Automated forward selection for Generalized Linear Models with Categorical and Numerical Variables using PROC GENMOD

Generalized linear models are a powerful tool to measure relationships between variables, as they can handle nonnormal distributions without altering the properties of variables involved. When applied to risk factor analysis, they can help determine the most important factors contributing to the incidence, prevalence or acquisition of a particular medical condition. This paper presents a particular case in which the aforementioned factors are unknown and a selection must be made from a pool containing both numerical and class variables. Since the model uses an option that is only present in the GENMOD procedure (and not in the LOGISTIC procedure, for example), an algorithm for selecting variables needed to be created from scratch. The proposed macro was built in such a case. Several factors, both numerical and categorical, were tested using forward selection and defined criteria for entering the model and for keeping the variable in the model. The macro also selects the numeric and categorical variables to include only the later in the class statement of the PROC GENMOD.