Plots for Influential and Poorly Fitting Covariate Patterns in the LOGISTIC Procedure

Logistic regression is a popular tool among epidemiologists and other data analysts and while the SAS~ system is widely used, few investigators make use of the readily available plots that assess model fit. The purpose of this presentation is to demonstrate the steps necessary to generate these plots. We provide a brief overview of logistic regression diagnostics and introduce a sample data set. Using the output statement in PROC LOGISTIC we show how to obtain the estimated logistic probabilities, change in Pearson chi-square, and Cook distance parameters so simple plots can be generated for assessing the presence of illfitting and/or influential observations. The plots provide a graphical approach for assessing model adequacy. While interpretation of the plots is somewhat subjective, these simple plots are a valuable tool for screening out inadequately fitting logistic regression models which may lead to incorrect or misleading inferences.