Causal Inference
Application and description of Bayesian doubly robust conditional quantile treatment effect estimation: A use case with sensitivity analysis Hayden L. Smith* Hayden Smith Smith Smith Smith Smith Smith Smith UnityPoint Health – Des Moines
Background: Treatment effect heterogeneity is a central topic in precision medicine and policy evaluation. Common means-based causal effect estimators (e.g., conditional average treatment effects (CATE)) summarize outcomes at the population level and fail to capture distributional differences. Doubly robust conditional quantile treatment effects (CQTE) assist in overcoming this limitation.
Methods: A Bayesian CQTE approach was used to examine for effect modification (EM) in an analgesia-first sedation protocol implemented at a Midwestern Intensive Care Unit (ICU) in the United States. The study had a pre-/post-study design examining changes in hourly morphine milligram equivalents (MME) administered to patients with mechanical ventilation. The EM variable of interest was whether or not the patient had a primary complaint related to a respiratory condition. Stabilized inverse propensity scores were estimated via Bayesian additive regression trees and a conditional outcome model fit separately for the pre- and post- intervention groups using Bayesian quantile regression (i.e., tau: 0.1-0.9) with asymmetric Laplace likelihoods. Posterior predictions were combined with propensity scores to construct doubly robust estimates aggregated by the EM variable.
Results: Post-protocol MME changes in the ICU revealed heterogeneity in treatment effects across the outcome distributions (Figure). Specifically, effect modifier subgroups showed differences across quantiles, indicating different benefits based on the modifier. Sensitivity analyses will be presented at the conference and will include comparisons to QTE, CATE, and ATE, as well as varied propensity score and uncertainty methods.
Conclusion: CQTE provide a flexible framework for evaluating heterogeneity across outcome distributions in non-randomized settings. By combining propensity score weighting with conditional quantile modeling, the process mitigates confounding bias and allows for subgroup-specific inference.
