Predictive Bayesian inference and dynamic treatment regimes.

Area of Science:

• Causal Inference
• Longitudinal Data Analysis
• Bayesian Statistics

Background:

• Optimal dynamic treatment regimes (DTRs) are crucial for personalized medicine.
• Model-based estimation of DTRs allows outcome comparisons but is prone to misspecification.
• Frequentist estimators face challenges in nonregular estimation problems.

Purpose of the Study:

• To develop a robust Bayesian approach for estimating optimal DTRs.
• To enable direct probabilistic comparisons between DTRs, including static regimes.
• To address model misspecification issues, particularly the 'null-paradox'.

Main Methods:

• Utilized dynamic programming and Monte Carlo integration within a Bayesian predictive framework.
• Introduced a 'null-robust' reparametrization for longitudinal settings to ensure correct inference under the null hypothesis.
• Justified and incorporated inverse probability of treatment weighting (IPTW) within the Bayesian setting for confounding control.

Main Results:

• The proposed Bayesian approach circumvents issues with frequentist estimators.
• The 'null-robust' reparametrization ensures valid inferences even with potential model misspecification.
• Causal inference is framed as posterior predictive inference, facilitating confounding control via IPTW.

Conclusions:

• The Bayesian predictive approach offers a robust method for estimating and comparing DTRs.
• Null-robust reparametrization is essential for reliable causal inference in longitudinal studies.
• Integrating IPTW within a Bayesian framework enhances the validity of causal effect estimation.