table Names
PROC GENMOD assigns a name to each table that it creates. You can use these names to reference
the table when using the Output Delivery System (ODS) to select tables and create output data sets.
These names are listed separately in Table 37.8 for a maximum likelihood analysis, in Table 37.9 for
a Bayesian analysis, and in Table 37.10 for an Exact analysis. For more information about ODS, see
Chapter 20, “Using the Output Delivery System.”
Table 37.8
ODS Tables Produced in PROC GENMOD for a Classical Analysis
ODS Table Name
Description
Statement
Option
AssessmentSummary
ClassLevels
Contrasts
ContrastCoef
ConvergenceStatus
CorrB
Model assessment summary
Classification variable levels
Tests of contrasts
Contrast coefficients
Convergence status
Parameter estimate correlation matrix
Parameter estimate covariance matrix
Estimates of contrasts
ASSESS
CLASS
CONTRAST
CONTRAST
MODEL
MODEL
Default
Default
Default
E
Default
CORRB
MODEL
COVB
ESTIMATE
Default
CovB
Estimates
ODS Table Names F 2569
Table 37.8 continued
ODS Table Name
Description
Statement
Option
EstimateCoef
GEEEmpPEst
Contrast coefficients
GEE parameter estimates
with empirical standard errors
GEE QIC fit criteria
GEE log odds ratio model
information
GEE model information
GEE parameter estimates
with model-based standard
errors
GEE model-based correlation matrix
GEE model-based covariance matrix
GEE empirical correlation
matrix
GEE empirical covariance
matrix
GEE working correlation
matrix
Iteration history for contrasts
Iteration history for likelihood ratio confidence intervals
Iteration history for parameter estimates
Iteration history for GEE parameter estimates
Iteration history for Type 3
statistics
Likelihood ratio confidence
intervals
Coefficients for least squares
means
Least squares means differences
Least squares means
Lagrange statistics
Last evaluation of the generalized gradient and Hessian
Last evaluation of the gradient and Hessian
ESTIMATE
REPEATED
E
Default
REPEATED
REPEATED
Default
LOGOR=
REPEATED
REPEATED
Default
MODELSE
REPEATED
MCORRB
REPEATED
MCOVB
REPEATED
ECORRB
REPEATED
ECOVB
REPEATED
CORRW
GEEFitCriteria
GEELogORInfo
GEEModInfo
GEEModPEst
GEENCorr
GEENCov
GEERCorr
GEERCov
GEEWCorr
IterContrasts
IterLRCI
IterParms
IterParmsGEE
IterType3
LRCI
LSMeanCoef
LSMeanDiffs
LSMeans
LagrangeStatistics
LastGEEGrad
LastGradHess
MODEL CON- ITPRINT
TRAST
MODEL
LRCI ITPRINT
MODEL
ITPRINT
MODEL
REPEATED
MODEL
ITPRINT
MODEL
LRCI ITPRINT
LSMEANS
E
LSMEANS
DIFF
LSMEANS
MODEL
MODEL
REPEATED
MODEL
Default
NOINT | NOSCALE
ITPRINT
TYPE3 ITPRINT
ITPRINT
2570 F Chapter 37: The GENMOD Procedure
Table 37.8
continued
ODS Table Name
Description
Statement
Option
LinDep
CONTRAST
Default
MODEL
MODEL
Default
Default without REPEATED
Default
CONTRAST
Default
ObStats
Linearly dependent rows of
contrasts
Model information
Goodness-of-fit statistics
Number of observations
summary
Nonestimable rows of contrasts
Observation-wise statistics
MODEL
ParameterEstimates
Parameter estimates
MODEL
OBSTATS | CL |
PREDICTED |
RESIDUALS | XVARS
Default without REPEATED |
PRINTMLE with REPEATED
Default
DIST=MULTINOMIAL |
DIST=BINOMIAL
TYPE1
TYPE3
Default
ModelInfo
Modelfit
NObs
NonEst
ParmInfo
ResponseProfiles
Parameter indices
Frequency counts for multinomial and binary models
Type1
Type 1 tests
Type3
Type 3 tests
ZeroParameterEstimates Parameter estimates for zeroinflated model
Table 37.9
MODEL
MODEL
MODEL
MODEL
ZEROMODEL
ODS Tables Produced in PROC GENMOD for a Bayesian Analysis
ODS Table Name
Description
Statement
Option
AutoCorr
Autocorrelations of the posterior samples
Classification variable levels
Prior distribution of the regression coefficients
Convergence status of maximum likelihood estimation
Correlation matrix of the
posterior samples
Effective sample size
Fit statistics
Gelman and Rubin convergence diagnostics
Geweke convergence diagnostics
Heidelberger and Welch convergence diagnostics
Initial values of the Markov
chains
BAYES
Default
CLASS
BAYES
Default
Default
MODEL
Default
BAYES
SUMMARY=CORR
BAYES
BAYES
BAYES
Default
Default
DIAG=GELMAN
BAYES
Default
BAYES
DIAG=HEIDELBERGER
BAYES
Default
ClassLevels
CoeffPrior
ConvergenceStatus
Corr
ESS
FitStatistics
Gelman
Geweke
Heidelberger
InitialValues
ODS Table Names F 2571
Table 37.9 continued
ODS Table Name
Description
Statement
Option
IterParms
Iteration history for parameter estimates
Last evaluation of the gradient and Hessian for maximum likelihood estimation
Monte Carlo standard errors
Model information
Number of observations
Maximum likelihood estimates of model parameters
Parameter indices
Prior distribution for scale
and shape
HPD and equal-tail intervals
of the posterior samples
Posterior samples (for ODS
output data set only)
Summary statistics of the
posterior samples
Raftery and Lewis convergence diagnostics
MODEL
ITPRINT
MODEL
ITPRINT
BAYES
PROC
MODEL
DIAG=MCSE
Default
Default
Default
MODEL
BAYES
Default
Default
BAYES
Default
LastGradHess
MCError
ModelInfo
NObs
ParameterEstimates
ParmInfo
ParmPrior
PostIntervals
PosteriorSample
PostSummaries
Raftery
BAYES
BAYES
Default
BAYES
DIAG=RAFTERY
Table 37.10 ODS Tables Produced in PROC GENMOD for an Exact Analysis
ODS Table Name
Description
Statement
Option
ExactOddsRatio
Exact odds ratios
EXACT
ExactParmEst
Parameter estimates
EXACT
ExactTests
NStrataIgnored
Conditional exact tests
Number of uninformative
strata
Number of strata with specific response frequencies
Event and nonevent frequencies for each stratum
Sufficient statistics
EXACT
STRATA
ESTIMATE=ODDS,
ESTIMATE=BOTH
ESTIMATE,
ESTIMATE=PARM,
ESTIMATE=BOTH
Default
Default
STRATA
Default
STRATA
INFO
EXACT
OUTDIST=
StrataSummary
StrataInfo
SuffStats
2572 F Chapter 37: The GENMOD Procedure
ODS Graphics
To request graphics with PROC GENMOD, you must first enable ODS Graphics by specifying
the ODS GRAPHICS ON statement. See Chapter 21, “Statistical Graphics Using ODS,” for more
information. Some graphs are produced by default; other graphs are produced by using statements
and options. You can reference every graph produced through ODS Graphics with a name. The
names of the graphs that PROC GENMOD generates are listed in Table 37.11, along with the
required statements and options.
ODS Graph Names
PROC GENMOD 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 37.11.
To request these graphs, you must specify the ODS GRAPHICS ON statement in addition to the
options indicated in Table 37.11.
Table 37.11
ODS Graphics Produced by PROC GENMOD
ODS Graph Name
Description
Statement Option
ADPanel
Autocorrelation function
and density panel
Autocorrelation function
panel
Autocorrelation function
plot
Cluster Cook’s D by cluster number
Cluster DFFIT by cluster
number
Cluster leverage by cluster number
Cook’s distance
Panel of aggregates of
residuals
Model assessment based
on aggregates of residuals
Deviance residuals by
linear predictor
Deviance values
Cluster DFBeta by cluster number
DFBeta
BAYES
PLOTS=(AUTOCORR DENSITY)
BAYES
PLOTS= AUTOCORR
BAYES
PLOTS(UNPACK)=AUTOCORR
PROC
PLOTS=
PROC
PLOTS=
PROC
PLOTS=
PROC
ASSESS
PLOTS=
CRPANEL
ASSESS
Default
PROC
PLOTS=
PROC
PROC
PLOTS=
PLOTS=
PROC
PLOTS=
AutocorrPanel
AutocorrPlot
ClusterCooksDPlot
ClusterDFFITPlot
ClusterLeveragePlot
CooksDPlot
CumResidPanel
CumulativeResiduals
DevianceResidByXBeta
DevianceResidualPlot
DFBETAByCluster
DFBETAPlot
ODS Graphics F 2573
Table 37.11 continued
ODS Table Name
Description
DiagnosticPlot
Panel of residuals, in- PROC
fluence, and diagnostic MODEL
statistics
REPEATED
Leverage
PROC
Likelihood residuals by PROC
linear predictor
Likelihood residuals
PROC
Pearson residuals by lin- PROC
ear predictor
Pearson residuals
PROC
Predicted values
PROC
Raw residuals by linear PROC
predictor
Raw residuals
PROC
Standardized deviance PROC
residuals by linear
predictor
Standardized deviance PROC
residuals
Standardized cluster DF- PROC
Beta by cluster number
Standardized DFBeta
PROC
Standardized Pearson PROC
residuals by linear
predictor
Standardized Pearson PROC
residuals
Trace and autocorrela- BAYES
tion function panel
Trace, autocorrelation, BAYES
and density function
panel
Trace and density panel
BAYES
Trace panel
BAYES
Trace plot
BAYES
Zero-inflation probabili- PROC
ties
LeveragePlot
LikeResidByXBeta
LikeResidualPlot
PearsonResidByXBeta
PearsonResidualPlot
PredictedByObservation
RawResidByXBeta
RawResidualPlot
StdDevianceResidByXBeta
StdDevianceResidualPlot
StdDFBETAByCluster
StdDFBETAPlot
StdPearsonResidByXBeta
StdPearsonResidualPlot
TAPanel
TADPanel
TDPanel
TracePanel
TracePlot
ZeroInflationProbPlot
Statement Option
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=
PLOTS=(TRACE AUTOCORR)
Default
PLOTS=(TRACE DENSITY)
PLOTS=TRACE
PLOTS(UNPACK)=TRACE
PLOTS=
2574 F Chapter 37: The GENMOD Procedure
Examples: GENMOD Procedure
The following examples illustrate some of the capabilities of the GENMOD procedure. These are not
intended to represent definitive analyses of the data sets presented here. You should refer to the texts
cited in the references for guidance on complete analysis of data by using generalized linear models.
Example 37.1: Logistic Regression
In an experiment comparing the effects of five different drugs, each drug is tested on a number of
different subjects. The outcome of each experiment is the presence or absence of a positive response
in a subject. The following artificial data represent the number of responses r in the n subjects for
the five different drugs, labeled A through E. The response is measured for different levels of a
continuous covariate x for each drug. The drug type and the continuous covariate x are explanatory
variables in this experiment. The number of responses r is modeled as a binomial random variable for
each combination of the explanatory variable values, with the binomial number of trials parameter
equal to the number of subjects n and the binomial probability equal to the probability of a response.
The following DATA step creates the data set:
data drug;
input drug$ x r
datalines;
A .1
1 10
A
B .2
3 13
B
C .04 0 10
C
D .34 5 10
D
E .2 12 20
E
;
run;
n @@;
.23 2
.3
4
.15 0
.6
5
.34 15
12
15
11
9
20
A
B
C
D
E
.67 1
.45 5
.56 1
.7
8
.56 13
9
16
12
10
15
B
C
.78
.7
E
.8
5
2
13
12
17
20
A logistic regression for these data is a generalized linear model with response equal to the binomial
proportion r/n. The probability distribution is binomial, and the link function is logit. For these data,
drug and x are explanatory variables. The probit and the complementary log-log link functions are
also appropriate for binomial data.
PROC GENMOD performs a logistic regression on the data in the following SAS statements:
proc genmod data=drug;
class drug;
model r/n = x drug / dist = bin
link = logit
lrci;
run;
Since these data are binomial, you use the events/trials syntax to specify the response in the MODEL
statement. Profile likelihood confidence intervals for the regression parameters are computed using
the LRCI option.
Example 37.1: Logistic Regression F 2575
General model and data information is produced in Output 37.1.1.
Output 37.1.1 Model Information
The GENMOD Procedure
Model Information
Data Set
Distribution
Link Function
Response Variable (Events)
Response Variable (Trials)
WORK.DRUG
Binomial
Logit
r
n
The five levels of the CLASS variable DRUG are displayed in Output 37.1.2.
Output 37.1.2 CLASS Variable Levels
Class Level Information
Class
Levels
drug
5
Values
A B C D E
In the “Criteria For Assessing Goodness Of Fit” table displayed in Output 37.1.3, the value of the
deviance divided by its degrees of freedom is less than 1. A p-value is not computed for the deviance;
however, a deviance that is approximately equal to its degrees of freedom is a possible indication of
a good model fit. Asymptotic distribution theory applies to binomial data as the number of binomial
trials parameter n becomes large for each combination of explanatory variables. McCullagh and
Nelder (1989) caution against the use of the deviance alone to assess model fit. The model fit for each
observation should be assessed by examination of residuals. The OBSTATS option in the MODEL
statement produces a table of residuals and other useful statistics for each observation.
Output 37.1.3 Goodness-of-Fit Criteria
Criteria For Assessing Goodness Of Fit
Criterion
DF
Value
Value/DF
Deviance
Scaled Deviance
Pearson Chi-Square
Scaled Pearson X2
Log Likelihood
Full Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)
12
12
12
12
5.2751
5.2751
4.5133
4.5133
-114.7732
-23.7343
59.4686
67.1050
64.8109
0.4396
0.4396
0.3761
0.3761
2576 F Chapter 37: The GENMOD Procedure
In the “Analysis Of Parameter Estimates” table displayed in Output 37.1.4, chi-square values for
the explanatory variables indicate that the parameter values other than the intercept term are all
significant. The scale parameter is set to 1 for the binomial distribution. When you perform an
overdispersion analysis, the value of the overdispersion parameter is indicated here. See the section
“Overdispersion” on page 2521 for a discussion of overdispersion.
Output 37.1.4 Parameter Estimates
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
DF
Estimate
Standard
Error
Intercept
x
drug
drug
drug
drug
drug
Scale
1
1
1
1
1
1
0
0
0.2792
1.9794
-2.8955
-2.0162
-3.7952
-0.8548
0.0000
1.0000
0.4196
0.7660
0.6092
0.4052
0.6655
0.4838
0.0000
0.0000
A
B
C
D
E
Likelihood Ratio
95% Confidence
Limits
-0.5336
0.5038
-4.2280
-2.8375
-5.3111
-1.8072
0.0000
1.0000
1.1190
3.5206
-1.7909
-1.2435
-2.6261
0.1028
0.0000
1.0000
Wald
Chi-Square
Pr > ChiSq
0.44
6.68
22.59
24.76
32.53
3.12
.
