📝 SummarXiv

Abstract

This paper focuses on the Bayesian Network Propensity Score (BNPS), a novel approach for estimating treatment effects in observational studies characterized by unknown (and likely unbalanced) designs and complex dependency structures among covariates. Traditional methods, such as logistic regression, often impose rigid parametric assumptions that may lead to misspecification errors, compromising causal inference. Recent classical and machine learning alternatives, such as boosted CART, random forests, and Stable Balancing Weights, seem to be attractive in a predictive perspective, but they typically lack asymptotic properties, such as consistency, efficiency, and valid variance estimation. In contrast, the recently proposed BNPS to estimate propensity scores uses Bayesian Networks to flexibly model conditional dependencies while preserving essential statistical properties such as consistency, asymptotic normality and asymptotic efficiency. Combined with the H\'ajek estimator, BNPS enables robust estimation of the Average Treatment Effect (ATE) in scenarios with strong covariate interactions and unknown data-generating mechanisms. Through extensive simulations across fifteen realistic scenarios and varying sample sizes, BNPS consistently outperforms benchmark methods in both empirical rejection rates and coverage accuracy. Finally, an application to a real-world dataset of 7,162 prostate cancer patients from San Raffaele Hospital (Milan, Italy) demonstrates BNPS's practical value in assessing the impact of pelvic lymph node dissection on hospitalization duration and biochemical recurrence. The findings support BNPS as a statistically robust, interpretable and transparent alternative for causal inference in complex observational settings, enhancing the reliability of evidence from real-world biomedical data.