In this article, we present a technique and critical test statistic for assessing the fit of a binary-dependent variable model (e.g., a logit or probit). We examine how closely a model's predicted probabilities match the observed frequency of events in the data set, and whether these deviations are systematic or merely noise. Our technique allows researchers to detect problems with a model's specification that obscure substantive understanding of the underlying data-generating process, such as missing interaction terms or unmodeled nonlinearities. We also show that these problems go undetected by the fit statistics most commonly used in political science.