Risk stratification studies, which are aimed at early identification of patients at high risk of clinical complications, are based on comparatively complex techniques for the analysis of retrospectively collected clinical data. Many of these studies contain serious misconceptions and errors in their analysis. This paper is a methodological review which lists the most frequent errors and omissions in retrospective stratification studies. The paper explains the principle of receiver operator characteristics and positive and negative predictive accuracy, and demonstrates the necessity of using a representative patient population for risk stratification. The danger of misleading conclusions resulting from a multivariate analysis, and several suggestions on how to organise and evaluate these studies are also discussed.