0.5057
0.0098
<.0001
<.0001
<.0001
0.0773
.
NOTE: The scale parameter was held fixed.
The preceding table contains the profile likelihood confidence intervals for the explanatory variable
parameters requested with the LRCI option. Wald confidence intervals are displayed by default.
Profile likelihood confidence intervals are considered to be more accurate than Wald intervals (see
Aitkin et al. (1989)), especially with small sample sizes. You can specify the confidence coefficient
with the ALPHA= option in the MODEL statement. The default value of 0.05, corresponding to 95%
confidence limits, is used here. See the section “Confidence Intervals for Parameters” on page 2525
for a discussion of profile likelihood confidence intervals.
Example 37.2: Normal Regression, Log Link
Consider the following data, where x is an explanatory variable and y is the response variable. It
appears that y varies nonlinearly with x and that the variance is approximately constant. A normal
distribution with a log link function is chosen to model these data; that is, log.i / D x0i ˇ so that
i D exp.x0i ˇ/.
Example 37.2: Normal Regression, Log Link F 2577
data nor;
input x y;
datalines;
0 5
0 7
0 9
1 7
1 10
1 8
2 11
2 9
3 16
3 13
3 14
4 25
4 24
5 34
5 32
5 30
;
run;
The following SAS statements produce the analysis with the normal distribution and log link:
proc genmod data=nor;
model y = x / dist
link
output out
=
pred
=
resraw
=
reschi
=
resdev
=
stdreschi =
stdresdev =
reslik
=
run;
= normal
= log;
Residuals
Pred
Resraw
Reschi
Resdev
Stdreschi
Stdresdev
Reslik;
The OUTPUT statement is specified to produce a data set that contains predicted values and residuals
for each observation. This data set can be useful for further analysis, such as residual plotting.
The results from these statements are displayed in Output 37.2.1.
Output 37.2.1 Log-Linked Normal Regression
The GENMOD Procedure
Model Information
Data Set
Distribution
Link Function
Dependent Variable
WORK.NOR
Normal
Log
y
2578 F Chapter 37: The GENMOD Procedure
Output 37.2.1 continued
Criteria For Assessing Goodness Of Fit
Criterion
DF
Value
Value/DF
Deviance
Scaled Deviance
Pearson Chi-Square
Scaled Pearson X2
Log Likelihood
Full Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)
14
14
14
14
52.3000
16.0000
52.3000
16.0000
-32.1783
-32.1783
70.3566
72.3566
72.6743
3.7357
1.1429
3.7357
1.1429
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
DF
Estimate
Standard
Error
Intercept
x
Scale
1
1
1
1.7214
0.3496
1.8080
0.0894
0.0206
0.3196
Wald 95%
Confidence Limits
1.5461
0.3091
1.2786
Wald
Chi-Square
Pr > ChiSq
370.76
286.64
<.0001
<.0001
1.8966
0.3901
2.5566
NOTE: The scale parameter was estimated by maximum likelihood.
The PROC GENMOD scale parameter, in the case of the normal distribution, is the standard deviation.
By default, the scale parameter is estimated by maximum likelihood. You can specify a fixed standard
deviation by using the NOSCALE and SCALE= options in the MODEL statement.
proc print data=Residuals ;
run;
Output 37.2.2 Data Set of Predicted Values and Residuals
Obs x
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
0
0
0
1
1
1
2
2
3
3
3
4
4
5
5
5
y
5
7
9
7
10
8
11
9
16
13
14
25
24
34
32
30
Pred
5.5921
5.5921
5.5921
7.9324
7.9324
7.9324
11.2522
11.2522
15.9612
15.9612
15.9612
22.6410
22.6410
32.1163
32.1163
32.1163
Reschi
Resraw
Resdev
-0.59212
1.40788
3.40788
-0.93243
2.06757
0.06757
-0.25217
-2.25217
0.03878
-2.96122
-1.96122
2.35897
1.35897
1.88366
-0.11634
-2.11634
-0.59212
1.40788
3.40788
-0.93243
2.06757
0.06757
-0.25217
-2.25217
0.03878
-2.96122
-1.96122
2.35897
1.35897
1.88366
-0.11634
-2.11634
-0.59212
1.40788
3.40788
-0.93243
2.06757
0.06757
-0.25217
-2.25217
0.03878
-2.96122
-1.96122
2.35897
1.35897
1.88366
-0.11634
-2.11634
Stdreschi Stdresdev
-0.34036
0.80928
1.95892
-0.54093
1.19947
0.03920
-0.14686
-1.31166
0.02249
-1.71738
-1.13743
1.37252
0.79069
1.22914
-0.07592
-1.38098
-0.34036
0.80928
1.95892
-0.54093
1.19947
0.03920
-0.14686
-1.31166
0.02249
-1.71738
-1.13743
1.37252
0.79069
1.22914
-0.07592
-1.38098
Reslik
-0.34036
0.80928
1.95892
-0.54093
1.19947
0.03920
-0.14686
-1.31166
0.02249
-1.71738
-1.13743
1.37252
0.79069
1.22914
-0.07592
-1.38098
Example 37.3: Gamma Distribution Applied to Life Data F 2579
The data set of predicted values and residuals (Output 37.2.2) is created by the OUTPUT statement.
You can use the PLOTS= option in the PROC GENMOD statement to create plots of predicted values
and residuals. Note that raw, Pearson, and deviance residuals are equal in this example. This is a
characteristic of the normal distribution and is not true in general for other distributions.
Example 37.3: Gamma Distribution Applied to Life Data
Life data are sometimes modeled with the gamma distribution. Although PROC GENMOD does not
analyze censored data or provide other useful lifetime distributions such as the Weibull or lognormal,
it can be used for modeling complete (uncensored) data with the gamma distribution, and it can
provide a statistical test for the exponential distribution against other gamma distribution alternatives.
See Lawless (2003) or Nelson (1982) for applications of the gamma distribution to life data.
The following data represent failure times of machine parts, some of which are manufactured by
manufacturer A and some by manufacturer B.
data A;
input [email protected]@
mfg = 'A';
datalines;
620 470 260 89
103 100 39
460
218 393 106 158
403 103 69
158
399 1274 32
12
548 381 203 871
317 85
1410 250
32
421 32
343
1792 47
95
76
1585 253 6
860
537 101 385 176
164 16
1267 352
1279 356 751 500
151 24
689 1119
763 555 14
45
;
run;
data B;
input [email protected]@
mfg = 'B';
datalines;
1747 945 12
1453
20
41
35
69
1090 1868 294 96
142 892 1307 310
403 860 23
406
561 348 130 13
317 304 79
1793
9
256 201 733
;
388
284
152
818
134
193
41
376
515
89
11
160
803
1733
776
242
1285
477
947
660
531
1101
1512
72
1055
565
195
560
2194
1
;
14
195
618
230
1054
230
536
510
150
89
44
30
1935
250
12
660
2580 F Chapter 37: The GENMOD Procedure
122
667
405
113
646
195
246
55
6
35
380
;
run;
27
761
998
25
575
1061
323
729
1566
181
609
273
1096
1409
940
219
174
198
813
459
147
546
1231
43
61
28
303
377
234
1216
946
116
182
44
278
848
304
388
39
1618
764
141
289
87
407
41
38
10
308
539
794
19
data lifdat;
set A B;
run;
The following SAS statements use PROC GENMOD to compute Type 3 statistics to test for differences between the two manufacturers in machine part life. Type 3 statistics are identical to Type 1
statistics in this case, since there is only one effect in the model. The log link function is selected to
ensure that the mean is positive.
proc genmod data = lifdat;
class mfg;
model lifetime = mfg / dist=gamma
link=log
type3;
run;
Example 37.3: Gamma Distribution Applied to Life Data F 2581
The output from these statements is displayed in Output 37.3.1.
Output 37.3.1 Gamma Model of Life Data
The GENMOD Procedure
Model Information
Data Set
Distribution
Link Function
Dependent Variable
WORK.LIFDAT
Gamma
Log
lifetime
Class Level Information
Class
Levels
mfg
Values
2
A B
Criteria For Assessing Goodness Of Fit
Criterion
Deviance
Scaled Deviance
Pearson Chi-Square
Scaled Pearson X2
Log Likelihood
Full Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)
DF
Value
Value/DF
199
199
199
199
287.0591
237.5335
211.6870
175.1652
-1432.4177
-1432.4177
2870.8353
2870.9572
2880.7453
1.4425
1.1936
1.0638
0.8802
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
Intercept
mfg
mfg
Scale
A
B
DF
Estimate
Standard
Error
1
1
0
1
6.1302
0.0199
0.0000
0.8275
0.1043
0.1559
0.0000
0.0714
Wald 95%
Confidence Limits
5.9257
-0.2857
0.0000
0.6987
6.3347
0.3255
0.0000
0.9800
Wald
Chi-Square
Pr > ChiSq
3451.61
0.02
.
<.0001
0.8985
.
NOTE: The scale parameter was estimated by maximum likelihood.
LR Statistics For Type 3 Analysis
Source
mfg
DF
ChiSquare
Pr > ChiSq
1
0.02
0.8985
The p-value of 0.8985 for the chi-square statistic in the Type 3 table indicates that there is no
significant difference in the part life between the two manufacturers.
2582 F Chapter 37: The GENMOD Procedure
Using the following statements, you can refit the model without using the manufacturer as an effect.
The LRCI option in the MODEL statement is specified to compute profile likelihood confidence
intervals for the mean life and scale parameters.
proc genmod data = lifdat;
model lifetime = / dist=gamma
link=log
lrci;
run;
Output 37.3.2 displays the results of fitting the model with the mfg effect omitted.
Output 37.3.2 Refitting of the Gamma Model: Omitting the mfg Effect
The GENMOD Procedure
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
DF
Estimate
Standard
Error
Intercept
Scale
1
1
6.1391
0.8274
0.0775
0.0714
Likelihood Ratio
95% Confidence
Limits
5.9904
0.6959
6.2956
0.9762
Wald
Chi-Square
Pr > ChiSq
6268.10
<.0001
NOTE: The scale parameter was estimated by maximum likelihood.
The intercept is the estimated log mean of the fitted gamma distribution, so that the mean life of the
parts is
D exp.INTERCEPT/ D exp.6:1391/ D 463:64
The SCALE parameter used in PROC GENMOD is the inverse of the gamma dispersion parameter,
and it is sometimes called the gamma index parameter. See the section “Response Probability
Distributions” on page 2510 for the definition of the gamma probability density function. A value of
1 for the index parameter corresponds to the exponential distribution . The estimated value of the
scale parameter is 0.8274. The 95% profile likelihood confidence interval for the scale parameter is
(0.6959, 0.9762), which does not contain 1. The hypothesis of an exponential distribution for the
data is, therefore, rejected at the 0.05 level. A confidence interval for the mean life is
.exp.5:99/; exp.6:30// D .399:57; 542:18/
Example 37.4: Ordinal Model for Multinomial Data
This example illustrates how you can use the GENMOD procedure to fit a model to data measured
on an ordinal scale. The following statements create a SAS data set called Icecream. The data set
contains the results of a hypothetical taste test of three brands of ice cream. The three brands are
rated for taste on a five-point scale from very good (vg) to very bad (vb). An analysis is performed to
Example 37.4: Ordinal Model for Multinomial Data F 2583
assess the differences in the ratings of the three brands. The variable taste contains the ratings, and
the variable brand contains the brands tested. The variable count contains the number of testers rating
each brand in each category.
The following statements create the Icecream data set:
data Icecream;
input count brand$ taste$;
datalines;
70 ice1 vg
71 ice1 g
151 ice1 m
30 ice1 b
46 ice1 vb
20 ice2 vg
36 ice2 g
130 ice2 m
74 ice2 b
70 ice2 vb
50 ice3 vg
55 ice3 g
140 ice3 m
52 ice3 b
50 ice3 vb
;
run;
The following statements fit a cumulative logit model to the ordinal data with the variable taste as
the response and the variable brand as a covariate. The variable count is used as a FREQ variable.
proc genmod data=Icecream rorder=data;
freq count;
class brand;
model taste = brand / dist=multinomial
link=cumlogit
aggregate=brand
type1;
estimate 'LogOR12' brand 1 -1 / exp;
estimate 'LogOR13' brand 1 0 -1 / exp;
estimate 'LogOR23' brand 0 1 -1 / exp;
run;
The AGGREGATE=BRAND option in the MODEL statement specifies the variable brand as defining
multinomial populations for computing deviances and Pearson chi-squares. The RORDER=DATA
option specifies that the taste variable levels be ordered by their order of appearance in the input
data set—that is, from very good (vg) to very bad (vb). By default, the response is sorted in
increasing ASCII order. Always check the “Response Profiles” table to verify that response levels are
appropriately ordered. The TYPE1 option requests a Type 1 test for the significance of the covariate
brand.
2584 F Chapter 37: The GENMOD Procedure
If j .x/ D Pr.taste j / is the cumulative probability of the j th or lower taste category, then the
odds ratio comparing x1 to x2 is as follows:
j .x1 /=.1
j .x2 /=.1
j .x1 //
D expŒ.x1
j .x2 //
x2 /0 ˇ
See McCullagh and Nelder (1989, Chapter 5) for details on the cumulative logit model. The
ESTIMATE statements compute log odds ratios comparing each of brands. The EXP option in the
ESTIMATE statements exponentiates the log odds ratios to form odds ratio estimates. Standard
errors and confidence intervals are also computed.
Output 37.4.1 displays general information about the model and data, the levels of the CLASS
variable brand, and the total number of occurrences of the ordered levels of the response variable
taste.
Output 37.4.1 Ordinal Model Information
The GENMOD Procedure
Model Information
Data Set
Distribution
Link Function
Dependent Variable
Frequency Weight Variable
WORK.ICECREAM
Multinomial
Cumulative Logit
taste
count
Class Level Information
Class
Levels
brand
3
Values
ice1 ice2 ice3
Response Profile
Ordered
Value
1
2
3
4
5
taste
vg
g
m
b
vb
Total
Frequency
140
162
421
156
166
Example 37.4: Ordinal Model for Multinomial Data F 2585
Output 37.4.2 displays estimates of the intercept terms and covariates and associated statistics. The
intercept terms correspond to the four cumulative logits defined on the taste categories in the order
1
/,
shown in Output 37.4.1. That is, Intercept1 is the intercept for the first cumulative logit, log. 1 pp
1
p Cp
1
2
Intercept2 is the intercept for the second cumulative logit, log. 1 .p
/, and so forth.
