A1253
Title: Experimenting with finite to infinite populations
Authors: Jonathan Chipman - University of Utah (United States) [presenting]
Oleksandr Sverdlov - Novartis (United States)
Diane Uschner - Roche (United States)
Abstract: ANOVA is a common inferential strategy for randomized trials and assumes observations are drawn from an infinite population. However, trial participants are often considered a finite population based on the inclusion/exclusion criteria, location of the trial, and single-point timing of the trial. A finite population central limit theorem provides a degree of reassurance to assume normality when carrying out complete randomization with fixed equal allocation. Yet, in practice, many trials restrict randomization to reduce the risk of chronological bias by using a maximum tolerable imbalance procedure (MTI) or permuted block design (PBD). The impact on Type I error when using MTI or PBD for a finite population is not well studied. Through extensive simulations, common restrictions to randomization are observed to impact the Type I error convergence rate. When using ANOVA in a finite population, Type I error is more well controlled when implementing complete randomization than when using MTI or PBD. Randomization-based inference ensures an exact 5\% inference under all settings and is reflective of the population from which patients are drawn.
