UCLA

Logistic regression has been widely employed in the life sciences where the response variable of interest is the presence or absence of some characteristic or condition. Despite the popularity of logistic regression approaches and the simplicity that comes with implementing methods in software, the tools in place for model evaluation remain rather limited. Here we explore goodness-of-fit assessment for logistic regression models. In particular, following a review of numerical summaries such as likelihood ratio tests, Hosmer-Lemeshow tests, information criteria, and residual deviance, we focus on plots of fitted versus actual percentages and explore how the power of such graphical tests appears to depend on the choice of bin size.