Methods for leveraging secondary endpoints in the analysis of randomized trials

Abstract

Randomized controlled trials are the gold standard for evaluating the efficacy of an intervention. However, enrolling human participants can be expensive and slow, and put participants at risk of negative outcomes. Careful study design and planning is needed in order to efficiently use participants' data and not enroll more participants than necessary. This leads to a balance between designing trials that are powered to answer the most important questions versus trials that address less relevant questions with fewer participants enrolled. For example, there is often a trade-off between selecting the most scientifically relevant primary endpoint versus a less relevant, but more powerful, endpoint. Additionally, subpopulation analyses are often considered to answer additional questions regarding treatment effect heterogeneity; however, powering trials for these analyses requires far more participants than are required to analyze the average treatment effect. This motivates a large literature on improving the efficiency of randomized trials by optimally using the information that each participant provides. In practice, trials often measure and analyze several secondary endpoints, which augment the analysis of the primary endpoint; if multiple endpoints demonstrate a similar treatment effect, one might be more confident in the results of the trial than if the primary endpoint was considered in isolation. However, there is limited work on leveraging information from secondary efficacy endpoints to increase efficiency. This dissertation develops two novel methods to estimate the average treatment effect on the primary endpoint while leveraging data from secondary endpoints. Chapter 2 considers how secondary endpoints can be used to enhance dynamic data borrowing methods to more precisely estimate subpopulation average treatment effects. The proposed model incorporates secondary endpoints into the likelihood to more readily discern whether two subpopulations have homogeneous effects. This results in smaller biases than conventional methods when subpopulations have heterogeneous effects, but higher efficiency when subpopulations have homogeneous effects. Chapter 3 develops a treatment effect estimator based on a constrained joint model for primary and secondary endpoints. This estimator gains efficiency over the standard treatment effect estimator when the model is correctly specified and is robust to model misspecification. Chapter 4 considers how these two methods can be used within tobacco regulatory science and analyzes data from a recent trial of very low nicotine content cigarettes. The average treatment effect on abstinence from smoking is efficiently estimated within two subpopulations: people who smoke menthol cigarettes and people who smoke non-menthol cigarettes. These methods empower trialists to more efficiently use participant data in answering high impact scientific questions.