Statistical Considerations for Clinical Trials Aiming to Identify Individualized Treatment Rules

Abstract

Clinical trials have traditionally been designed to identify the best treatment on average or a treatment that results in benefit for a majority of a population. However, treatment that works well for a majority may not work at all for the minority, which motivates the need for personalized treatment rules in modern clinical trials. For a single stage randomized control trial, identification of an individualized treatment rule (ITR) is often set as an aim, particularly in behavioral intervention trials. ITRs assign treatment as a function of patients' clinical information which contrasts with a static treatment rule that assign everyone the same treatment. Much of the focus on ITRs revolves around identifying rules that are close to a theoretical optimal rule, which could lead to identifying rules that perform worse than the optimal static rule particularly in the absence of substantial effect heterogeneity. This limitation motivates new methods that reliably recommend the estimated optimal static rule when evidence of effect heterogeneity is lacking, and considerations for sample size regarding reliable identification of a beneficial ITR, which is an ITR that performs better than the optimal static rule. To address these limitations, we introduce a Monte Carlo integration based calculation of the probability to identify a beneficial ITR which requires specification of a data generating model. We also introduce an approach to selecting the penalty parameter in a LASSO model such that the static rule is identified with high probability in the absence of treatment effect heterogeneity which mitigates the risk of identifying a harmful ITR. A Dynamic Treatment Regime (DTR) is a clinical tool to guide the treatment decisions of clinicians which assigns treatment at each decision point over time based on patient characteristics including prior response to treatment. A Sequential Multiple Assignment Randomized Trial (SMART) aims to identify optimal DTR through randomization at multiple time points. We introduce beneficial DTRs as DTRs that performs better than the estimated optimal embedded DTR. This definition implies that in the absence of treatment effect heterogeneity, an identified more deeply tailored DTR would be harmful in a population. To address this, we introduce a permutation test method to select the penalty parameter in a LASSO model such that no treatment interaction coefficients are selected for regression using Q-learning in the absence of treatment effect heterogeneity with specified probability at each stage of treatment assignment. The use of Q-learning, however, presents challenges in that the stage one model is frequently incorrectly specified. IQ-learning avoids this by not directly modeling the Q-function at the first stage of treatment. However, variable selection methods have not been considered when using IQ-learning and we apply a group LASSO method where the penalty parameter is selected through the same permutation-based methods. We apply all of our methods to two separate SMARTs, the Program for LUng Cancer Screening and TObacco Cessation (PLUTO) which aims to identify DTRs to assist with smoking cessation and the M-Bridge study which aimed to estimate an optimal DTR to prevent binge drinking in college freshman.