1 Cp2 /
Output 37.4.2 Parameter Estimates
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
Intercept1
Intercept2
Intercept3
Intercept4
brand
brand
brand
Scale
ice1
ice2
ice3
DF
Estimate
Standard
Error
1
1
1
1
1
1
0
0
-1.8578
-0.8646
0.9231
1.8078
0.3847
-0.6457
0.0000
1.0000
0.1219
0.1056
0.1060
0.1191
0.1370
0.1397
0.0000
0.0000
Wald 95% Confidence
Limits
-2.0967
-1.0716
0.7154
1.5743
0.1162
-0.9196
0.0000
1.0000
-1.6189
-0.6576
1.1308
2.0413
0.6532
-0.3719
0.0000
1.0000
Wald
Chi-Square
232.35
67.02
75.87
230.32
7.89
21.36
.
Analysis Of Maximum Likelihood
Parameter Estimates
Parameter
Intercept1
Intercept2
Intercept3
Intercept4
brand
brand
brand
Scale
Pr > ChiSq
ice1
ice2
ice3
<.0001
<.0001
<.0001
<.0001
0.0050
<.0001
.
NOTE: The scale parameter was held fixed.
The Type 1 test displayed in Output 37.4.3 indicates that Brand is highly significant; that is, there are
significant differences among the brands. The log odds ratios and odds ratios in the “ESTIMATE
Statement Results” table indicate the relative differences among the brands. For example, the odds
ratio of 2.8 in the “Exp(LogOR12)” row indicates that the odds of brand 1 being in lower taste
categories is 2.8 times the odds of brand 2 being in lower taste categories. Since, in this ordering,
the lower categories represent the more favorable taste results, this indicates that brand 1 scored
significantly better than brand 2. This is also apparent from the data in this example.
2586 F Chapter 37: The GENMOD Procedure
Output 37.4.3 Type 1 Tests and Odds Ratios
LR Statistics For Type 1 Analysis
Source
Intercepts
brand
Deviance
DF
ChiSquare
Pr > ChiSq
65.9576
9.8654
2
56.09
<.0001
Contrast Estimate Results
Mean
Estimate
Label
LogOR12
Exp(LogOR12)
LogOR13
Exp(LogOR13)
LogOR23
Exp(LogOR23)
Mean
Confidence Limits
0.7370
0.6805
0.7867
0.5950
0.5290
0.6577
0.3439
0.2850
0.4081
L'Beta
Estimate
Standard
Error
Alpha
1.0305
2.8024
0.3847
1.4692
-0.6457
0.5243
0.1401
0.3926
0.1370
0.2013
0.1397
0.0733
0.05
0.05
0.05
0.05
0.05
0.05
Contrast Estimate Results
Label
LogOR12
Exp(LogOR12)
LogOR13
Exp(LogOR13)
LogOR23
Exp(LogOR23)
L'Beta
Confidence Limits
0.7559
2.1295
0.1162
1.1233
-0.9196
0.3987
1.3050
3.6878
0.6532
1.9217
-0.3719
0.6894
ChiSquare
Pr > ChiSq
54.11
<.0001
7.89
0.0050
21.36
<.0001
Example 37.5: GEE for Binary Data with Logit Link Function
Output 37.5.1 displays a partial listing of a SAS data set of clinical trial data comparing two treatments
for a respiratory disorder. See “Gee Model for Binary Data” in the SAS/STAT Sample Program
Library for the complete data set. These data are from Stokes, Davis, and Koch (2000).
Patients in each of two centers are randomly assigned to groups receiving the active treatment or
a placebo. During treatment, respiratory status, represented by the variable outcome (coded here
as 0=poor, 1=good), is determined for each of four visits. The variables center, treatment, sex, and
baseline (baseline respiratory status) are classification variables with two levels. The variable age
(age at time of entry into the study) is a continuous variable.
Explanatory variables in the model are Intercept (xij1 ), treatment (xij 2 ), center (xij 3 ), sex (xij 4 ), age
(xij 5 ), and baseline (xij 6 ), so that x 0 D Œxij1 ; xij 2 ; : : : ; xij 6  is the vector of explanatory variables.
Indicator variables for the classification explanatory variables can be automatically generated by
listing them in the CLASS statement in PROC GENMOD. To be consistent with the analysis in
Stokes, Davis, and Koch (2000), the four classification explanatory variables are coded as follows
Example 37.5: GEE for Binary Data with Logit Link Function F 2587
via options in the CLASS statement:
xij 2 D
xij 4 D
0 placebo
1 active
0 male
1 female
0 center 1
1 center 2
xij 3 D
xij 6 D
00
11
Suppose yij represents the respiratory status of patient i at the j th visit, j D 1; : : : ; 4, and ij D
E.yij / represents the mean of the respiratory status. Since the response data are binary, you can use
the variance function for the binomial distribution v.ij / D ij .1 ij / and the logit link function
g.ij / D log.ij =.1 ij //. The model for the mean is g.ij / D xij 0 ˇ, where ˇ is a vector of
regression parameters to be estimated.
Output 37.5.1 Respiratory Disorder Data
O
b
s
c
e
n
t
e
r
i
d
t
r
e
a
t
m
e
n
t
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
3
3
3
3
4
4
4
4
5
5
5
5
P
P
P
P
P
P
P
P
A
A
A
A
P
P
P
P
P
P
P
P
s
e
x
M
M
M
M
M
M
M
M
M
M
M
M
M
M
M
M
F
F
F
F
a
g
e
b
a
s
e
l
i
n
e
v
i
s
i
t
1
v
i
s
i
t
2
v
i
s
i
t
3
v
i
s
i
t
4
v
i
s
i
t
o
u
t
c
o
m
e
46
46
46
46
28
28
28
28
23
23
23
23
44
44
44
44
13
13
13
13
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
1
1
1
1
0
0
0
0
1
1
1
1
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
0
1
1
1
1
The GEE solution is requested with the REPEATED statement in the GENMOD procedure. The
option SUBJECT=ID(CENTER) specifies that the observations in a single cluster be uniquely
2588 F Chapter 37: The GENMOD Procedure
identified by center and id within center. The option TYPE=UNSTR specifies the unstructured
working correlation structure. The MODEL statement specifies the regression model for the mean
with the binomial distribution variance function. The following SAS statements perform the GEE
model fit:
proc genmod data=resp descend;
class id treatment(ref="P") center(ref="1") sex(ref="M")
baseline(ref="0") / param=ref;
model outcome=treatment center sex age baseline / dist=bin;
repeated subject=id(center) / corr=unstr corrw;
run;
These statements first fit the generalized linear (GLM) model specified in the MODEL statement.
The parameter estimates from the generalized linear model fit are not shown in the output, but they
are used as initial values for the GEE solution. The DESCEND option in the PROC GENMOD
statement specifies that the probability that outcome D 1 be modeled. If the DESCEND option had
not been specified, the probability that outcome D 0 would be modeled by default.
Information about the GEE model is displayed in Output 37.5.2. The results of GEE model fitting
are displayed in Output 37.5.3. Model goodness-of-fit criteria are displayed in Output 37.5.4. If you
specify no other options, the standard errors, confidence intervals, Z scores, and p-values are based
on empirical standard error estimates. You can specify the MODELSE option in the REPEATED
statement to create a table based on model-based standard error estimates.
Output 37.5.2 Model Fitting Information
The GENMOD Procedure
GEE Model Information
Correlation Structure
Subject Effect
Number of Clusters
Correlation Matrix Dimension
Maximum Cluster Size
Minimum Cluster Size
Unstructured
id(center) (111 levels)
111
4
4
4
Output 37.5.3 Results of Model Fitting
Working Correlation Matrix
Row1
Row2
Row3
Row4
Col1
Col2
Col3
Col4
1.0000
0.3351
0.2140
0.2953
0.3351
1.0000
0.4429
0.3581
0.2140
0.4429
1.0000
0.3964
0.2953
0.3581
0.3964
1.0000
Example 37.6: Log Odds Ratios and the ALR Algorithm F 2589
Output 37.5.3 continued
Analysis Of GEE Parameter Estimates
Empirical Standard Error Estimates
Parameter
Intercept
treatment
center
sex
age
baseline
A
2
F
1
Estimate
Standard
Error
-0.8882
1.2442
0.6558
0.1128
-0.0175
1.8981
0.4568
0.3455
0.3512
0.4408
0.0129
0.3441
95% Confidence
Limits
-1.7835
0.5669
-0.0326
-0.7512
-0.0427
1.2237
0.0071
1.9214
1.3442
0.9768
0.0077
2.5725
Z Pr > |Z|
-1.94
3.60
1.87
0.26
-1.36
5.52
0.0519
0.0003
0.0619
0.7981
0.1728
<.0001
Output 37.5.4 Model Fit Criteria
GEE Fit Criteria
QIC
QICu
512.3416
499.6081
The nonsignificance of age and sex make them candidates for omission from the model.
Example 37.6: Log Odds Ratios and the ALR Algorithm
Since the respiratory data in Example 37.5 are binary, you can use the ALR algorithm to model the
log odds ratios instead of using working correlations to model associations. In this example, a “fully
parameterized cluster” model for the log odds ratio is fit. That is, there is a log odds ratio parameter
for each unique pair of responses within clusters, and all clusters are parameterized identically.
The following statements fit the same regression model for the mean as in Example 37.5 but use a
regression model for the log odds ratios instead of a working correlation. The LOGOR=FULLCLUST
option specifies a fully parameterized log odds ratio model.
proc genmod data=resp descend;
class id treatment(ref="P") center(ref="1") sex(ref="M")
baseline(ref="0") / param=ref;
model outcome=treatment center sex age baseline / dist=bin;
repeated subject=id(center) / logor=fullclust;
run;
2590 F Chapter 37: The GENMOD Procedure
The results of fitting the model are displayed in Output 37.6.1 along with a table that shows the
correspondence between the log odds ratio parameters and the within-cluster pairs. Model goodnessof-fit criteria are shown in Output 37.6.2. The QIC for the ALR model shown in Output 37.6.2 is
511.86, whereas the QIC for the unstructured working correlation model shown in Output 37.5.4 is
512.34, indicating that the ALR model is a slightly better fit.
Output 37.6.1 Results of Model Fitting
The GENMOD Procedure
Log Odds Ratio
Parameter Information
Parameter
Group
Alpha1
Alpha2
Alpha3
Alpha4
Alpha5
Alpha6
(1,
(1,
(1,
(2,
(2,
(3,
2)
3)
4)
3)
4)
4)
Analysis Of GEE Parameter Estimates
Empirical Standard Error Estimates
Parameter
Intercept
treatment
center
sex
age
baseline
Alpha1
Alpha2
Alpha3
Alpha4
Alpha5
Alpha6
A
2
F
1
Estimate
Standard
Error
-0.9266
1.2611
0.6287
0.1024
-0.0162
1.8980
1.6109
1.0771
1.5875
2.1224
1.8818
2.1046
0.4513
0.3406
0.3486
0.4362
0.0125
0.3404
0.4892
0.4834
0.4735
0.5022
0.4686
0.4949
95% Confidence
Limits
-1.8111
0.5934
-0.0545
-0.7526
-0.0407
1.2308
0.6522
0.1297
0.6594
1.1381
0.9634
1.1347
-0.0421
1.9287
1.3119
0.9575
0.0084
2.5652
2.5696
2.0246
2.5155
3.1068
2.8001
3.0745
Output 37.6.2 Model Fit Criteria
GEE Fit Criteria
QIC
QICu
511.8589
499.6516
Z Pr > |Z|
-2.05
3.70
1.80
0.23
-1.29
5.58
3.29
2.23
3.35
4.23
4.02
4.25
0.0400
0.0002
0.0713
0.8144
0.1977
<.0001
0.0010
0.0259
0.0008
<.0001
<.0001
<.0001
Example 37.6: Log Odds Ratios and the ALR Algorithm F 2591
You can fit the same model by fully specifying the z matrix. The following statements create a data
set containing the full z matrix:
data zin;
keep id center z1-z6 y1 y2;
array zin(6) z1-z6;
set resp ;
by center id;
if first.id
then do;
t = 0;
do m = 1 to 4;
do n = m+1 to 4;
do j = 1 to 6;
zin(j) = 0;
end;
y1 = m;
y2 = n;
t + 1;
zin(t) = 1;
output;
end;
end;
end;
run;
proc print data=zin (obs=12);
Output 37.6.3 displays the full z matrix for the first two clusters. The z matrix is identical for all
clusters in this example.
Output 37.6.3 Full z Matrix Data Set
Obs
z1
z2
z3
z4
z5
z6
center
id
y1
y2
1
2
3
4
5
6
7
8
9
10
11
12
1
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
1
1
1
2
2
3
1
1
1
2
2
3
2
3
4
3
4
4
2
3
4
3
4
4
2592 F Chapter 37: The GENMOD Procedure
The following statements fit the model for fully parameterized clusters by fully specifying the z
matrix. The results are identical to those shown previously.
proc genmod data=resp descend;
class id treatment(ref="P") center(ref="1") sex(ref="M")
baseline(ref="0") / param=ref;
model outcome=treatment center sex age baseline / dist=bin;
repeated subject=id(center) / logor=zfull
zdata=zin
zrow =(z1-z6)
ypair=(y1 y2) ;
run;
Example 37.7: Log-Linear Model for Count Data
In this example the data, from Thall and Vail (1990), concern the treatment of people suffering from
epileptic seizure episodes. These data are also analyzed in Diggle, Liang, and Zeger (1994). The data
consist of the number of epileptic seizures in an eight-week baseline period, before any treatment,
and in each of four two-week treatment periods, in which patients received either a placebo or the
drug Progabide in addition to other therapy. A portion of the data is displayed in Table 37.12. See
“Gee Model for Count Data, Exchangeable Correlation” in the SAS/STAT Sample Program Library
for the complete data set.
Table 37.12
Epileptic Seizure Data
Patient ID
Treatment
Baseline
Visit1
Visit2
Visit3
Visit4
104
106
107
.
.
.
101
102
103
.
.
.
Placebo
Placebo
Placebo
11
11
6
5
3
2
3
5
4
3
3
0
3
3
5
Progabide
Progabide
Progabide
76
38
19
11
8
0
14
7
4
9
9
3
8
4
0
Model the data as a log-linear model with V ./ D (the Poisson variance function) and
log.E.Yij // D ˇ0 C xi1 ˇ1 C xi 2 ˇ2 C xi1 xi 2 ˇ3 C log.tij /
where
Yij D number of epileptic seizures in interval j
Example 37.7: Log-Linear Model for Count Data F 2593
tij D length of interval j
1 W weeks 8 16 (treatment)
xi1 D
0 W weeks 0 8 (baseline)
1 W progabide group
xi 2 D
0 W placebo group
The correlations between the counts are modeled as rij D ˛, i ¤ j (exchangeable correlations). For
comparison, the correlations are also modeled as independent (identity correlation matrix). In this
model, the regression parameters have the interpretation in terms of the log seizure rate displayed in
Table 37.13.
