Bayesian Propensity Score Predictive Inference for Extending Trial Findings to Target Populations
Abstract
The CODIACS Vanguard randomized trial compared a depression care intervention with usual care for patients with persistent depressive symptoms following acute coronary syndrome. We assess whether its findings can be extended to broader patient populations represented by two cohorts constructed from the All of Us electronic health records. This extension is challenging when treatment effects vary across covariates whose distributions differ between trial and target populations. We develop Bayesian Propensity Score Predictive Inference (PSPI), an interaction-based outcome modeling framework that combines Bayesian additive regression trees for potentially high-dimensional covariates with natural cubic splines of the sampling propensity score. Two variants include a separate spline in the treatment effect component alone or in both the prognostic and treatment effect components. The spline components allow separate regularization of selection-related variation and thus improve potential outcome estimation within the observed range and allow an explicit extrapolation rule beyond their boundary knots under substantial covariate shift. An inverse Gumbel transformation of the sampling propensity score balances compression near zero and the extent of extrapolation. Simulation studies show that spline-based PSPI models outperform competing methods in the settings studied, achieving lower root mean squared errors and near-nominal coverage, particularly under substantial covariate shift. In the application, PSPI estimates larger intervention benefits in both target populations than the unadjusted trial estimate.
Related
- SCOPE Before You Sample: Conformal Scores Guide Posterior Exploration for Fast and Calibrated Bayesian Predictive Inference
- Concentration and Calibration in Predictive Bayesian Inference
- Predictive-Context LOO: Bayesian Model Comparison with Pretrained Predictive Learners
- Efficient Bayesian inference for multiple network data
- Bayesian inference under model misspecification
BibTeX
@misc{zou2026bayesian,
title = {{Bayesian Propensity Score Predictive Inference for Extending Trial Findings to Target Populations}},
author = {Jungang Zou and Joseph E. Schwartz and Roderick J. A. Little and Nathalie Moise and Qixuan Chen},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.31915},
eprint = {2610.31915},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.31915},
}
RIS
TY - UNPB
TI - Bayesian Propensity Score Predictive Inference for Extending Trial Findings to Target Populations
AU - Jungang Zou
AU - Joseph E. Schwartz
AU - Roderick J. A. Little
AU - Nathalie Moise
AU - Qixuan Chen
PY - 2026
DA - 2026/10/22
PB - arXiv
DO - 10.48550/arxiv.2610.31915
UR - https://arxiv.org/abs/2610.31915
AB - The CODIACS Vanguard randomized trial compared a depression care intervention with usual care for patients with persistent depressive symptoms following acute coronary syndrome. We assess whether its findings can be extended to broader patient populations represented by two cohorts constructed from the All of Us electronic health records. This extension is challenging when treatment effects vary across covariates whose distributions differ between trial and target populations. We develop Bayesian Propensity Score Predictive Inference (PSPI), an interaction-based outcome modeling framework that combines Bayesian additive regression trees for potentially high-dimensional covariates with natural cubic splines of the sampling propensity score. Two variants include a separate spline in the treatment effect component alone or in both the prognostic and treatment effect components. The spline components allow separate regularization of selection-related variation and thus improve potential outcome estimation within the observed range and allow an explicit extrapolation rule beyond their boundary knots under substantial covariate shift. An inverse Gumbel transformation of the sampling propensity score balances compression near zero and the extent of extrapolation. Simulation studies show that spline-based PSPI models outperform competing methods in the settings studied, achieving lower root mean squared errors and near-nominal coverage, particularly under substantial covariate shift. In the application, PSPI estimates larger intervention benefits in both target populations than the unadjusted trial estimate.
ER -