Primary Submission Category: Design of Experiments
Evaluate Power and Sample Size in Efficient Randomized Control Trial Design
Authors: Lauren Liao, Emilie Højbjerre-Frandsen, Alejandro Schuler,
Presenting Author: Lauren Liao*
To design a prospective study, power and sample size calculations are necessary to ensure a successful randomized trial targeting an effect of interest. Traditionally, calculations from unadjusted estimators are most commonly used in application. While semi-parametrically efficient estimators have been proved to reduce the variance while yielding more robust results, prospective studies leveraging efficient estimators are limited. To encourage a more efficient trial design, Schuler (2021) proposed a new formula for power calculation to leverage historical data to estimate sample size needed for semi-parametrically efficient estimators. However, this formula can overestimate the power using targeted minimum loss based estimation (TMLE) as suggested for the difference-in-means estimator. Instead, we recommend leveraging prognostic covariate adjustment as suggested in Liao et al. (2023) in combination with TMLE to minimize the likelihood of having an underpowered study. We showcase the practical usage of the power calculation developed by Schuler (2021) in a simulation to consider different machine learning algorithms and data generating distributions. We also illustrate the power calculation and validation with a case study using data from Novo Nordisk A/S.