Logistic regression has become the standard method for analyzing binary, multinomial or ordinal scale response data in epidemiologic research. An important, but often overlooked, assumption concerning the continuous independent variables in the model is that they must be "linear in the logit". If this assumption is violated, the variable cannot be included "as is". Instead, some appropriate transformation or categorization of the variable must be considered. In this talk we will consider various methods for determining whether a variable is linear in the logit and will consider modeling implications when it is not.
Stanley Lemeshow (Professor and Dean of the College of Public Health at The Ohio State University, Columbus)