Table 37.13 Interpretation of Regression Parameters
Treatment
Visit
Placebo
Baseline
1–4
Baseline
1–4
Progabide
log.E.Yij /=tij /
ˇ0
ˇ0 C ˇ1
ˇ0 C ˇ2
ˇ0 C ˇ1 C ˇ2 C ˇ3
The difference between the log seizure rates in the pretreatment (baseline) period and the treatment
periods is ˇ1 for the placebo group and ˇ1 C ˇ3 for the Progabide group. A value of ˇ3 < 0 indicates
a reduction in the seizure rate.
Output 37.7.1 lists the first 14 observations of the data, which are arranged as one visit per observation:
Output 37.7.1 Partial Listing of the Seizure Data
Obs
id
y
visit
trt
bline
age
1
2
3
4
5
6
7
8
9
10
11
12
13
14
104
104
104
104
106
106
106
106
107
107
107
107
114
114
5
3
3
3
3
5
3
3
2
4
0
5
4
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
11
11
11
11
11
11
11
11
6
6
6
6
8
8
31
31
31
31
30
30
30
30
25
25
25
25
36
36
Some further data manipulations create an observation for the baseline measures, a log time interval
variable for use as an offset, and an indicator variable for whether the observation is for a baseline
measurement or a visit measurement. Patient 207 is deleted as an outlier, as in the Diggle, Liang,
and Zeger (1994) analysis.
2594 F Chapter 37: The GENMOD Procedure
The following statements prepare the data for analysis with PROC GENMOD:
data new;
set thall;
output;
if visit=1 then do;
y=bline;
visit=0;
output;
end;
run;
data new;
set new;
if id ne 207;
if visit=0 then do;
x1=0;
ltime=log(8);
end;
else do;
x1=1;
ltime=log(2);
end;
run;
For comparison with the GEE results, an ordinary Poisson regression is first fit. The results are
shown in Output 37.7.2.
Output 37.7.2 Maximum Likelihood Estimates
The GENMOD Procedure
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
DF
Estimate
Standard
Error
Intercept
x1
trt
x1*trt
Scale
1
1
1
1
0
1.3476
0.1108
-0.1080
-0.3016
1.0000
0.0341
0.0469
0.0486
0.0697
0.0000
Wald 95%
Confidence Limits
1.2809
0.0189
-0.2034
-0.4383
1.0000
1.4144
0.2027
-0.0127
-0.1649
1.0000
Wald
Chi-Square
Pr > ChiSq
1565.44
5.58
4.93
18.70
<.0001
0.0181
0.0264
<.0001
NOTE: The scale parameter was held fixed.
The GEE solution is requested with the REPEATED statement in the GENMOD procedure. The
SUBJECT=ID option indicates that the variable id describes the observations for a single cluster,
and the CORRW option displays the working correlation matrix. The TYPE= option specifies the
correlation structure; the value EXCH indicates the exchangeable structure.
Example 37.7: Log-Linear Model for Count Data F 2595
The following statements perform the analysis:
proc genmod data=new;
class id;
model y=x1 | trt / d=poisson offset=ltime;
repeated subject=id / corrw covb type=exch;
run;
These statements first fit a generalized linear model (GLM) to these data by maximum likelihood.
The estimates are not shown in the output, but are used as initial values for the GEE solution.
Information about the GEE model is displayed in Output 37.7.3. The results of fitting the model
are displayed in Output 37.7.4. Compare these with the model of independence displayed in Output 37.7.2. The parameter estimates are nearly identical, but the standard errors for the independence
case are underestimated. The coefficient of the interaction term, ˇ3 , is highly significant under the
independence model and marginally significant with the exchangeable correlations model.
Output 37.7.3 GEE Model Information
The GENMOD Procedure
GEE Model Information
Correlation Structure
Subject Effect
Number of Clusters
Correlation Matrix Dimension
Maximum Cluster Size
Minimum Cluster Size
Exchangeable
id (58 levels)
58
5
5
5
Output 37.7.4 GEE Parameter Estimates
Analysis Of GEE Parameter Estimates
Empirical Standard Error Estimates
Parameter Estimate
Intercept
x1
trt
x1*trt
1.3476
0.1108
-0.1080
-0.3016
Standard
Error
0.1574
0.1161
0.1937
0.1712
95% Confidence
Limits
1.0392
-0.1168
-0.4876
-0.6371
1.6560
0.3383
0.2716
0.0339
Z Pr > |Z|
8.56
0.95
-0.56
-1.76
<.0001
0.3399
0.5770
0.0781
2596 F Chapter 37: The GENMOD Procedure
Table 37.14 displays the regression coefficients, standard errors, and normalized coefficients that
result from fitting the model with independent and exchangeable working correlation matrices.
Table 37.14
Results of Model Fitting
Variable
Correlation Structure
Intercept
Exchangeable
Independent
Exchangeable
Independent
Exchangeable
Independent
Exchangeable
Independent
Visit .x1 /
Treat .x2 /
x1 x2
Coef.
Std. Error
Coef./S.E.
1.35
1.35
0.11
0.11
0.11
0.11
0.30
0.30
0.16
0.03
0.12
0.05
0.19
0.05
0.17
0.07
8.56
39.52
0.95
2.36
0.56
2.22
1.76
4.32
The fitted exchangeable correlation matrix is specified with the CORRW option and is displayed in
Output 37.7.5.
Output 37.7.5 Working Correlation Matrix
Working Correlation Matrix
Row1
Row2
Row3
Row4
Row5
Col1
Col2
Col3
Col4
Col5
1.0000
0.5941
0.5941
0.5941
0.5941
0.5941
1.0000
0.5941
0.5941
0.5941
0.5941
0.5941
1.0000
0.5941
0.5941
0.5941
0.5941
0.5941
1.0000
0.5941
0.5941
0.5941
0.5941
0.5941
1.0000
If you specify the COVB option, you produce both the model-based (naive) and the empirical (robust)
covariance matrices. Output 37.7.6 contains these estimates.
Output 37.7.6 Covariance Matrices
Covariance Matrix (Model-Based)
Prm1
Prm2
Prm3
Prm4
Prm1
Prm2
Prm3
Prm4
0.01223
0.001520
-0.01223
-0.001520
0.001520
0.01519
-0.001520
-0.01519
-0.01223
-0.001520
0.02495
0.005427
-0.001520
-0.01519
0.005427
0.03748
Example 37.8: Model Assessment of Multiple Regression Using Aggregates of Residuals F 2597
Output 37.7.6 continued
Covariance Matrix (Empirical)
Prm1
Prm2
Prm3
Prm4
Prm1
Prm2
Prm3
Prm4
0.02476
-0.001152
-0.02476
0.001152
-0.001152
0.01348
0.001152
-0.01348
-0.02476
0.001152
0.03751
-0.002999
0.001152
-0.01348
-0.002999
0.02931
The two covariance estimates are similar, indicating an adequate correlation model.
Example 37.8: Model Assessment of Multiple Regression Using
Aggregates of Residuals
This example illustrates the use of cumulative residuals to assess the adequacy of a normal linear
regression model. Neter et al. (1996, Section 8.2) describe a study of 54 patients undergoing a certain
kind of liver operation in a surgical unit. The data consist of the survival time and certain covariates.
After a model selection procedure, they arrived at the following model:
Y D ˇ0 C ˇ1 X1 C ˇ2 X2 C ˇ3 X3 C where Y is the logarithm (base 10) of the survival time; X1 , X2 , X3 are blood-clotting score,
prognostic index, and enzyme function, respectively; and is a normal error term. A listing of
the SAS data set containing the data is shown in Output 37.8.1. The variables Y, X1, X2, and X3
correspond to Y , X1 , X2 , and X3 , and LogX1 is log(X1 ). The PROC GENMOD fit of the model is
shown in Output 37.8.2. The analysis first focuses on the adequacy of the functional form of X1 ,
blood-clotting score.
2598 F Chapter 37: The GENMOD Procedure
Output 37.8.1 Surgical Unit Example Data
Obs
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
Y
X1
X2
X3
LogX1
2.3010
2.0043
2.3096
2.0043
2.7067
1.9031
1.9031
2.1038
2.3054
2.3075
2.5172
1.8129
2.9191
2.5185
2.2253
2.3365
1.9395
1.5315
2.3324
2.2355
2.0374
2.1335
1.8451
2.3424
2.4409
2.1584
2.2577
2.7589
1.8573
2.2504
1.8513
1.7634
2.0645
2.4698
2.0607
2.2648
2.0719
2.0792
2.1790
2.1703
1.9777
1.8751
2.6840
2.1847
2.2810
2.0899
2.4928
2.5999
2.1987
2.4914
2.0934
2.0969
2.2967
2.4955
6.7
5.1
7.4
6.5
7.8
5.8
5.7
3.7
6.0
3.7
6.3
6.7
5.8
5.8
7.7
7.4
6.0
3.7
7.3
5.6
5.2
3.4
6.7
5.8
6.3
5.8
5.2
11.2
5.2
5.8
3.2
8.7
5.0
5.8
5.4
5.3
2.6
4.3
4.8
5.4
5.2
3.6
8.8
6.5
3.4
6.5
4.5
4.8
5.1
3.9
6.6
6.4
6.4
8.8
62
59
57
73
65
38
46
68
67
76
84
51
96
83
62
74
85
51
68
57
52
83
26
67
59
61
52
76
54
76
64
45
59
72
58
51
74
8
61
52
49
28
86
56
77
40
73
86
67
82
77
85
59
78
81
66
83
41
115
72
63
81
93
94
83
43
114
88
67
68
28
41
74
87
76
53
68
86
100
73
86
90
56
59
65
23
73
93
70
99
86
119
76
88
72
99
88
77
93
84
106
101
77
103
46
40
85
72
0.82607
0.70757
0.86923
0.81291
0.89209
0.76343
0.75587
0.56820
0.77815
0.56820
0.79934
0.82607
0.76343
0.76343
0.88649
0.86923
0.77815
0.56820
0.86332
0.74819
0.71600
0.53148
0.82607
0.76343
0.79934
0.76343
0.71600
1.04922
0.71600
0.76343
0.50515
0.93952
0.69897
0.76343
0.73239
0.72428
0.41497
0.63347
0.68124
0.73239
0.71600
0.55630
0.94448
0.81291
0.53148
0.81291
0.65321
0.68124
0.70757
0.59106
0.81954
0.80618
0.80618
0.94448
Example 37.8: Model Assessment of Multiple Regression Using Aggregates of Residuals F 2599
In order to assess the adequacy of the fitted multiple regression model, the ASSESS statement in the
following SAS statements is used to create the plots of cumulative residuals against X1 shown in
Output 37.8.3 and Output 37.8.4 and the summary table in Output 37.8.5:
ods graphics on;
proc genmod data=Surg;
model Y = X1 X2 X3 / scale=Pearson;
assess var=(X1) / resample=10000
seed=603708000
crpanel ;
run;
Output 37.8.2 Regression Model for Linear X1
The GENMOD Procedure
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
DF
Estimate
Standard
Error
Intercept
X1
X2
X3
Scale
1
1
1
1
0
0.4836
0.0692
0.0093
0.0095
0.0469
0.0426
0.0041
0.0004
0.0003
0.0000
Wald 95%
Confidence Limits
0.4001
0.0612
0.0085
0.0089
0.0469
0.5672
0.0772
0.0100
0.0101
0.0469
Wald
Chi-Square
Pr > ChiSq
128.71
288.17
590.45
966.07
<.0001
<.0001
<.0001
<.0001
NOTE: The scale parameter was estimated by the square root of Pearson's
Chi-Square/DOF.
See Lin, Wei, and Ying (2002) for details about model assessment that uses cumulative residual
plots. The RESAMPLE= keyword specifies that a p-value be computed based on a sample of
10,000 simulated residual paths. A random number seed is specified by the SEED= keyword for
reproducibility. If you do not specify the seed, one is derived from the time of day. The keyword
CRPANEL specifies that the panel of four cumulative residual plots shown in Output 37.8.4 be created,
each with two simulated paths. The single residual plot with 20 simulated paths in Output 37.8.3 is
created by default.
These graphical displays are requested by specifying the ODS GRAPHICS statement and the ASSESS
statement. For general information about ODS Graphics, see Chapter 21, “Statistical Graphics Using
ODS.” For specific information about the graphics available in the GENMOD procedure, see the
section “ODS Graphics” on page 2572.
2600 F Chapter 37: The GENMOD Procedure
Output 37.8.3 Cumulative Residual Plot for Linear X1 Fit
Example 37.8: Model Assessment of Multiple Regression Using Aggregates of Residuals F 2601
Output 37.8.4 Cumulative Residual Panel Plot for Linear X1 Fit
Output 37.8.5 Summary of Model Assessment
Assessment Summary
Assessment
Variable
X1
Maximum
Absolute
Value
Replications
Seed
Pr >
MaxAbsVal
0.0380
10000
603708000
0.1084
The p-value of 0.1084 reported on Output 37.8.3 and Output 37.8.5 suggests that a more adequate
model might be possible. The observed cumulative residuals in Output 37.8.3 and Output 37.8.4,
represented by the heavy lines, seem atypical of the simulated curves, represented by the light lines,
reinforcing the conclusion that a more appropriate functional form for X1 is possible.
The cumulative residual plots in Output 37.8.6 provide guidance in determining a more appropriate
functional form. The four curves were created from simple forms of model misspecification by using
simulated data. The mean models of the data and the fitted model are shown in Table 37.15.
2602 F Chapter 37: The GENMOD Procedure
Output 37.8.6 Typical Cumulative Residual Patterns
Table 37.15
Model Misspecifications
Plot
Data E(Y )
Fitted Model E(Y )
(a)
(b)
(c)
(d)
log(X)
X C X2
X C X2 C X3
I.X > 5/
X
X
X C X2
X
The observed cumulative residual pattern in Output 37.8.3 and Output 37.8.4 most resembles the
behavior of the curve in plot (a) of Output 37.8.6, indicating that log(X1 ) might be a more appropriate
term in the model than X1 .
Example 37.8: Model Assessment of Multiple Regression Using Aggregates of Residuals F 2603
The following SAS statements fit a model with LogX1 in place of X1 and request a model assessment:
proc genmod data=Surg;
model Y = LogX1 X2 X3 / scale=Pearson;
assess var=(LogX1) / resample=10000
seed=603708000;
run;
ods graphics off;
The revised model fit is shown in Output 37.8.7, the p-value from the simulation is 0.4777, and the
cumulative residuals plotted in Output 37.8.8 show no systematic trend. The log transformation for
X1 is more appropriate. Under the revised model, the p-values for testing the functional forms of X2
and X3 are 0.20 and 0.63, respectively; and the p-value for testing the linearity of the model is 0.65.
Thus, the revised model seems reasonable.
Output 37.8.7 Multiple Regression Model with Log(X1)
The GENMOD Procedure
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
DF
Estimate
Standard
Error
Intercept
LogX1
X2
X3
Scale
1
1
1
1
0
0.1844
0.9121
0.0095
0.0096
0.0434
0.0504
0.0491
0.0004
0.0003
0.0000
Wald 95%
Confidence Limits
0.0857
0.8158
0.0088
0.0090
0.0434
0.2832
1.0083
0.0102
0.0101
0.0434
Wald
Chi-Square
Pr > ChiSq
13.41
345.05
728.62
1139.73
0.0003
<.0001
<.0001
<.0001
NOTE: The scale parameter was estimated by the square root of Pearson's
Chi-Square/DOF.
