Conditional and Unconditional Treatment Effects in Randomized Clinical Trials: Estimands, Estimation, and Interpretation

Abstract

ICH E9(R1) specifies the importance of precisely defining the treatment effect for clinical trials—to inform patient choices and facilitate evidence-based decision-making. FDA’s guidance on covariate adjustment encourages the judicious use of baseline covariates to enhance efficiency. Careful consideration is required when adjusting for covariates in nonlinear models such as logistic regression and Cox regression. For these nonlinear models, including baseline covariates can change the targeted treatment effect (estimand). It is also crucial that the proposed statistical analyses align with the estimand of interest. Covariate-adjusted estimators used for unconditional treatment effect are typically robust to misspecification of the used regression models. Despite extensive literature and recommendations by the FDA on the statistical theory and properties of these methods, the real-life application of such methodologies is still limited. Moreover, a few open questions require further discussion, especially for time-to-event outcomes. In this article, we present causal estimands definition of conditional and unconditional treatment effects in randomized clinical trials. We also evaluate different estimation methods for estimating conditional and unconditional treatment effects for the binary and time-to-event endpoints. We also discuss practical considerations in choosing the conditional versus unconditional effect, implementing the estimation methods, and interpreting results.