Causal Inference Methods and Intermediate Endpoints in Randomized Clinical Trials

Abstract

In clinical research and randomized clinical trials, intermediate endpoints can serve several purposes. It is possible that an intermediate marker may serve as a surrogate S for a true clinical outcome of interest T with the goal of making the trial run more efficiently or cost-effectively. Rigorous assessment as to whether a proposed surrogate endpoint is valid is challenging, however.

Chapter II extends causal inference approaches to validate a candidate surrogate outcome using potential outcomes. Using the principal surrogacy criteria, we incorporate baseline covariates in the setting of normally-distributed endpoints. In particular, our setting of interest allows us to assume the surrogate under the placebo, S(0), is zero-valued. We develop methods to incorporate conditional independence and other modeling assumptions and explore their impact on the assessment of surrogacy. We demonstrate our approach via simulation of data that mimics an ongoing study of a muscular dystrophy gene therapy.

Chapter III also considers the motivating clinical trial for muscular dystrophy, whereas now the true outcomes T(0), T(1) are measured longitudinally. We develop a mixed model approach that can potentially gain estimation efficiency. Further, it may be possible to measure additional T and S outcomes in a delayed treatment start or cross-over trial design. In this situation, subjects who are first administered the placebo may be given the gene therapy at a later time. This chapter addresses models and metrics for validation in such a trial. We also consider how to define the quantities for validation such that they may depend on time.

In Chapter IV, we extend these ideas to the surrogate validation framework with time-to-event data. We develop a method that incorporates the censoring and semi-competing risk structure that is often encountered with multiple survival endpoints. We consider novel ways to define the parameters measuring the association between outcomes and relevant principal strata using a illness-death framework. We model conditional hazards while maintaining a valid causal interpretation by viewing this through the lens of a causal multi-state model. Finally, we apply our proposed methods to a prostate cancer randomized clinical trial.