The Logic and Logistics of Logistic Regression

Abstract

Although logistic regression models are widely used in multivariable analyses with dichotomous outcomes, many of their features, which can be very helpful tools in better understanding the data, are often underutilized. In addition to the widely used and reported odds ratios and p-values, PROC LOGISTIC generates a plethora of statistics from which one can gain further insight, make stronger analytical inferences, and more easily identify errors in the model construction. Model options will be explored, and resulting outputs from real-world examples will be explained in detail. Topics covered include model fit statistics, maximum likelihood estimates, effect modification, Receiver Operator Characteristic (ROC) curves, and the Hosmer-Lemeshow Goodness of Fit Test.