Causal Inference for Health Policy Support Systems: A Directed Acyclic Graphs Based Approach

Abstract

The accurate estimation of causal effects is fundamental for evidence-based health policy decisions. Traditional correlational approaches often produce biased estimates due to uncontrolled confounding, leading to potentially misleading policy recommendations. This study proposes and evaluates a causal inference framework based on Directed Acyclic Graphs (DAGs) for assessing health intervention effects in observational settings. Using simulated data from a nutrition intervention program (n=1,500), we systematically compared DAG-based causal estimation methods against conventional correlational approaches. Our simulations incorporated realistic confounding structures including socioeconomic status, parental education, and baseline health metrics. Results demonstrate that naive correlational analysis overestimated the intervention effect by 25.6% (10.05 vs. 8.00 points), while DAG-based backdoor adjustment methods yielded estimates within 5.4% of the true effect (8.43, 95% CI: 7.88-8.95). Inverse probability weighting produced intermediate results (9.83 points). Sensitivity analyses revealed that causal estimates remained robust across varying degrees of unmeasured confounding, whereas correlational estimates deteriorated rapidly. These findings suggest that DAG-based frameworks provide more reliable effect estimates for health policy evaluation, particularly when randomized trials are infeasible. The methodology presented offers practical guidance for public health researchers and policymakers seeking to make evidence-informed decisions from observational data. Implementation of these causal inference tools could substantially improve the validity of policy impact assessments in resource-limited settings where experimental designs are often impractical.