Efficient screening of covariates in population models using Wald's approximation to the likelihood ratio test

Abstract

We propose an efficient algorithm for screening covariates in population model building using Wald's approximation to the likelihood ratio test (LRT) statistic in conjunction with Schwarz's Bayesian criterion. The algorithm can be applied to a full model fit of k covariate parameters to calculate the approximate LRT for all 2k - 1 possible restricted models. The algorithm's efficiency also permits internal validation of the model selection process via bootstrap methods. We illustrate the use of this algorithm for both model selection and validation with data from a Daypro pediatric study. The algorithm is easily implemented using standard statistical software such as SAS/IML and S-Plus. A SAS/IML macro to perform the algorithm is provided.