Online Program

Type 3 Statistics in SAS Procedures: What Do They Really Mean? (303028)

Keywords: categorical, logistic, Type 3 statistics, GEE, SAS

Many SAS procedures that model categorical outcomes produce type 3 statistics to evaluate individual effects. For example, given two binary predictors in a logistic regression, part of the default PROC LOGISTIC output is a table of type 3 analysis of effects with a Wald chi-square, degrees of freedom, and probability for each effect. This is essentially a test of the joint effect of the parameters. As such, these tests of effects should be independent of how the categorical variables are coded and of the order of specification in the model statement. This is, however, not what happens when interaction is included. Using indicator vs. effect coding or changing the referent group results in different Wald chi-square statistics (and significance levels) for the main effects of predictors, although the statistic for the interaction is the same in all formulations. Though one could argue only the interaction is important, the change in significance of individual effects as the coding scheme changes is disconcerting. We explore what the type 3 analysis of effects is doing, how these results compare to those from the ESTIMATE/CONTRAST, and how to interpret the results.