Jointly modeling multiple endpoints for efficient treatment effect estimation in randomized controlled trials

TL;DR

This paper introduces a joint modeling approach for multiple endpoints in randomized controlled trials, improving treatment effect estimation efficiency and robustness, demonstrated through tobacco research data.

Contribution

It develops a novel estimator that leverages multiple endpoints, enhancing efficiency and robustness compared to standard methods.

Findings

Reduces standard error by 27% in tobacco study

Provides a robust joint modeling framework for multiple endpoints

Improves confidence in treatment effect estimates

Abstract

Randomized controlled trials are the gold standard for evaluating the efficacy of an intervention. However, there is often a trade-off between selecting the most scientifically relevant primary endpoint versus a less relevant, but more powerful, endpoint. For example, in the context of tobacco regulatory science many trials evaluate cigarettes per day as the primary endpoint instead of abstinence from smoking due to limited power. Additionally, it is often of interest to consider subgroup analyses to answer additional questions; such analyses are rarely adequately powered. In practice, trials often collect multiple endpoints. Heuristically, if multiple endpoints demonstrate a similar treatment effect we would be more confident in the results of this trial. However, there is limited research on leveraging information from secondary endpoints besides using composite endpoints which can be difficult to interpret. In this paper, we develop an estimator for the treatment effect on the primary endpoint based on a joint model for primary and secondary efficacy endpoints. This estimator gains efficiency over the standard treatment effect estimator when the model is correctly specified but is robust to model misspecification via model averaging. We illustrate our approach by estimating the effect of very low nicotine content cigarettes on the proportion of Black people who smoke who achieve abstinence and find our approach reduces the standard error by 27%.