New proposed criteria for the clinical diagnosis of Alzheimer's disease increasingly incorporate biomarkers, most of which are normally measured on a continuous scale. Operationalizing such criteria thus requires continuous biomarkers to be dichotomized, which in turns requires the selection of a cut-point at which to dichotomize. In this article, we review the statistical principles underlying the choice of cut-points, describe some of the most commonly adopted statistical approaches used to estimate cut-points, highlight potential pitfalls in some of the approaches and characterize in what sense the estimated cut-point from each approach is optimal. We also emphasize that how a cut-point is selected must be made in reference to how the resulting dichotomized biomarker is to be used, and in particular what actions will follow from a positive or negative test result.