Primary Submission Category: Randomized Studies
Sequential Adaptive Designs that Learn Optimal Individualized Treatment Rules by Utilizing Surrogate Outcomes
Authors: Wenxin Zhang, Aaron Hudson, Maya Petersen, Mark van der Laan,
Presenting Author: Wenxin Zhang*
Randomized trials with covariate-adjusted response-adaptive (CARA) designs can be appealing because they allow assigning less patients with inferior treatments based on prior patients’ responses to treatment and current patients’ covariates. To implement a CARA design, one requires the outcome to be observed shortly after treatment is administered. However, when there is a long follow-up period until the outcome of interest is observed, there may be insufficient information to learn the treatment effect on the outcome conditional on patients’ covariates, therefore making it challenging to implement a CARA design. In this work, we study a setting in which multiple surrogate outcomes are observed in advance of the final outcome of interest. One can then consider implementing a CARA design by assigning treatment based on the conditional average treatment effect on the surrogate outcomes. We discuss benefits and drawbacks of using surrogate outcomes with different follow-up time in a CARA design. And we propose a target causal parameter to evaluate utility of a surrogate in this setting and estimate that under the targeting maximum likelihood estimation (TMLE) framework. We also develop a CARA design with a data-adaptive strategy for choosing and utilizing the optimal surrogate to assign treatments. We illustrate the performance of our proposed adaptive design in terms of minimizing the chance of participants receiving inferior treatments through a range of simulation studies.