2604 F Chapter 37: The GENMOD Procedure
Output 37.8.8 Cumulative Residual Plot with Log(X1)
Example 37.9: Assessment of a Marginal Model for Dependent Data
This example illustrates the use of cumulative residuals to assess the adequacy of a marginal model
for dependent data fit by generalized estimating equations (GEEs). The assessment methods are
applied to CD4 count data from an AIDS clinical trial reported by Fischl, Richman, and Hansen
(1990) and reanalyzed by Lin, Wei, and Ying (2002). The study randomly assigned 360 HIV patients
to the drug AZT and 351 patients to placebo. CD4 counts were measured repeatedly over the course
of the study. The data used here are the 4328 measurements taken in the first 40 weeks of the study.
The analysis focuses on the time trend of the response. The first model considered is
E.yi k / D ˇ0 C ˇ1 Ti k C ˇ2 Ti2k C ˇ3 Ri Ti k C ˇ4 Ri Ti2k
where Ti k is the time (in weeks) of the kth measurement on the i th patient, yi k is the CD4 count
at Tik for the i th patient, and Ri is the indicator of AZT for the i th patient. Normal errors and an
independent working correlation are assumed.
Example 37.9: Assessment of a Marginal Model for Dependent Data F 2605
The following statements create the SAS data set cd4:
data cd4;
input Id Y Time Time2 TrtTime TrtTime2;
Time3 = Time2 * Time;
TrtTime3 = TrtTime2 * Time;
datalines;
1
264.00024
-0.28571
0.08163
1
175.00070
4.14286
17.16327
1
306.00150
8.14286
66.30612
1
331.99835
12.14286
147.44898
1
309.99929
16.14286
260.59184
1
185.00077
28.71429
824.51020
1
175.00070
40.14286
1611.44898
-0.28571
4.14286
8.14286
12.14286
16.14286
28.71429
40.14286
0.08163
17.16327
66.30612
147.44898
260.59184
824.51020
1611.44898
... more lines ...
711
711
;
run;
488.00224
240.00026
12.14286
18.14286
147.44898
329.16327
12.14286
18.14286
147.44898
329.16327
The following SAS statements fit the preceding model, create the cumulative residual plot in
Output 37.9.1, and compute a p-value for the model.
These graphical displays are requested by specifying the ODS GRAPHICS statement and the ASSESS
statement. For general information about ODS Graphics, see Chapter 21, “Statistical Graphics Using
ODS.” For specific information about the graphics available in the GENMOD procedure, see the
section “ODS Graphics” on page 2572.
Here, the SAS data set variables Time, Time2, TrtTime, and TrtTime2 correspond to Ti k , Ti2k , Ri Ti k ,
and Ri Ti2k , respectively. The variable Id identifies individual patients.
ods graphics on;
proc genmod data=cd4;
class Id;
model Y = Time Time2 TrtTime TrtTime2;
repeated sub=Id;
assess var=(Time) / resample
seed=603708000;
run;
ods graphics off;
2606 F Chapter 37: The GENMOD Procedure
Output 37.9.1 Cumulative Residual Plot for Quadratic Time Fit
The cumulative residual plot in Output 37.9.1 displays cumulative residuals versus time for the model
and 20 simulated realizations. The associated p-value, also shown in Output 37.9.1, is 0.18. These
results indicate that a more satisfactory model might be possible. The observed cumulative residual
pattern most resembles plot (c) in Output 37.8.6, suggesting cubic time trends.
The following SAS statements fit the model, create the plot in Output 37.9.2, and compute a p-value
for a model with the additional terms Ti3k and Ri Ti3k :
ods graphics on;
proc genmod data=cd4;
class Id;
model Y = Time Time2 Time3 TrtTime TrtTime2 TrtTime3;
repeated sub=Id;
assess var=(Time) / resample
seed=603708000;
run;
ods graphics off;
Example 37.9: Assessment of a Marginal Model for Dependent Data F 2607
Output 37.9.2 Cumulative Residual Plot for Cubic Time Fit
The observed cumulative residual pattern appears more typical of the simulated realizations, and the
p-value is 0.45, indicating that the model with cubic time trends is more appropriate.
2608 F Chapter 37: The GENMOD Procedure
Example 37.10: Bayesian Analysis of a Poisson Regression Model
This example illustrates a Bayesian analysis of a log-linear Poisson regression model. Consider
the following data on patients from clinical trials. The data set is a subset of the data described in
Ibrahim, Chen, and Lipsitz (1999).
data Liver;
input X1-X6 Y;
datalines;
19.1358
50.0110
23.5970
18.4959
20.0474
56.7699
28.0277
59.7836
28.6851
74.1589
18.8092
31.0630
28.7201
52.9178
21.3669
61.6603
23.7332
42.2904
51.000
3.429
3.429
4.000
5.714
2.286
37.286
54.143
0.571
0
0
1
0
1
0
1
0
1
0
0
1
0
0
1
0
1
0
1
1
0
1
1
1
1
1
1
3
9
6
6
1
61
6
6
21
2.571
4.429
1
1
0
0
0
0
1
6
... more lines ...
19.1327
17.3010
;
run;
65.3425
51.4493
The primary interest is in prediction of the number of cancerous liver nodes when a patient enters
the trials, by using six other baseline characteristics. The number of nodes is modeled by a Poisson
regression model with the six baseline characteristics as covariates. The response and regression
variables are as follows:
Y
X1
X2
X3
X4
X5
X6
Number of Cancerous Liver Nodes
Body Mass Index
Age, in Years
Time Since Diagnosis of Disease, in Weeks
Two Biochemical Markers (each classified as normal=1 or abnormal=0)
Anti Hepatitis B Antigen
Associated Jaundice (yes=1, no=0)
Two analyses are performed using PROC GENMOD. The first analysis uses noninformative normal
prior distributions, and the second analysis uses an informative normal prior for one of the regression
parameters.
In the following BAYES statement, COEFFPRIOR=NORMAL specifies a noninformative independent normal prior distribution with zero mean and variance 106 for each parameter.
The initial analysis is performed using PROC GENMOD to obtain Bayesian estimates of the
regression coefficients by using the following SAS statements:
Example 37.10: Bayesian Analysis of a Poisson Regression Model F 2609
ods graphics ON;
proc genmod data=Liver;
model Y = X1-X6 / dist=Poisson link=log;
bayes seed=1 coeffprior=normal;
run;
Maximum likelihood estimates of the model parameters are computed by default. These are shown
in the “Analysis of Maximum Likelihood Parameter Estimates” table in Output 37.10.1.
Output 37.10.1 Maximum Likelihood Parameter Estimates
The GENMOD Procedure
Bayesian Analysis
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
DF
Estimate
Standard
Error
Intercept
X1
X2
X3
X4
X5
X6
Scale
1
1
1
1
1
1
1
0
2.4508
-0.0044
-0.0135
-0.0029
-0.2715
0.3215
0.2077
1.0000
0.2284
0.0080
0.0024
0.0022
0.0795
0.0832
0.0827
0.0000
Wald 95% Confidence
Limits
2.0032
-0.0201
-0.0181
-0.0072
-0.4272
0.1585
0.0456
1.0000
2.8984
0.0114
-0.0088
0.0014
-0.1157
0.4845
0.3698
1.0000
NOTE: The scale parameter was held fixed.
Noninformative independent normal prior distributions with zero means and variances of 106 were
used in the initial analysis. These are shown in Output 37.10.2.
Output 37.10.2 Regression Coefficient Priors
The GENMOD Procedure
Bayesian Analysis
Independent Normal Prior for Regression Coefficients
Parameter
Mean
Precision
Intercept
X1
X2
X3
X4
X5
X6
0
0
0
0
0
0
0
1E-6
1E-6
1E-6
1E-6
1E-6
1E-6
1E-6
2610 F Chapter 37: The GENMOD Procedure
Initial values for the Markov chain are listed in the “Initial Values and Seeds” table in Output 37.10.3.
The random number seed is also listed so that you can reproduce the analysis. Since no seed was
specified, the seed shown was derived from the time of day.
Output 37.10.3 MCMC Initial Values and Seeds
Initial Values of the Chain
Chain
Seed
Intercept
X1
X2
X3
X4
1
1
2.450813
-0.00435
-0.01347
-0.00291
-0.27149
Initial Values of the Chain
X5
X6
0.321507
0.207713
Summary statistics for the posterior sample are displayed in the “Fit Statistics,” “Descriptive Statistics
for the Posterior Sample,” “Interval Statistics for the Posterior Sample,” and “Posterior Correlation
Matrix” tables in Output 37.10.4, Output 37.10.5, Output 37.10.6, and Output 37.10.7, respectively.
Since noninformative prior distributions for the regression coefficients were used, the mean and
standard deviations of the posterior distributions for the model parameters are close to the maximum
likelihood estimates and standard errors.
Output 37.10.4 Fit Statistics
Fit Statistics
DIC (smaller is better)
pD (effective number of parameters)
829.729
6.966
Output 37.10.5 Descriptive Statistics
The GENMOD Procedure
Bayesian Analysis
Posterior Summaries
Parameter
N
Mean
Standard
Deviation
25%
Intercept
X1
X2
X3
X4
X5
X6
10000
10000
10000
10000
10000
10000
10000
2.4520
-0.00473
-0.0134
-0.00309
-0.2705
0.3193
0.2095
0.2268
0.00801
0.00236
0.00220
0.0792
0.0834
0.0834
2.2997
-0.0100
-0.0150
-0.00455
-0.3241
0.2629
0.1538
Percentiles
50%
2.4521
-0.00465
-0.0134
-0.00305
-0.2697
0.3180
0.2086
75%
2.6053
0.000759
-0.0118
-0.00158
-0.2172
0.3762
0.2653
Example 37.10: Bayesian Analysis of a Poisson Regression Model F 2611
Output 37.10.6 Interval Statistics
Posterior Intervals
Parameter
Alpha
Intercept
X1
X2
X3
X4
X5
X6
0.050
0.050
0.050
0.050
0.050
0.050
0.050
Equal-Tail Interval
2.0169
-0.0210
-0.0181
-0.00757
-0.4250
0.1552
0.0477
HPD Interval
2.9056
0.0106
-0.00878
0.00109
-0.1132
0.4821
0.3749
2.0069
-0.0212
-0.0181
-0.00764
-0.4232
0.1647
0.0490
2.8923
0.0103
-0.00885
0.000989
-0.1119
0.4905
0.3758
Output 37.10.7 Posterior Sample Correlation Matrix
Posterior Correlation Matrix
Parameter
Intercept
X1
X2
X3
X4
X5
X6
Intercept
X1
X2
X3
X4
X5
X6
1.000
-0.705
-0.430
-0.046
-0.225
-0.180
-0.415
-0.705
1.000
-0.211
-0.013
-0.068
0.067
0.128
-0.430
-0.211
1.000
-0.006
0.070
0.057
0.118
-0.046
-0.013
-0.006
1.000
0.016
-0.055
-0.089
-0.225
-0.068
0.070
0.016
1.000
-0.011
0.089
-0.180
0.067
0.057
-0.055
-0.011
1.000
-0.042
-0.415
0.128
0.118
-0.089
0.089
-0.042
1.000
Posterior sample autocorrelations for each model parameter are shown in Output 37.10.8. The
autocorrelation after 10 lags is negligible for all parameters, indicating good mixing in the Markov
chain.
Output 37.10.8 Posterior Sample Autocorrelations
The GENMOD Procedure
Bayesian Analysis
Posterior Autocorrelations
Parameter
Lag 1
Lag 5
Lag 10
Lag 50
Intercept
X1
X2
X3
X4
X5
X6
0.0551
0.0894
0.1197
0.0324
0.0309
0.0402
0.0696
-0.0134
-0.0054
-0.0170
-0.0036
0.0056
0.0015
-0.0047
-0.0101
-0.0080
0.0061
-0.0033
0.0053
-0.0111
-0.0024
0.0012
0.0019
0.0006
-0.0160
0.0115
0.0123
0.0006
2612 F Chapter 37: The GENMOD Procedure
The p-values for the Geweke test statistics shown in Output 37.10.9 all indicate convergence of the
MCMC. See the section “Assessing Markov Chain Convergence” on page 155 for more information
about convergence diagnostics and their interpretation.
Output 37.10.9 Geweke Diagnostic Statistics
Geweke Diagnostics
Parameter
z
Pr > |z|
Intercept
X1
X2
X3
X4
X5
X6
0.9855
-1.0835
-0.3847
0.6715
0.1328
1.0698
-0.1647
0.3244
0.2786
0.7005
0.5019
0.8943
0.2847
0.8692
The effective sample sizes for each parameter are shown in Output 37.10.10.
Output 37.10.10 Effective Sample Sizes
Effective Sample Sizes
Parameter
ESS
Autocorrelation
Time
Efficiency
Intercept
X1
X2
X3
X4
X5
X6
9245.8
8179.5
8067.8
9390.6
9157.6
9665.2
8778.7
1.0816
1.2226
1.2395
1.0649
1.0920
1.0346
1.1391
0.9246
0.8179
0.8068
0.9391
0.9158
0.9665
0.8779
Example 37.10: Bayesian Analysis of a Poisson Regression Model F 2613
Trace, autocorrelation, and density plots for the seven model parameters are shown in Output 37.10.11
through Output 37.10.17. All indicate satisfactory convergence of the Markov chain.
Output 37.10.11 Diagnostic Plots for Intercept
2614 F Chapter 37: The GENMOD Procedure
Output 37.10.12 Diagnostic Plots for X1
Example 37.10: Bayesian Analysis of a Poisson Regression Model F 2615
Output 37.10.13 Diagnostic Plots for X2
2616 F Chapter 37: The GENMOD Procedure
Output 37.10.14 Diagnostic Plots for X3
Example 37.10: Bayesian Analysis of a Poisson Regression Model F 2617
Output 37.10.15 Diagnostic Plots for X4
2618 F Chapter 37: The GENMOD Procedure
Output 37.10.16 Diagnostic Plots for X5
Example 37.10: Bayesian Analysis of a Poisson Regression Model F 2619
Output 37.10.17 Diagnostic Plots for X6
In order to illustrate the use of an informative prior distribution, suppose that researchers expect that
a unit increase in body mass index (X1) will be associated with an increase in the mean number of
nodes of between 10% and 20%, and they want to incorporate this prior knowledge in the Bayesian
analysis. For log-linear models, the mean and linear predictor are related by log.i / D xi0 ˇ. If
X11 and X12 are two values of body mass index, 1 and 2 are the two mean values, and all other
covariates remain equal for the two values of X1, then
1
D exp.ˇ.X11
2
X12 //
so that for a unit change in X1,
1
D exp.ˇ/
2
1
If 1:1 1:2, then 1:1 exp.ˇ/ 1:2, or 0:095 ˇ 0:182. This gives you guidance in
2
specifying a prior distribution for the ˇ for body mass index. Taking the mean of the prior normal
2620 F Chapter 37: The GENMOD Procedure
distribution to be the midrange of the values of ˇ, and taking ˙ 2 to be the extremes of the range,
an N.0:1385; 0:0005/ is the resulting prior distribution. The second analysis uses this informative
normal prior distribution for the coefficient of X1 and uses independent noninformative normal
priors with zero means and variances equal to 106 for the remaining model regression parameters.
