Extension of individual model averaging assessments to unbalanced designs and dose-response

Abstract

1Recent investigations assessed two non-linear mixed effect (NLME) model based approaches to test for drug effect on real data in the context of balanced two-arms designs. The standard approach (STD) showed type I error inflation and biased drug effect estimates contrary to the proposed alternative, individual model averaging (IMA), which had controlled type I error and unbiased drug effect estimates. The current study is an extension of the performances assessment of these two approaches to unbalanced designs and dose-response studies. The type I error rate and drug effect estimates were assessed for unbalanced designs, using placebo Alzheimer disease assessment scale cognitive (ADAS-cog) scores from 800 individuals. The bias in the drug effect estimates was assessed for dose response scenarios, on data modified by the addition of various dose-response scenarios (Emax= 2.5, 5, and 10). The generalization of IMA to any randomization ratio of two-arms studies was also presented, together with an alternative parameterization of IMA: saturated IMA (sIMA). Similarly to what was observed in balanced designs, both IMA and sIMA had controlled type I errors and unbiased drug effect estimates in unbalanced designs, whereas STD had uncontrolled type I error and biased drug estimates. For the dose-response studies STD had a systematic bias towards the underestimation of the drug effect estimates. IMA and sIMA were unbiased in the scenarios with high maximum effect but their performances were hindered at the lowest maximum drug effect scenario, because of the closeness in magnitude between the drug effect addition and the placebo model misspecification.