Missing data mechanisms in a dose-finding adaptive trial

Abstract

Although adaptive trials have become popular, little research has been done to investigate the effect of missing data in adaptive trials. We consider three different types of missing data mechanisms-missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR)-and introduce the "mixture missing mechanism" (MMM) to an adaptive three-period crossover study that uses the "maximizing procedure." These results are compared to a traditional nonadaptive equal allocation crossover study. Simulations suggest that certain missing data mechanisms can result in biased estimates. For equal allocation, the bias is uniform between treatments so treatment comparisons are unbiased. However, for the maximizing procedure, the bias is not uniform between treatments, so treatment comparisons are biased. For the MNAR and MMM mechanisms, unusually large bias occurs in the placebo group, leading to a substantial loss of power.