In the following BAYES statement, COEFFPRIOR=NORMAL(INPUT=NormalPrior) specifies the
normal prior distribution for the regression coefficients with means and variances contained in the
data set NormalPrior.
An analysis is performed using PROC GENMOD to obtain Bayesian estimates of the regression
coefficients by using the following SAS statements:
data NormalPrior;
input _type_ $ Intercept X1-X6;
datalines;
Var 1e6
0.0005
1e6
1e6
Mean 0.0
0.1385
0.0
0.0
;
run;
1e6
0.0
1e6
0.0
1e6
0.0
proc genmod data=Liver;
model Y = X1-X6 / dist=Poisson link=log;
bayes seed=1 plots=none coeffprior=normal(input=NormalPrior) ;
run;
ods graphics off;
The prior distributions for the regression parameters are shown in Output 37.10.18.
Output 37.10.18 Regression Coefficient Priors
The GENMOD Procedure
Bayesian Analysis
Independent Normal Prior for Regression Coefficients
Parameter
Mean
Precision
Intercept
X1
X2
X3
X4
X5
X6
0
0.1385
0
0
0
0
0
1E-6
2000
1E-6
1E-6
1E-6
1E-6
1E-6
Example 37.10: Bayesian Analysis of a Poisson Regression Model F 2621
Initial values for the MCMC are shown in Output 37.10.19. The initial values of the covariates are
joint estimates of their posterior modes. The prior distribution for X1 is informative, so the initial
value of X1 is further from the MLE than the rest of the covariates. Initial values for the rest of the
covariates are close to their MLEs, since noninformative prior distributions were specified for them.
Output 37.10.19 MCMC Initial Values and Seeds
Initial Values of the Chain
Chain
Seed
Intercept
X1
X2
X3
X4
1
1
2.14282
0.010595
-0.01434
-0.00301
-0.28062
Initial Values of the Chain
X5
X6
0.334983
0.231213
Goodness-of-fit, summary, and interval statistics are shown in Output 37.10.20. Except for X1, the
statistics shown in Output 37.10.20 are very similar to the previous statistics for noninformative
priors shown in Output 37.10.4 through Output 37.10.7. The point estimate for X1 is now positive.
This is expected because the prior distribution on ˇ1 is quite informative. The distribution reflects the
belief that the coefficient is positive. The N.0:1385; 0:0005/ distribution places the majority of its
probability density on positive values. As a result, the posterior density of ˇ1 places more likelihood
on positive values than in the noninformative case.
Output 37.10.20 Fit Statistics
Fit Statistics
DIC (smaller is better)
pD (effective number of parameters)
833.134
6.861
The GENMOD Procedure
Bayesian Analysis
Posterior Summaries
Parameter
N
Mean
Standard
Deviation
25%
Intercept
X1
X2
X3
X4
X5
X6
10000
10000
10000
10000
10000
10000
10000
2.1393
0.0104
-0.0143
-0.00318
-0.2801
0.3336
0.2333
0.2160
0.00685
0.00236
0.00218
0.0798
0.0834
0.0822
1.9929
0.00583
-0.0159
-0.00463
-0.3342
0.2772
0.1791
Percentiles
50%
2.1417
0.0106
-0.0143
-0.00313
-0.2807
0.3337
0.2327
75%
2.2845
0.0151
-0.0127
-0.00170
-0.2258
0.3902
0.2892
2622 F Chapter 37: The GENMOD Procedure
Output 37.10.20 continued
Posterior Intervals
Parameter
Alpha
Intercept
X1
X2
X3
X4
X5
X6
0.050
0.050
0.050
0.050
0.050
0.050
0.050
Equal-Tail Interval
1.7161
-0.00323
-0.0189
-0.00754
-0.4348
0.1705
0.0696
2.5599
0.0236
-0.00960
0.00101
-0.1223
0.4970
0.3968
HPD Interval
1.7075
-0.00264
-0.0189
-0.00754
-0.4311
0.1661
0.0655
2.5507
0.0241
-0.00972
0.000963
-0.1196
0.4915
0.3904
Example 37.11: Exact Poisson Regression
The following data, taken from Cox and Snell (1989, pp. 10–11), consists of the number, Notready,
of ingots that are not ready for rolling, out of Total tested, for several combinations of heating time
and soaking time:
data ingots;
input Heat Soak Notready
lnTotal= log(Total);
datalines;
7 1.0 0 10 14 1.0 0 31 27
7 1.7 0 17 14 1.7 0 43 27
7 2.2 0 7 14 2.2 2 33 27
7 2.8 0 12 14 2.8 0 31 27
7 4.0 0 9 14 4.0 0 19 27
;
Total @@;
1.0
1.7
2.2
2.8
4.0
1
4
0
1
1
56
44
21
22
16
51
51
51
51
1.0
1.7
2.2
4.0
3 13
0 1
0 1
0 1
The following invocation of PROC GENMOD fits an asymptotic (unconditional) Poisson regression
model to the data. The variable Notready is specified as the response variable, and the continuous
predictors Heat and Soak are defined in the CLASS statement as categorical predictors that use
reference coding. Specifying the offset variable as lnTotal enables you to model the ratio Notready/Total.
proc genmod data=ingots;
class Heat Soak / param=ref;
model Notready=Heat Soak / offset=lnTotal dist=poisson link=log;
exact Heat Soak / joint estimate;
exactoptions statustime=10;
run;
Example 37.11: Exact Poisson Regression F 2623
The EXACT statement is specified to additionally fit an exact conditional Poisson regression model.
Specifying the lnTotal offset variable models the ratio Notready/Total; in this case, the Total variable
contains the largest possible response value for each observation. The JOINT option produces a joint
test for the significance of the covariates, along with the usual marginal tests. The ESTIMATE option
produces exact parameter estimates for the covariates. The STATUSTIME=10 option is specified
in the EXACTOPTIONS statement for monitoring the progress of the results; this example can
take several minutes to complete due to the JOINT option. If you run out of memory, see the SAS
Companion for your system for information about how to increase the available memory.
The “Criteria For Assessing Goodness Of Fit” table is displayed in Output 37.11.1. Comparing the
deviance of 10.9363 with its asymptotic chi-square with 11 degrees of freedom distribution, you find
that the p-value is 0.084. This indicates that the specified model fits the data reasonably well.
Output 37.11.1 Unconditional Goodness of Fit Criteria
The GENMOD Procedure
Criteria For Assessing Goodness Of Fit
Criterion
DF
Value
Value/DF
Deviance
Scaled Deviance
Pearson Chi-Square
Scaled Pearson X2
Log Likelihood
Full Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)
11
11
11
11
10.9363
10.9363
9.3722
9.3722
-7.2408
-12.9038
41.8076
56.2076
49.3631
0.9942
0.9942
0.8520
0.8520
2624 F Chapter 37: The GENMOD Procedure
From the “Analysis Of Parameter Estimates” table in Output 37.11.2, you can see that only two of the
Heat parameters are deemed significant. Looking at the standard errors, you can see that the unconditional analysis had convergence difficulties with the Heat=7 parameter (Standard Error=264324.6),
which means you cannot fit this unconditional Poisson regression model to this data.
Output 37.11.2 Unconditional Maximum Likelihood Parameter Estimates
Analysis Of Maximum Likelihood Parameter Estimates
Parameter
Intercept
Heat
Heat
Heat
Soak
Soak
Soak
Soak
Scale
7
14
27
1
1.7
2.2
2.8
DF
Estimate
Standard
Error
1
1
1
1
1
1
1
1
0
-1.5700
-27.6129
-3.0107
-1.7180
-0.2454
0.5572
0.4079
-0.1301
1.0000
1.1657
264324.6
1.0025
0.7691
1.1455
1.1217
1.2260
1.4234
0.0000
Wald 95% Confidence
Limits
-3.8548
-518094
-4.9756
-3.2253
-2.4906
-1.6412
-1.9951
-2.9199
1.0000
Analysis Of Maximum
Likelihood Parameter
Estimates
Parameter
Intercept
Heat
Heat
Heat
Soak
Soak
Soak
Soak
Scale
Pr > ChiSq
7
14
27
1
1.7
2.2
2.8
NOTE: The scale parameter was held fixed.
0.1780
0.9999
0.0027
0.0255
0.8304
0.6193
0.7394
0.9272
0.7147
518039.0
-1.0458
-0.2106
1.9998
2.7557
2.8109
2.6597
1.0000
Wald
Chi-Square
1.81
0.00
9.02
4.99
0.05
0.25
0.11
0.01
Example 37.11: Exact Poisson Regression F 2625
Following the output from the asymptotic analysis, the exact conditional Poisson regression results
are displayed, as shown in Output 37.11.3.
Output 37.11.3 Exact Tests
The GENMOD Procedure
Exact Conditional Analysis
Conditional Exact Tests
Effect
Test
Joint
Score
Probability
Score
Probability
Score
Probability
Heat
Soak
Statistic
18.3665
1.294E-6
15.8259
0.000175
1.4612
0.00735
--- p-Value --Exact
Mid
0.0137
0.0471
0.0023
0.0063
0.8683
0.8176
0.0137
0.0471
0.0022
0.0062
0.8646
0.8139
The Joint test in the “Conditional Exact Tests” table in Output 37.11.3 is produced by specifying
the JOINT option in the EXACT statement. The p-values for this test indicate that the parameters
for Heat and Soak are jointly significant as explanatory effects in the model. If the Heat variable is
the only explanatory variable in your model, then the rows of this table labeled as “Heat” show the
joint significance of all the Heat effect parameters in that reduced model. In this case, a model that
contains only the Heat parameters still explains a significant amount of the variability; however, you
can see that a model that contains only the Soak parameters would not be significant.
The “Exact Parameter Estimates” table in Output 37.11.4 displays parameter estimates and tests of
significance for the levels of the CLASS variables. Again, the Heat=7 parameter has some difficulties;
however, in the exact analysis, a median unbiased estimate is computed for the parameter instead of
a maximum likelihood estimate. The confidence limits show that the Heat variable contains some
explanatory power, while the categorical Soak variable is insignificant and can be dropped from the
model.
Output 37.11.4 Exact Parameter Estimates
Exact Parameter Estimates
Parameter
Heat
Heat
Heat
Soak
Soak
Soak
Soak
Estimate
7
14
27
1
1.7
2.2
2.8
-2.7552*
-3.0255
-1.7846
-0.3231
0.5375
0.4035
-0.1661
Standard
Error
.
1.0128
0.8065
1.1717
1.1284
1.2347
1.4214
95% Confidence
Limits
-Infinity
-5.7450
-3.6779
-2.8673
-1.8056
-2.5785
-4.5490
-0.4267
-0.6194
0.2260
3.6754
4.4588
4.5054
4.2168
NOTE: * indicates a median unbiased estimate.
p-Value
0.0199
0.0113
0.0844
1.0000
1.0000
1.0000
1.0000
2626 F Chapter 37: The GENMOD Procedure
N OTE : If you want to make predictions from the exact results, you can obtain an estimate for the
intercept parameter by specifying the INTERCEPT keyword in the EXACT statement. You should
also remove the JOINT option to reduce the amount of time and memory consumed.
References
Agresti, A. (2002), Categorical Data Analysis, Second Edition, New York: John Wiley & Sons.
Aitkin, M., Anderson, D., Francis, B., and Hinde, J. (1989), Statistical Modelling in GLIM, Oxford:
Oxford Science Publications.
Akaike, H. (1979), “A Bayesian Extension of the Minimum AIC Procedure of Autoregressive Model
Fitting,” Biometrika, 66, 237–242.
Akaike, H. (1981), “Likelihood of a Model and Information Criteria,” Journal of Econometrics, 16,
3–14.
Boos, D. (1992), “On Generalized Score Tests,” The American Statistician, 46, 327–333.
Cameron, A. C. and Trivedi, P. K. (1998), Regression Analysis of Count Data, Cambridge: Cambridge
University Press.
Carey, V., Zeger, S. L., and Diggle, P. (1993), “Modelling Multivariate Binary Data with Alternating
Logistic Regressions,” Biometrika, 80, 517–526.
Collett, D. (2003), Modelling Binary Data, Second Edition, London: Chapman & Hall.
Cook, R. D. and Weisberg, S. (1982), Residuals and Influence in Regression, New York: Chapman &
Hall.
Cox, D. R. and Snell, E. J. (1989), The Analysis of Binary Data, Second Edition, London: Chapman
& Hall.
Davison, A. C. and Snell, E. J. (1991), “Residuals and Diagnostics,” in D. V. Hinkley, N. Reid, and
E. J. Snell, eds., Statistical Theory and Modelling, London: Chapman & Hall.
Diggle, P. J., Liang, K. Y., and Zeger, S. L. (1994), Analysis of Longitudinal Data, Oxford: Clarendon
Press.
Dobson, A. (1990), An Introduction to Generalized Linear Models, London: Chapman & Hall.
Firth, D. (1991), “Generalized Linear Models,” in D. V. Hinkley, N. Reid, and E. J. Snell, eds.,
Statistical Theory and Modelling, London: Chapman & Hall.
Fischl, M. A., Richman, D. D., and Hansen, N. (1990), “The Safety and Efficacy of Zidovudine
(AZT) in the Treatment of Subjects with Mildly Symptomatic Human Immunodeficiency Virus
Type I (HIV) Infection,” Annals of Internal Medicine, 112, 727–737.
Gamerman, D. (1997), “Efficient Sampling from the Posterior Distribution in Generalized Linear
Models,” Statistical Computing, 7, 57–68.
References F 2627
Gilks, W. (2003), “Adaptive Metropolis Rejection Sampling (ARMS),” software from MRC
Biostatistics Unit, Cambridge, UK, http://www.maths.leeds.ac.uk/~wally.gilks/
adaptive.rejection/web_page/Welcome.html.
Gilks, W. R., Best, N. G., and Tan, K. K. C. (1995), “Adaptive Rejection Metropolis Sampling with
Gibbs Sampling,” Applied Statistics, 44, 455–472.
Gilks, W. R., Richardson, S., and Spiegelhalter, D. J. (1996), Markov Chain Monte Carlo in Practice,
London: Chapman & Hall.
Gilks, W. R. and Wild, P. (1992), “Adaptive Rejection Sampling for Gibbs Sampling,” Applied
Statistics, 41, 337–348.
