A Sequential Multiple Assignment Randomized Trial Design with Response-Adaptive Tailoring Function

We present a novel sequential multiple assignment randomized trial (SMART) design that integrates response-adaptive randomization with tailoring functions (RA-TF-SMART). We develop percentile-based and Z-score RA-TFs that incorporate both within-patient and between-patient adaptation to map continuous outcomes to randomization probabilities. We apply Q-learning, tree-based reinforcement learning, and G-estimation to estimate dynamic treatment regimens (DTRs). We compare our RA-TF-SMART designs to balanced randomized SMARTs (BR-SMARTs), tailoring function SMARTs (TF-SMARTs), and generalized outcome-adaptive SMARTs (GO-SMARTs). This study addresses limitations in SMART methodology by presenting designs where randomization probability does not require dichotomization of continuous outcomes and utilizes both individual patient outcomes and accumulated treatment efficacy data from prior participants. RA-TF-SMARTs offer a flexible framework that maximizes benefit for trial participants while maintaining statistical validity for post-trial DTR estimation.