Robust Treatment-Effect Learning Beyond Binary Treatments: Stabilized Weighting with Deep Representation

Modern scientific, medical, and policy decisions increasingly rely on observational data to evaluate the effects of interventions such as medical treatments, public programs, and behavioral exposures. Many of these interventions are multi-level or continuous, such as medication dosage or program participation intensity, rather than simple binary choices. However, existing statistical methods, including widely used propensity score approaches, often become unstable or unreliable in these settings, particularly when applied to high-dimensional data with complex relationships. This project addresses these challenges by developing new methods for reliable and interpretable causal analysis. The research will enhance the ability to draw credible conclusions from large-scale data, with applications in public health, healthcare delivery, social policy programs, and digital health. By improving the stability and accuracy of treatment-effect estimation, the project supports evidence-based decision making that can improve health outcomes, inform policy design, and promote societal well-being. In addition, the project will contribute to workforce development through student training, interdisciplinary collaboration, and outreach activities that broaden participation in data science and statistics. This research advances the theory and methodology of causal inference for complex data with multi-level and continuous treatments, where classical inverse propensity weighting methods can fail due to unstable weights and ill-conditioned estimation. It develops a unified and scalable framework for treatment-effect estimation beyond binary interventions by leveraging stabilized weighting, empirical likelihood, and deep learning representations. The project introduces new methodologies for counterfactual distribution estimation, dose–response analysis, and general loss-based causal inference. The proposed approach replaces unstable inverse weighting with stabilized weighting strategies, including empirical likelihood weighting and minimum-variance weighting, which control variance inflation and prevent extreme weights. It further incorporates modern deep architectures, such as Transformer-based models, to flexibly estimate nuisance functions in high-dimensional settings. Theoretical contributions include non-asymptotic generalization bounds, convergence rates, and asymptotic inference guarantees, establishing a rigorous foundation for combining deep learning with semiparametric causal inference. The project will also develop easy-to-implement software packages to facilitate practical implementation and broad dissemination of the proposed methods. The resulting methods will be validated on large-scale observational datasets, providing robust, interpretable, and scalable tools for causal learning in complex real-world environments. 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: 2610432 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Shujie Ma | Institution: University of California-Riverside, RIVERSIDE, CA | Award Amount: $199,065 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2610432 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2610432.html