Abstract: |
Maximal chi-squared methods and receiver operating characteristic (ROC) curves-based methods are commonly used to dichotomize a continous predictor when the outcome is binary. In this article we consider this problem when the outcome of interest is subject to right censoring, with a twofold goal. First, we propose maximal concordance, a measure similar to the area under an ROC curve, as a metric for selecting an optimal cutpoint with censored endpoints. To support this, we show that selecting the cutpoint that maximizes the concordance probability is equivalent to maximizing the Youden index, a popular criterion when the ROC curve is used to choose a threshold with binary outcomes. Second, we compare the performance of two concordance-based metrics (c-index and concordance probability estimate) and the performance of three chi-squared-based metrics (Wald, log-rank, and partial likelihood ratio statistics). In our simulations performed under a variety of assumptions, maximizing the partial likelihood ratio test statistic has the best performance. © 2013 Copyright Grace Scientific Publishing, LLC. |