Unbiased and Efficient Estimation of Causal Treatment Effects in Cross-over Trials

TL;DR

This paper introduces a causal inference framework for cross-over trials, demonstrating that simple Gaussian linear models can unbiasedly estimate treatment effects even when they do not fully capture the data's complexity.

Contribution

It develops a G-computation approach combined with weighted least squares for unbiased causal effect estimation in cross-over trials, applicable with simple models.

Findings

Unbiased estimates achieved with G-computation and weighted least squares.

Simple working models perform comparably to complex ones in simulations.

Cross-over trials can be analyzed efficiently without bias using straightforward models.

Abstract

We introduce causal inference reasoning to cross-over trials, with a focus on Thorough QT (TQT) studies. For such trials, we propose different sets of assumptions and consider their impact on the modelling strategy and estimation procedure. We show that unbiased estimates of a causal treatment effect are obtained by a G-computation approach in combination with weighted least squares predictions from a working regression model. Only a few natural requirements on the working regression and weighting matrix are needed for the result to hold. It follows that a large class of Gaussian linear mixed working models lead to unbiased estimates of a causal treatment effect, even if they do not capture the true data generating mechanism. We compare a range of working regression models in a simulation study where data are simulated from a complex data generating mechanism with input parameters estimated on a real TQT data set. In this setting, we find that for all practical purposes working models adjusting for baseline QTc measurements have comparable performance. Specifically, this is observed for working models that are by default too simplistic to capture the true data generating mechanism. Cross-over trials and particularly TQT studies can be analysed efficiently using simple working regression models without biasing the estimates for the causal parameters of interest.