Blinded sample size recalculation in randomized controlled trials with analysis of covariance

TL;DR

This paper introduces a new blinded sample size recalculation method for ANCOVA in randomized trials that maintains power and controls error rates even under model misspecification, with practical implementation tools.

Contribution

It proposes a distribution-agnostic, blinded recalculation approach for ANCOVA-based sample size determination, addressing limitations of existing methods.

Findings

Method achieves nominal power in simulations.

It controls type I error without inflation.

Applied successfully to HIV trial data.

Abstract

In randomized controlled trials, covariate adjustment can improve statistical power and reduce the required sample size compared with unadjusted estimators. Several regulatory agencies have released guidance on covariate adjustment, which has recently attracted attention in biopharmaceutical research. Analysis of covariance (ANCOVA) is often used to adjust for baseline covariates when outcomes are continuous. To design a sample size based on ANCOVA, it is necessary to prespecify the association between the outcome and baseline covariates, as well as among the baseline covariates themselves. However, determining these parameters at the design stage is challenging. Although it may be possible to adaptively assess these during the trial and recalculate the required sample size, existing sample size recalculation methods assume that the joint distribution of the outcome and baseline covariates is multivariate normal, which is not always the case in practice. In this study, we propose a blinded sample size recalculation method for the ANCOVA estimator based on the asymptotic relative efficiency under minimal distributional assumptions, thus accommodating arbitrary model misspecification. The proposed method is able to achieve the nominal power and reduce the required sample size without inflating the type I error rate in the final analysis. We conducted simulations to evaluate the performance of the proposed method under various scenarios and applied it to data from an HIV clinical trial. We provide R and SAS macro codes to implement the proposed method.