Leveraging Baseline Information to Improve Inference in Adaptive Randomized Experiments with Small Sample Size

• Activity Number: 423
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Po-Han Brian Chen*+ and Rosenblum Michael
• Companies: The Johns Hopkins University and The Johns Hopkins University
• Keywords: Covariate-adaptive design ; Confidence interval ; Baseline information ; Small sample ; Randomized trial

Abstract

We investigate how baseline variables can be used to improve inference in randomized experiments with small sample size. We focus on constructing small sample 95% confidence intervals for the mean treatment effect. When there is no baseline information (or it is ignored), standard exact methods for constructing confidence intervals can be used but they are generally conservative. We aim to incorporate baseline information to construct valid confidence intervals with shorter widths. First, we explore how much efficiency can be gained using standard randomization but doing an analysis adjusting for the baseline variables. Second, we explore covariate-adaptive randomization that attempts to balance predictive variables between the study arms. In our algorithm for constructing confidence intervals, the more predictive the baseline variables are, the shorter the resulting confidence intervals. We do a simulation study to explore how much shorter we can make our confidence intervals than those from exact methods that ignore baseline variables, and still get correct coverage probability.