Collaborative Research: New Bayesian Causal Methods for Personalized Decision-Making Under Unmeasured Confounding

Data-driven personalized decision-making has become increasingly important across many fields, such as health sciences where tailoring treatments to individual patients can improve effectiveness and reduce adverse effects. Achieving reliable personalized decisions requires understanding cause-and-effect relationships between actions and outcomes. However, most real-world data sources, such as electronic medical records, health surveys, and social media data, are observational rather than randomized, making causal relationships difficult to establish. In these settings, hidden or unmeasured factors may influence both the actions individuals take and the outcomes they experience, leading to biased conclusions and unreliable recommendations if not properly addressed. This project will address this fundamental challenge by developing new statistical methods for learning optimal personalized decision rules from observational data when important confounding factors are not fully observed. The project will consider both single-stage and sequential decisions, with particular attention to continuous treatments such as medication dosages. A motivating application is kidney transplantation, where optimizing immunosuppressive therapy over time is essential to reduce the risk of graft failure while minimizing harmful side effects. By enabling more reliable individualized decision-making, this project will advance statistical science, machine learning, and artificial intelligence, support the training of students in modern data science, and contribute to improved health outcomes and broader societal well-being. This project aims to develop novel Bayesian causal methods for estimating treatment effects and optimizing individualized decision rules from observational data with unmeasured confounding. A Bayesian joint modeling framework will be introduced for treatment, outcome, observed covariates, and latent confounders, leveraging mild distributional assumptions to enable causal identification without relying on additional data sources required by many existing approaches, such as instrumental or proxy variables. The project will also develop a dynamic Bayesian causal modeling framework for longitudinal data, where treatment decisions and unmeasured confounders evolve over time. This framework will support the estimation of adaptive treatment regimes that respond to an individual’s evolving treatment history, outcomes, and characteristics. In addition, the project will design optimization methods for both single-stage and sequential decision-making, using posterior uncertainty to improve robustness in finite and unbalanced observational data settings. The methods will be evaluated through simulation studies and applied to large-scale real-world kidney transplantation data for studying optimal personalized and dynamic immunosuppressive dosing strategies. To facilitate broad dissemination, open-source software will be developed for implementation. The resulting framework and tools will provide a general approach to reliable personalized decision-making in biomedicine and other fields that rely on complex observational data. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2610269 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Wei Jin | Institution: Trustees of Boston University, BOSTON, MA | Award Amount: $73,241 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2610269 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2610269.html