Hardin, J. W. and Hilbe, J. M. (2003), Generalized Estimating Equations, Boca Raton, FL: Chapman
& Hall/CRC.
Hilbe, J. (1994), “Log Negative Binomial Regression Using the GENMOD Procedure,” in Proceedings of the Nineteenth Annual SAS Users Group International Conference, Cary, NC: SAS Institute
Inc.
Hilbe, J. M. (2007), Negative Binomial Regression, New York: Cambridge University Press.
Hilbe, J. M. (2009), Logistic Regression Models, London: 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.
Ibrahim, J. G., Chen, M.-H., and Lipsitz, S. R. (1999), “Monte Carlo EM for Missing Covariates in
Parametric Regression Models,” Biometrics, 55, 591–596.
Ibrahim, J. G., Chen, M. H., and Sinha, D. (2001), Bayesian Survival Analysis, New York: SpringerVerlag.
Ibrahim, J. G. and Laud, P. W. (1991), “On Bayesian Analysis of Generalized Linear Models Using
Jeffreys’ Prior,” Journal of the American Statistical Association, 86, 981–986.
Lambert, D. (1992), “Zero-Inflated Poisson Regression Models with an Application to Defects in
Manufacturing,” Technometrics, 34, 1–14.
Lawless, J. E. (1987), “Negative Binomial and Mixed Poisson Regression,” The Canadian Journal of
Statistics, 15, 209–225.
Lawless, J. F. (2003), Statistical Model and Methods for Lifetime Data, Second Edition, New York:
John Wiley & Sons.
Liang, K. Y. and Zeger, S. L. (1986), “Longitudinal Data Analysis Using Generalized Linear Models,”
Biometrika, 73, 13–22.
Lin, D. Y., Wei, L. J., and Ying, Z. (2002), “Model-Checking Techniques Based on Cumulative
Residuals,” Biometrics, 58, 1–12.
Lipsitz, S. H., Fitzmaurice, G. M., Orav, E. J., and Laird, N. M. (1994), “Performance of Generalized
Estimating Equations in Practical Situations,” Biometrics, 50, 270–278.
2628 F Chapter 37: The GENMOD Procedure
Lipsitz, S. H., Kim, K., and Zhao, L. (1994), “Analysis of Repeated Categorical Data Using
Generalized Estimating Equations,” Statistics in Medicine, 13, 1149–1163.
Littell, R. C., Freund, R. J., and Spector, P. C. (1991), SAS System for Linear Models, Third Edition,
Cary, NC: SAS Institute Inc.
Long, J. S. (1997), Regression Models for Categorical and Limited Dependent Variables, Thousand
Oaks, CA: Sage Publications.
McCullagh, P. (1983), “Quasi-likelihood Functions,” Annals of Statistics, 11, 59–67.
McCullagh, P. and Nelder, J. A. (1989), Generalized Linear Models, Second Edition, London:
Chapman & Hall.
Mehta, C. R., Patel, N., and Senchaudhuri, P. (1992), “Exact Stratified Linear Rank Tests for Ordered
Categorical and Binary Data,” Journal of Computational and Graphical Statistics, 1, 21–40.
Miller, M. E., Davis, C. S., and Landis, J. R. (1993), “The Analysis of Longitudinal Polytomous Data:
Generalized Estimating Equations and Connections with Weighted Least Squares,” Biometrics, 49,
1033–1044.
Myers, R. H., Montgomery, D. C., and Vining, G. (2002), Generalized Linear Models with Applications in Engineering and the Sciences, New York: John Wiley & Sons.
Nelder, J. A. and Wedderburn, R. W. M. (1972), “Generalized Linear Models,” Journal of the Royal
Statistical Society, Series A, 135, 370–384.
Nelson, W. (1982), Applied Life Data Analysis, New York: John Wiley & Sons.
Neter, J., Kutner, M. H., Nachtsheim, C. J., and Wasserman, W. (1996), Applied Linear Statistical
Models, Fourth Edition, Chicago: Irwin.
Pan, W. (2001), “Akaike’s Information Criterion in Generalized Estimating Equations,” Biometrics,
57, 120–125.
Pregibon, D. (1981), “Logistic Regression Diagnostics,” Annals of Statistics, 9, 705–724.
Preisser, J. S. and Qaqish, B. F. (1996), “Deletion Diagnostics for Generalised Estimating Equations,”
Biometrika, 83, 551–562.
Rao, C. R. (1973), Linear Statistical Inference, New York: John Wiley & Sons.
Rotnitzky, A. and Jewell, N. P. (1990), “Hypothesis Testing of Regression Parameters in Semiparametric Generalized Linear Models for Cluster Correlated Data,” Biometrika, 77, 485–497.
Royall, R. M. (1986), “Model Robust Inference Using Maximum Likelihood Estimators,” International Statistical Review, 54, 221–226.
Searle, S. R. (1971), Linear Models, New York: John Wiley & Sons.
Simonoff, J. S. (2003), Analyzing Categorical Data, New York: Springer-Verlag.
Spiegelhalter, D. J., Best, N. G., Carlin, B. P., and Van der Linde, A. (2002), “Bayesian Measures of
Model Complexity and Fit,” Journal of the Royal Statistical Society, Series B, 64(4), 583–616,
with discussion.
References F 2629
Stokes, M. E., Davis, C. S., and Koch, G. G. (2000), Categorical Data Analysis Using the SAS
System, Second Edition, Cary, NC: SAS Institute Inc.
Thall, P. F. and Vail, S. C. (1990), “Some Covariance Models for Longitudinal Count Data with
Overdispersion,” Biometrics, 46, 657–671.
Ware, J. H., Dockery, S. A. I., Speizer, F. E., and Ferris, B. G., Jr. (1984), “Passive Smoking,
Gas Cooking, and Respiratory Health of Children Living in Six Cities,” American Review of
Respiratory Diseases, 129, 366–374.
White, H. (1982), “Maximum Likelihood Estimation of Misspecified Models,” Econometrica, 50,
1–25.
Williams, D. A. (1987), “Generalized Linear Model Diagnostics Using the Deviance and Single
Case Deletions,” Applied Statistics, 36, 181–191.
Zeger, S. L., Liang, K. Y., and Albert, P. S. (1988), “Models for Longitudinal Data: A Generalized
Estimating Equation Approach,” Biometrics, 44, 1049–1060.
Subject Index
adjusted residuals
GENMOD procedure, 2529
aggregates of residuals, 2597, 2604
Akaike’s information criterion
(GENMOD), 2519
aliasing
GENMOD procedure, 2438
ALR algorithm
GENMOD procedure, 2589
alternating logistic regressions (ALR)
GENMOD procedure, 2589
bar (|) operator
GENMOD procedure, 2522
Bayesian analysis linear regression
GENMOD procedure, 2440
Bayesian information criterion
(GENMOD), 2519
binomial distribution
GENMOD procedure, 2513
case deletion diagnostics
GENMOD procedure, 2544
CATMOD procedure
log-linear models, 2435
classification variables
GENMOD procedure, 2522
sort order of levels (GENMOD), 2457
confidence intervals
confidence coefficient (GENMOD), 2492
fitted values of the mean (GENMOD), 2496,
2528
profile likelihood (GENMOD), 2495, 2525
Wald (GENMOD), 2498, 2526
continuous variables
GENMOD procedure, 2522
contrasts
GENMOD procedure, 2481
convergence criterion
GENMOD procedure, 2492, 2503
correlated data
GEE (GENMOD), 2429, 2532
correlation
matrix (GENMOD), 2492, 2517
covariance matrix
GENMOD procedure, 2492, 2517
crossed effects
GENMOD procedure, 2522
cumulative residuals, 2597, 2604
design matrix
GENMOD procedure, 2523
deviance
definition (GENMOD), 2433
GENMOD procedure, 2491
scaled (GENMOD), 2517, 2518
deviance information criterion, 2551
deviance residuals
GENMOD procedure, 2528, 2529
diagnostics
GENMOD procedure, 2493, 2544
DIC, 2551
dispersion parameter
estimation (GENMOD), 2432, 2519, 2524,
2525
GENMOD procedure, 2520
weights (GENMOD), 2509
effect
specification (GENMOD), 2522
effective number of parameters, 2551
estimability checking
GENMOD procedure, 2479
estimation
dispersion parameter (GENMOD), 2432
maximum likelihood (GENMOD), 2516
regression parameters (GENMOD), 2432
events/trials format for response
GENMOD procedure, 2491, 2513
exact conditional logistic regression, see exact
logistic regression
exact conditional Poisson regression, see exact
Poisson regression
exact logistic regression
GENMOD procedure, 2482, 2553
exact Poisson regression
GENMOD procedure, 2482, 2507, 2553,
2622
exponential distribution
GENMOD procedure, 2582
F statistics
GENMOD procedure, 2526, 2527
Fisher’s scoring method
GENMOD procedure, 2498, 2517
gamma distribution
GENMOD procedure, 2512
GEE, see generalized estimating equations
Generalized Estimating Equations (GEE), 2453
generalized estimating equations (GEE), 2503,
2532, 2586, 2592
generalized linear model
GENMOD procedure, 2431
theory (GENMOD), 2510
GENMOD procedure
adjusted residuals, 2529
AIC, 2519
Akaike’s information criterion, 2519
aliasing, 2438
Bayesian analysis linear regression, 2440
Bayesian information criterion, 2519
BIC, 2519
binomial distribution, 2513
built-in link function, 2432
built-in probability distribution, 2433
case deletion diagnostics, 2544
classification variables, 2522
confidence intervals, 2492
continuous variables, 2522
contrasts, 2481
convergence criterion, 2484, 2485, 2492,
2503
correlated data, 2429, 2532
correlation matrix, 2492, 2517
covariance matrix, 2492, 2517
crossed effects, 2522
design matrix, 2523
deviance, 2491
deviance definition, 2433
deviance residuals, 2529
diagnostics, 2493, 2544
dispersion parameter, 2520
dispersion parameter estimation, 2432, 2524,
2525
dispersion parameter weights, 2509
effect specification, 2522
estimability checking, 2479
events/trials format for response, 2491, 2513
exact logistic regression, 2482, 2553
exact Poisson regression, 2482, 2507, 2553,
2622
expected information matrix, 2517
exponential distribution, 2582
F statistics, 2526, 2527
Fisher’s scoring method, 2498, 2517
gamma distribution, 2512
GEE, 2429, 2453, 2503, 2532, 2586, 2589,
2592
generalized estimating equations (GEE),
2429
generalized linear model, 2431
geometric distribution, 2512
goodness of fit, 2517
gradient, 2516
Hessian matrix, 2516
information matrix, 2498
initial values, 2493, 2504
intercept, 2433, 2436, 2495
inverse Gaussian distribution, 2512
Lagrange multiplier statistics, 2527
life data, 2579
likelihood residuals, 2529
linear model, 2430
linear predictor, 2429, 2430, 2436, 2523,
2560
link function, 2429, 2431, 2514
log-likelihood functions, 2515
log-linear models, 2435
logistic regression, 2574
main effects, 2522
maximum likelihood estimation, 2516
_MEAN_ automatic variable, 2502
model checking, 2597, 2604
multinomial distribution, 2513
multinomial models, 2529
negative binomial distribution, 2512
nested effects, 2522
Newton-Raphson algorithm, 2516
normal distribution, 2511
observed information matrix, 2517
offset, 2497, 2560
offset variable, 2436
ordering of effects, 2457
ordinal data, 2582
output data sets, 2553, 2554
output ODS Graphics table names, 2572
output table names, 2568
overdispersion, 2521
Pearson residuals, 2529
Pearson’s chi-square, 2491, 2517, 2519
Poisson distribution, 2512
Poisson regression, 2435
polynomial effects, 2522
profile likelihood confidence intervals, 2495,
2525
programming statements, 2502
QIC, 2539
quasi-likelihood, 2521
quasi-likelihood functions, 2539
quasi-likelihood information criterion, 2539
raw residuals, 2528
regression parameters estimation, 2432
regressor effects, 2522
repeated measures, 2429, 2532
residuals, 2497, 2528, 2529
_RESP_ automatic variable, 2502
scale parameter, 2514
scaled deviance, 2517, 2518
score statistics, 2527
singular contrast matrix, 2479
stratified exact logistic regression, 2507
stratified exact Poisson regression, 2507
subpopulation, 2491
suppressing output, 2462
Type 1 analysis, 2434, 2523
Type 3 analysis, 2434, 2524
user-defined link function, 2487
variance function, 2433
Wald confidence intervals, 2498, 2526
working correlation matrix, 2504, 2506, 2532
_XBETA_ automatic variable, 2502
zero-inflated models, 2530
zero-inflated negative binomial distribution,
2513
zero-inflated Poisson distribution, 2513
geometric distribution
GENMOD procedure, 2512
goodness of fit
GENMOD procedure, 2517
gradient
GENMOD procedure, 2516
functions (GENMOD), 2515
log-linear models
CATMOD procedure, 2435
GENMOD procedure, 2435
logistic regression
GENMOD procedure, 2432, 2574
Hessian matrix
GENMOD procedure, 2516
offset
GENMOD procedure, 2497, 2560
offset variable
GENMOD procedure, 2436
ordinal model
GENMOD procedure, 2582
output data sets
GENMOD procedure, 2553, 2554
output ODS Graphics table names
GENMOD procedure, 2572
output table names
GENMOD procedure, 2568
overdispersion
GENMOD procedure, 2521
information matrix
expected (GENMOD), 2517
observed (GENMOD), 2517
initial values
GENMOD procedure, 2493, 2504
intercept
GENMOD procedure, 2433, 2436, 2495
inverse Gaussian distribution
GENMOD procedure, 2512
Lagrange multiplier
statistics (GENMOD), 2527
life data
GENMOD procedure, 2579
likelihood residuals
GENMOD procedure, 2529
linear model
GENMOD procedure, 2430, 2431
linear predictor
GENMOD procedure, 2429, 2430, 2436,
2523, 2560
link function
built-in (GENMOD), 2432, 2494
GENMOD procedure, 2429, 2431, 2514
user-defined (GENMOD), 2487
log-likelihood
main effects
GENMOD procedure, 2522
maximum likelihood
estimation (GENMOD), 2516
model assessment, 2597, 2604
model checking, 2597, 2604
multinomial
distribution (GENMOD), 2513
models (GENMOD), 2529
negative binomial distribution
GENMOD procedure, 2512
nested effects
GENMOD procedure, 2522
Newton-Raphson algorithm
GENMOD procedure, 2516
normal distribution
GENMOD procedure, 2511
parameter estimates
GENMOD procedure, 2565
