4. PROC GENMOD Commands

Many SAS users who apply statistical methods to data are familiar with 
procedures such as PROCs REG and GLM for regression and analysis of 
variance/covariance models and perhaps the relatively new PROC MIXED for 
fitting linear models for continuous data. PROC GENMOD can also compute 
fixed effect linear models with normally distributed residuals; however, 
its primary function is to work with data based on other types of 
distributions such as the binomial for dichotomous data and Poisson or 
negative binomial models for counts. In addition, a special feature of 
PROC GENMOD is that it can accommodate the analysis of correlated discrete 
data in a manner similar to how PROC MIXED works with correlated 
continuous data under a an estimation technique known as Generalized 
Estimating Equations (GEE).

PROC GENMOD can analyze simple as well as very complex statistical models. 
Other SAS procedures also exist which perform the same analysis; however, 
some features may exist in these procedures which are missing in PROC 
GENMOD but still is important to know the differences and advantages of 
each. Examples are given showing how PROC GENMOD will analyze various 
types of response data which do not meet the assumptions of normality of 
the residuals.

The following statements for the GENMOD procedure resemble the analogous 
statements in PROC GLM or PROC MIXED and have similar or the same 
functions. Items enclosed within the <> are optional.
  
PROC GENMOD < options > ;
BY variables ;     
CLASS variables / <options>; 
FREQ variable ;
MODEL response = < effects > < /options > ;  
CONTRAST 'label' effect < ...effect values > < /options > ;  
ESTIMATE 'label' effect < ...effect values > < /options > ;  
LSMEANS effects < / options > ;  
OUTPUT OUT=dataset  < keyword=name...keyword=name > ;                         
REPEATED SUBJECT= subject-effect < / options > ;  
RUN;

The functions of these statements and their relevant options will be 
covered in the several subsequent sections in the GENMOD web-page.

When working with classification data, the CLASS statement functions much 
like it does in PROC MIXED or GLM. That is, it applies GLM coding (dummy 
variables equal to 0 or 1) to the levels of the factors with the highest 
coded value serving as the reference category. Effect coding is also 
available, where the levels are set to -1 or +1 which is achieved with 
PARAM= option on the CLASS statement:

  CLASS fctr / PARAM= effect;

The FREQ statement is necessary when records containing variables with 
identical values are stored in one record in the file containing a 
variable which indicates how many duplicate records this one record 
represents.

Other statements include the ability to define your own models (i.e. you 
define the link, variance and deviance functions). These statements allow 
you to enter special mathematical relationships for the mean and variance 
of the response.

Several other SAS procedures are very helpful to summarize and explore the 
distributions of the data entered into or produced by PROC GENMOD. You 
should also be very familiar with the PRINT, SORT, TABULATE, TRANSPOSE, 
FREQ, and PLOT procedures, which are among the most important ones to 
know.


GENMOD Compared with other SAS Procedures

Other SAS procedures also compute generalized linear models. In some cases 
the outputs are exactly the same as GENMOD, others include important 
features not necessarily part of or even computable with GENMOD. For the 
various probability distributions, these procedures include:

Normal Distribution: PROCs REG, ANOVA, GLM, MIXED

Binomial and Multinomial: LOGISTIC, PROBIT, CATMOD

PROCs NLMIXED and GLIMMIX compute generalized linear models, but also 
include the capability to incorporate random effects, much like the types 
of problems PROC MIXED can handle with random effects for continuous data. 
One of the features of PROC NLMIXED is the ability to enter the likelihood 
function directly into a set of statements which allows you to work with 
zero-inflated count data or censored continuous data, as important 
examples. How these other procedures relate to PROC GENMOD will be 
incorporated into several examples which follow.