B2035
Title: Selecting an optimal cutpoint in Cox proportional hazards models with several covariates
Authors: Woojoo Lee - Inha University (Korea, South) [presenting]
Abstract: In survival analysis, a continuous covariate sometimes needs to be transformed to a binary variable to enhance interpretation of regression coefficients. For doing this, the key problem is to find an optimal cutoff and assess the statistical significance of the transformed binary variable correctly. A naive approach is using the cutoff giving the most significant p-value for the binary variable, but this shows highly distorted type I error. In order to overcome such problem, two novel testing methods were developed based on Brownian or Brownian bridge process. Although these methods assumed that there is only one continuous covariate in Cox proportional hazards model, which is often violated in practice, they are currently employed as standard methods even when there are several covariates. We investigate the performance of the two methods when there are several covariates in Cox proportional hazards model. Our numerical study shows that the two testing methods may not provide an optimal cutoff and suffer from distorted type I error. To mitigate these problems, we adopt a more recent testing method that can take into account other covariates in the Cox model. The numerical study shows that the recent testing method controls type error well at the required nominal level and shows smaller mean squared error of cutoff estimator than the two methods developed for only one continuous covariate case.
