During covariate modeling in pharmacometrics, computational time can be reduced by using a fast preselection tool to identify a subset of promising covariates that are to be tested with the more computationally demanding likelihood ratio test (LRT), which is considered to be the standard for covariate selection. There is however a lack of knowledge on best practices for covariate (pre)selection in pharmacometric repeated time-to-event (RTTE) models. Therefore, we aimed to systematically evaluate the performance of three covariate (pre)selection tools for RTTE models: the likelihood ratio test (LRT), the empirical Bayes estimates (EBE) test, and a novel Schoenfeld-like residual test. This was done in simulated datasets with and without a "true" time-constant covariate, and both in the presence and absence of high EBE shrinkage. In scenarios with a "true" covariate effect, all tools had comparable power to detect this effect. In scenarios without a "true" covariate effect, the false positive rates of the LRT and the Schoenfeld-like residual test were slightly inflated to 5.7% and 7.2% respectively, while the EBE test had no inflated false positive rate. The presence of high EBE shrinkage (> 40%) did not affect the performance of any of the covariate (pre)selection tools. We found the EBE test to be a fast and accurate tool for covariate preselection in RTTE models. The novel Schoenfeld-like residual test proposed here had a similar performance in the tested scenarios and might be applied more readily to time-varying covariates, such as drug concentration and dynamic biomarkers.
Keywords: covariate model building; empirical Bayes estimate; non-linear mixed effects modeling; repeated time-to-event.