Our novel application of machine learning meaningfully improves the analytic precision in randomized trials. This work can improve trial design: smaller sample sizes for the same statistical power (the probability of finding a significant effect). This work can also improve trial analysis: greater statistical power for the same sample size. Our results debunk some common misconceptions in trial analysis. In particular, our approach is model-robust; there is no risk of introducing bias. Our approach is guaranteed to improve efficiency; our data-adaptive approach avoids forced adjustment that is harmful to precision. Our approach is fully pre-specified, obviating concerns about a “fishing expedition” to find the most favorable analysis. In simulation studies and a real data analysis, we show meaningful improvements regardless of the outcome-type (i.e., for both binary and continuous outcomes) or randomization scheme (i.e., for both simple and stratified randomization).
March 6, 2024