A simulation study comparing performance of minimum P-value and Youden index as methods to find optimal cut-points of continuous variables

INTRODUCTION. In recent years the evaluation of the ability of a diagnostic (prognostic) variable to distinguish a diseased (at risk) from a non-diseased patient (not at risk) has been widely discussed. In the presence of a continuous variable, the clinical decision-making process often uses for classification a cut-point value. From the statistical perspective the critical point arises as to how determine this threshold value. Two analytical methods are often used in order to categorize continuous variables: the minimum P-value [1] - based on the maximization of a chi-square statistic - and the Youden index [2] method. The performance of these approaches has so far not been extensively compared.OBJECTIVES. Aim of our work is to compare the performance of the minimum P-value and the Youden index methods through a simulation study. Indeed these approaches are mathematically related. The chi-square statistic of the minimum P-value approach is a transformation of the Youden index function, accounting for variance in parameter estimation. CONCLUSIONS. We have presented a simulation study aimed to compare two common methods used to define cut-points of new biomarkers: the minimum P-value and the Youden index approach. We have shown that under the biomarker normality distribution assumption the Youden index approach performs better than the minimum P-value approach. The difference in performance between the minimum P-value and the Youden approach is due to the variance component included in the chi-square statistic. This aspect, that could be intuitively thought as an advantage of the minimum P-value approach, refers to the null hypothesis of absence of association between the true disease status and the classification variable. However, the identification of cut-points for dichotomization has a start point the possible presence of some discrimination potential of the variable needing categorization.