Pearson residuals
GENMOD procedure, 2528, 2529
Pearson’s chi-square
GENMOD procedure, 2491, 2517, 2519
Poisson distribution
GENMOD procedure, 2512
Poisson regression
GENMOD procedure, 2432, 2435
polynomial effects
GENMOD procedure, 2522
probability distribution
built-in (GENMOD), 2433, 2493
exponential family (GENMOD), 2510
user-defined (GENMOD), 2479
profile likelihood confidence intervals
GENMOD procedure, 2525
programming statements
GENMOD procedure, 2502
quasi-likelihood
functions (GENMOD), 2539
GENMOD procedure, 2521
quasi-likelihood information criterion
(GENMOD), 2539
raw residuals
GENMOD procedure, 2528
regressor effects
GENMOD procedure, 2522
repeated measures
GEE (GENMOD), 2429, 2532
residuals
GENMOD procedure, 2497, 2528, 2529
response variable
sort order of levels (GENMOD), 2461
scale parameter
GENMOD procedure, 2514
score statistics
GENMOD procedure, 2527
singularity criterion
contrast matrix (GENMOD), 2479
information matrix (GENMOD), 2498
standard error
GENMOD procedure, 2565
stratified exact logistic regression
GENMOD procedure, 2507
stratified exact Poisson regression
GENMOD procedure, 2507
subpopulation
GENMOD procedure, 2491
suppressing output
GENMOD procedure, 2462
Type 1 analysis
GENMOD procedure, 2434, 2523
Type 3 analysis
GENMOD procedure, 2434, 2524
variance function
GENMOD procedure, 2433
working correlation matrix
GENMOD procedure, 2504, 2506, 2532
zero-inflated
models (GENMOD), 2530
zero-inflated negative binomial
distribution (GENMOD), 2513
zero-inflated Poisson
distribution (GENMOD), 2513
Syntax Index
ABSFCONV option
MODEL statement (GENMOD), 2484
AGGREGATE= option
MODEL statement (GENMOD), 2491
ALPHA= option
ESTIMATE statement (GENMOD), 2481
EXACT statement (GENMOD), 2482
MODEL statement (GENMOD), 2492
ALPHAINIT= option
REPEATED statement (GENMOD), 2503
ASSESS statement
GENMOD procedure, 2462
DATA= option
PROC GENMOD statement, 2457
DESCENDING option
CLASS statement (GENMOD), 2474
DEVIANCE statement, GENMOD procedure,
2479, 2503
DIAGNOSTICS option
MODEL statement (GENMOD), 2493
DIST= option
MODEL statement (GENMOD), 2493
DSCALE
MODEL statement (GENMOD), 2497
BAYES statement
GENMOD procedure, 2463
BY statement
GENMOD procedure, 2473
E option
CONTRAST statement (GENMOD), 2479
ESTIMATE statement (GENMOD), 2481
ECORRB option
REPEATED statement (GENMOD), 2504
ECOVB option
REPEATED statement (GENMOD), 2504
EFFECTPLOT statement
GENMOD procedure, 2480
ERR= option
MODEL statement (GENMOD), 2493
ESTIMATE option
EXACT statement (GENMOD), 2482
ESTIMATE statement
GENMOD procedure, 2481
EXACT statement
GENMOD procedure, 2482
EXACTONLY option
PROC GENMODstatement, 2457
EXACTOPTIONS statement
GENMOD procedure, 2484
EXP option
ESTIMATE statement (GENMOD), 2481
EXPECTED option
MODEL statement (GENMOD), 2493
CHECKDEPENDENCY= option
STRATA statement (GENMOD), 2508
CICONV= option
MODEL statement (GENMOD), 2492
CL option
MODEL statement (GENMOD), 2492
CLASS statement
GENMOD procedure, 2474
CLTYPE= option
EXACT statement (GENMOD), 2482
CODING= option
MODEL statement (GENMOD), 2492
CONTRAST statement
GENMOD procedure, 2477
CONVERGE= option
MODEL statement (GENMOD), 2492
REPEATED statement (GENMOD), 2503
CONVH= option
MODEL statement (GENMOD), 2492
CORR= option
REPEATED statement (GENMOD), 2506
CORRB option
MODEL statement (GENMOD), 2492
REPEATED statement (GENMOD), 2504
CORRW option
REPEATED statement (GENMOD), 2504
COVB option
MODEL statement (GENMOD), 2492
REPEATED statement (GENMOD), 2504
CPREFIX= option
CLASS statement (GENMOD), 2474
FCONV= option
MODEL statement (GENMOD), 2485
FREQ statement
GENMOD procedure, 2487
FWDLINK statement, GENMOD procedure,
2487, 2503
GENMOD procedure
syntax, 2456
GENMOD procedure, ASSESS statement, 2462
GENMOD PROCEDURE, BAYES statement,
2463
GENMOD procedure, BAYES statement
STATISTICS= option, 2472
THINNING= option, 2473
GENMOD procedure, BY statement, 2473
GENMOD procedure, CLASS statement, 2474
CPREFIX= option, 2474
DESCENDING option, 2474
LPREFIX= option, 2474
MISSING option, 2474
ORDER= option, 2474
PARAM= option, 2475
REF= option, 2476
TRUNCATE option, 2476
GENMOD procedure, CONTRAST statement,
2477
E option, 2479
SINGULAR= option, 2479
WALD option, 2479
GENMOD procedure, DEVIANCE statement,
2479, 2503
GENMOD procedure, EFFECTPLOT statement,
2480
GENMOD procedure, ESTIMATE statement
ALPHA= option, 2481
E option, 2481
EXP option, 2481
SINGULAR= option, 2481
GENMOD procedure, EXACT statement, 2482
ALPHA= option, 2482
CLTYPE= option, 2482
ESTIMATE option, 2482
JOINT option, 2483
JOINTONLY option, 2483
MIDPFACTOR= option, 2483
ONESIDED option, 2483
OUTDIST= option, 2483
GENMOD procedure, EXACTOPTIONS
statement, 2484
GENMOD procedure, FREQ statement, 2481,
2487
GENMOD procedure, FWDLINK statement,
2487, 2503
GENMOD procedure, INVLINK statement, 2488,
2503
GENMOD procedure, LSMESTIMATE statement,
2489
GENMOD procedure, MODEL statement, 2491
ABSFCONV option, 2484
AGGREGATE= option, 2491
ALPHA= option, 2492
CICONV= option, 2492
CL option, 2492
CODING= option, 2492
CONVERGE= option, 2492
CONVH= option, 2492
CORRB option, 2492
COVB option, 2492
DIAGNOSTICS option, 2493
DIST= option, 2493
ERR= option, 2493
EXPECTED option, 2493
FCONV= option, 2485
INFLUENCE option, 2493
INITIAL= option, 2493
INTERCEPT= option, 2494
ITPRINT option, 2494
LINK= option, 2494
LRCI option, 2495
MAXIT= option, 2495
NOINT option, 2495
NOLOGSCALE option, 2486
NOSCALE option, 2495
OBSTATS option, 2495
OFFSET= option, 2497
PRED option, 2497
PREDICTED option, 2497
RESIDUALS option, 2497
SCALE= option, 2497
SCORING= option, 2498
SINGULAR= option, 2498
TYPE1 option, 2498
TYPE3 option, 2498
WALD option, 2498
WALDCI option, 2498
XCONV= option, 2487
XVARS option, 2499
GENMOD procedure, OUTPUT statement, 2499
keyword= option, 2499
OUT= option, 2499
GENMOD procedure, PROC GENMOD
statement, 2457
DATA= option, 2457
NAMELEN= option, 2457
ORDER= option, 2457
PLOTS= option, 2458
RORDER= option, 2461
GENMOD procedure, REPEATED statement,
2453, 2503
ALPHAINIT= option, 2503
CONVERGE= option, 2503
CORR= option, 2506
CORRB option, 2504
CORRW option, 2504
COVB option, 2504
ECORRB option, 2504
ECOVB option, 2504
INITIAL= option, 2504
INTERCEPT= option, 2504
LOGOR= option, 2504
MAXITER= option, 2505
MCORRB option, 2505
MCOVB option, 2505
MODELSE option, 2505
RUPDATE= option, 2505
SORTED option, 2505
SUBCLUSTER= option, 2505
SUBJECT= option, 2503
TYPE= option, 2506
V6CORR option, 2506
WITHIN= option, 2506
WITHINSUBJECT= option, 2506
YPAIR= option, 2506
ZDATA= option, 2507
ZROW= option, 2507
GENMOD procedure, SCWGT statement, 2509
GENMOD procedure, SLICE statement, 2507
GENMOD procedure, STORE statement, 2507
GENMOD procedure, STRATA statement, 2507
CHECKDEPENDENCY= option, 2508
INFO option, 2509
MISSING option, 2508
NOSUMMARY option, 2509
GENMOD procedure, VARIANCE statement,
2509
GENMOD procedure, WEIGHT statement, 2509
GENMOD procedure, ZEROMODEL statement,
2510
LINK= option, 2510
GENMODprocedure, PROC GENMODstatement
EXACTONLY option, 2457
INFLUENCE option
MODEL statement (GENMOD), 2493
INFO option
STRATA statement (GENMOD), 2509
INITIAL= option
MODEL statement (GENMOD), 2493
REPEATED statement (GENMOD), 2504
INTERCEPT= option
MODEL statement (GENMOD), 2494
REPEATED statement (GENMOD), 2504
INVLINK statement, GENMOD procedure, 2488,
2503
ITPRINT option
MODEL statement (GENMOD), 2494
JOINT option
EXACT statement (GENMOD), 2483
JOINTONLY option
EXACT statement (GENMOD), 2483
keyword= option
OUTPUT statement (GENMOD), 2499
LINK= option
MODEL statement (GENMOD), 2494
ZEROMODEL statement (GENMOD), 2510
LOGOR= option
REPEATED statement (GENMOD), 2504
LPREFIX= option
CLASS statement (GENMOD), 2474
LRCI option
MODEL statement (GENMOD), 2495
LSMESTIMATE statement
GENMOD procedure, 2489
MAXIT= option
MODEL statement (GENMOD), 2495
MAXITER= option
REPEATED statement (GENMOD), 2505
MCORRB option
REPEATED statement (GENMOD), 2505
MCOVB option
REPEATED statement (GENMOD), 2505
MIDPFACTOR= option
EXACT statement (GENMOD), 2483
MISSING option
CLASS statement (GENMOD), 2474
STRATA statement (GENMOD), 2508
MODEL statement
GENMOD procedure, 2491
MODELSE option
REPEATED statement (GENMOD), 2505
NAMELEN= option
PROC GENMOD statement, 2457
NOINT option
MODEL statement (GENMOD), 2495
NOLOGSCALE option
MODEL statement (GENMOD), 2486
NOSCALE option
MODEL statement (GENMOD), 2495
NOSUMMARY option
STRATA statement (GENMOD), 2509
OBSTATS option
MODEL statement (GENMOD), 2495
OFFSET= option
MODEL statement (GENMOD), 2497
ONESIDED option
EXACT statement (GENMOD), 2483
ORDER= option
CLASS statement (GENMOD), 2474
PROC GENMOD statement, 2457
OUT= option
OUTPUT statement (GENMOD), 2499
OUTDIST= option
EXACT statement (GENMOD), 2483
OUTPUT statement
GENMOD procedure, 2499
PARAM= option
CLASS statement (GENMOD), 2475
PLOTS= option
PROC GENMOD statement, 2458
PRED option
MODEL statement (GENMOD), 2497
PREDICTED option
MODEL statement (GENMOD), 2497
PROC GENMOD statement, see GENMOD
procedure
PSCALE
MODEL statement (GENMOD), 2497
REF= option
CLASS statement (GENMOD), 2476
REPEATED statement
GENMOD procedure, 2453, 2503
RESIDUALS option
MODEL statement (GENMOD), 2497
RORDER= option
PROC GENMOD statement, 2461
RUPDATE= option
REPEATED statement (GENMOD), 2505
SCALE= option
MODEL statement (GENMOD), 2497
SCORING= option
MODEL statement (GENMOD), 2498
SCWGT statement
GENMOD procedure, 2509
SINGULAR= option
CONTRAST statement (GENMOD), 2479
ESTIMATE statement (GENMOD), 2481
MODEL statement (GENMOD), 2498
SLICE statement
GENMOD procedure, 2507
SORTED option
REPEATED statement (GENMOD), 2505
STATISTICS= option
BAYES statement(GENMOD), 2472
STORE statement
GENMOD procedure, 2507
STRATA statement
GENMOD procedure, 2507
SUBCLUSTER= option
REPEATED statement (GENMOD), 2505
SUBJECT= option
REPEATED statement (GENMOD), 2503
THINNING= option
BAYES statement(GENMOD), 2473
TRUNCATE option
CLASS statement (GENMOD), 2476
TYPE1 option
MODEL statement (GENMOD), 2498
TYPE3 option
MODEL statement (GENMOD), 2498
TYPE= option
REPEATED statement (GENMOD), 2506
V6CORR option
REPEATED statement (GENMOD), 2506
VARIANCE statement, GENMOD procedure,
2509
WALD option
CONTRAST statement (GENMOD), 2479
MODEL statement (GENMOD), 2498
WALDCI option
MODEL statement (GENMOD), 2498
WEIGHT statement
GENMOD procedure, 2509
WITHIN= option
REPEATED statement (GENMOD), 2506
WITHINSUBJECT= option
REPEATED statement (GENMOD), 2506
XCONV= option
MODEL statement (GENMOD), 2487
XVARS option
MODEL statement (GENMOD), 2499
YPAIR= option
REPEATED statement (GENMOD), 2506
ZDATA= option
REPEATED statement (GENMOD), 2507
ZEROMODEL statement
GENMOD procedure, 2510
ZROW= option
REPEATED statement (GENMOD), 2507
Your Turn
We welcome your feedback.
If you have comments about this book, please send them to
[email protected] Include the full title and page numbers (if
applicable).
If you have comments about the software, please send them to
[email protected]
SAS Publishing Delivers!
®
Whether you are new to the work force or an experienced professional, you need to distinguish yourself in this rapidly
changing and competitive job market. SAS Publishing provides you with a wide range of resources to help you set
yourself apart. Visit us online at support.sas.com/bookstore.
®
SAS Press
®
Need to learn the basics? Struggling with a programming problem? You’ll find the expert answers that you
need in example-rich books from SAS Press. Written by experienced SAS professionals from around the
world, SAS Press books deliver real-world insights on a broad range of topics for all skill levels.
SAS Documentation
support.sas.com/saspress
®
To successfully implement applications using SAS software, companies in every industry and on every
continent all turn to the one source for accurate, timely, and reliable information: SAS documentation.
We currently produce the following types of reference documentation to improve your work experience:
• Online help that is built into the software.
• Tutorials that are integrated into the product.
• Reference documentation delivered in HTML and PDF – free on the Web.
• Hard-copy books.
support.sas.com/publishing
SAS Publishing News
®
Subscribe to SAS Publishing News to receive up-to-date information about all new SAS titles, author
podcasts, and new Web site features via e-mail. Complete instructions on how to subscribe, as well as
access to past issues, are available at our Web site.
support.sas.com/spn
SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc. in the USA and other countries. ® indicates USA registration.
Other brand and product names are trademarks of their respective companies. © 2009 SAS Institute Inc. All rights reserved. 518177_1US.0109
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement