EVAULATING THE ACCURACY OF A PREDICTIVE MODEL IN SURVIVAL ANALYSIS

Background. Risk prediction models can be used as an aid when determining patient management. Because incorrect predictions could have substantial consequences, it is very important to evaluate the predictive accuracy of a model. When the risk prediction model is the result of a survival analysis where the outcomes are potentially censored, there is no standard way of evaluating predictive accuracy. There have been many different measures proposed for this purpose, but they all have limitations. The most frequently used measure in practical applications is the concordance index (c-index) proposed by Harrell et al. (1982). The c-index, however, is biased when the data are censored, a problem that does not appear to be widely recognized. The purpose of this poster is to highlight and explore this bias. Methods. To study the properties of the c-index when data are censored, we conducted a simulation study. We simulated failure times and predictions from a risk model by generating data from a binormal distribution, which allowed us to prespecify the true level of concordance. We generated censoring times from a uniform distribution. By varying the various distribution parameters, the true level of concordance, the percentag