Operating characteristics of stepwise covariate selection in pharmacometric modeling

Stepwise covariate modeling (SCM) is a widely used tool in pharmacometric analyses to identify covariates that explain between-subject variability (BSV) in exposure and exposure-response relationships. However, this approach has several potential weaknesses, including over-estimated covariate effect and incorrect selection of covariates due to collinearity. In this work, we investigated the operating characteristics (i.e., accuracy, precision, and power) of SCM in a controlled setting by simulating sixteen scenarios with up to four covariate relationships. The SCM analysis showed a decrease in the power to detect the true covariates as model complexity increased. Furthermore, false highly correlated covariates were frequently selected in place of or in addition to the true covariates. Relative root mean square errors (RMRSE) ranged from 1 to 51% for the fixed effects parameters, increased with the number of covariates included in the model, and were slightly higher than the RMRSE obtained with a simple re-estimation exercise with the true model (i.e., stochastic simulation and estimation). RMRSE for BSV increased with the number of covariates included in the model, with a covariance parameter RMRSE of almost 135% in the most complex scenario. Loose boundary conditions on the continuous covariate power relation appeared to have an impact on the covariate model selection in SCM. A stricter boundary condition helped achieve high power (> 90%), even in the most complex scenario. Finally, reducing the sample size in terms of number of subjects or number of samples proved to have an impact on the power to detect the correct model.