A Unified Causal Inference Framework for the Desirability of Outcome Ranking Paradigm in Benefit-Risk Evaluation

Research Question and Approach

This paper addresses the challenge of evaluating treatments in clinical trials where efficacy and safety must be balanced simultaneously. The authors focus on the Desirability of Outcome Ranking (DOOR) paradigm, which integrates multiple clinical endpoints into a single, patient-centric ordinal scale. To estimate the DOOR probability—the probability that a patient on one treatment has a more desirable outcome than a patient on another—the authors develop a unified causal inference framework. This framework uses sequential risk-set hazards to model ordinal outcomes and derives the efficient influence function (EIF) to enable robust statistical inference. The researchers evaluated several estimators, including G-computation, inverse probability weighting (IPW), augmented IPW (AIPW), and Targeted Maximum Likelihood Estimation (TMLE), using both generalized linear models and the Super Learner ensemble method for nuisance function estimation.

The study finds that TMLE combined with Super Learner (TMLE-SL) consistently outperforms other methods in point estimation. When extending the framework to inference, the authors demonstrate that cross-fitted TMLE (CVTMLE-SL) provides the strongest performance regarding bias reduction, recovery of underlying ordinal distributions, and confidence interval coverage. The framework is shown to be flexible, accommodating complex relationships between covariates, treatment mechanisms, and ordinal outcomes without requiring rigid parametric assumptions like the proportional-odds model.

Traditional clinical trial analyses often evaluate efficacy and safety endpoints separately, which can obscure the overall clinical experience of a patient. By providing a rigorous, covariate-adjusted framework for DOOR, this research enables more reliable benefit-risk assessments in both randomized trials and observational studies. The use of machine-learning-based nuisance estimation (Super Learner) and doubly robust methods (TMLE) ensures that the resulting estimates are less sensitive to model misspecification, making the approach highly suitable for complex biomedical data where treatment effects and patient characteristics are heterogeneous.