[Paper Review] Estimating Bayesian Optimal Treatment Regimes for Dichotomous Outcomes using Observational Data

Thomas Klausch, Peter van de Ven+3 more|arXiv (Cornell University)|Sep 18, 2018

Advanced Causal Inference Techniques3 references3 citations

TL;DR

This paper proposes a Bayesian decision-theoretic framework for estimating optimal treatment regimes (OTR) in observational studies with dichotomous outcomes, using a loss function defined on the bivariate distribution of potential outcomes. It reduces unnecessary chemotherapy use in oropharyngeal cancer by 75% without sacrificing survival, offering a clinically meaningful trade-off between treatment burden and outcome efficacy through posterior predictive inference and sensitivity analysis on unobserved potential outcome associations.

ABSTRACT

Optimal treatment regimes (OTR) are individualised treatment assignment strategies that identify a medical treatment as optimal given all background information available on the individual. We discuss Bayes optimal treatment regimes estimated using a loss function defined on the bivariate distribution of dichotomous potential outcomes. The proposed approach allows considering more general objectives for the OTR than maximization of an expected outcome (e.g., survival probability) by taking into account, for example, unnecessary treatment burden. As a motivating example we consider the case of oropharynx cancer treatment where unnecessary burden due to chemotherapy is to be avoided while maximizing survival chances. Assuming ignorable treatment assignment we describe Bayesian inference about the OTR including a sensitivity analysis on the unobserved partial association of the potential outcomes. We evaluate the methodology by simulations that apply Bayesian parametric and more flexible non-parametric outcome models. The proposed OTR for oropharynx cancer reduces the frequency of the more burdensome chemotherapy assignment by approximately 75% without reducing the average survival probability. This regime thus offers a strong increase in expected quality of life of patients.

Motivation & Objective

• To develop a Bayesian framework for estimating optimal treatment regimes (OTR) in observational data with dichotomous outcomes.
• To incorporate patient-level treatment burden and outcome trade-offs via a flexible loss function on potential outcomes.
• To enable decision-making that balances survival probability and unnecessary treatment toxicity.
• To evaluate the method using simulations and a real-world oropharyngeal cancer dataset with known confounders.
• To provide posterior summaries of expected loss, survival, and treatment assignment probabilities for clinical decision support.

Proposed method

• The method uses a Bayesian decision-theoretic approach where the optimal treatment is selected to minimize the posterior expected loss.
• It models the joint distribution of potential outcomes (Y(1), Y(0)) under a conditional parametrization of the loss function, allowing for covariate-dependent loss structures.
• Posterior inference is based on two marginal Bayesian models for potential outcomes, enabling use of standard techniques like model selection and prior specification.
• The approach incorporates sensitivity analysis on the unobserved partial association between potential outcomes using a copula-based prior.
• It applies both parametric (logistic regression) and non-parametric (Bayesian Additive Regression Trees, BART) outcome models to assess robustness.
• Treatment assignment rules are derived from posterior predictive distributions, with uncertainty quantified via credible intervals and posterior probabilities.

Experimental results

Research questions

• RQ1 Can a Bayesian framework improve optimal treatment regime estimation by explicitly modeling treatment burden alongside survival outcomes in observational data?
• RQ2 How does the inclusion of a loss function that penalizes unnecessary treatment affect the resulting treatment regime compared to standard outcome-only optimization?
• RQ3 What is the impact of model misspecification and sample size on the bias and coverage of posterior estimates in OTR estimation?
• RQ4 How does the proposed method perform in reducing chemotherapy use in oropharyngeal cancer while maintaining survival rates?
• RQ5 To what extent can posterior uncertainty and credible intervals inform clinical decision-making for individualized treatment?

Key findings

• The proposed OTR reduced chemotherapy assignment by approximately 75% compared to the observed regime, from 55.2% to 13.4% of patients, without decreasing average survival probability.
• The average survival probability under the OTR was 75.9%, slightly higher than the observed regime’s 72.8%, indicating no loss in outcome efficacy.
• Posterior probabilities showed high certainty that radiotherapy alone was optimal for many patients, particularly in low-risk subgroups.
• Simulations revealed optimism bias in small samples, especially under low heterogeneity, but coverage was nominal for large samples in logistic models and conservative for BART.
• The method achieved strong trade-offs between minimizing treatment burden and maintaining survival, with posterior means and credible intervals providing robust decision support.
• The conditional parametrization of the loss function enabled clinically relevant penalization of unnecessary treatment, aligning with real-world clinical priorities.