Statistical Methods to Integrate Data to Improve Efficiency and Generalizability of Results from Sequential Multiple Assignment Randomized Trials

Patient heterogeneity necessitates treatment recommendations tailored to individual patient characteristics so the right treatment is given to the right person at the right time. Dynamic treatment regimes (DTRs) are treatment rules based on an individual's characteristics that may change over time, such as disease progression and response to previous treatments. Sequential multiple assignment randomized trials (SMARTs) are the gold standard for estimating optimal DTRs, but have limitations, such as sample size requirements and selection bias. We explore methods to integrate high quality observational data with SMART data to improve efficiency and generalizability. Motivation for this research comes from the personalized medicine aims of the Back Pain Consortium Research Program. One of its aims is to elucidate associations between patient characteristics and response to treatment. Clinicians and investigators Consortium-wide defined and developed standard operating procedures for collecting harmonized measures across different data domains for chronic low back pain, with the goal of integrating data across studies. We describe the rationale for these harmonized measures and associated data standards. To improve the efficiency of estimating optimal DTRs from SMARTs, we propose the multi-stage augmented Q-learning estimator (MAQE), which integrates observational study data with SMART data. The MAQE adapts a doubly robust augmentation estimator and embeds it in the Q-learning framework. Simulation studies show it has a similar or higher rate of correctly classifying test set patients to the optimal treatment sequence, and has a similar or higher value function, as compared with a standard Q-learning estimator without augmentation. Calibration weights (CWs) have been used in various applications to improve generalizability of trial results to a target population. CWs are generated such that the first and second moments of the covariate distributions in the trial will empirically match those of the target population. We incorporate stage-wise CWs in the analysis of a SMART, using summary statistics from an observational study assumed to have the same covariate profile as the target population to generate the weights. We show via extensive simulations that estimators that incorporate these stage-wise weights into estimation of optimal DTRs have better performance than standard Q-learning and the MAQE.