ADVANCING PUBLIC HEALTH WITH CAUSAL MACHINE LEARNING: ESTIMATING THE CAUSAL EFFECT OF SMOKING ON DIABETES RISK USING BRFSS 2015 DATA

Abstract

Distinguishing correlation from causation is critical for effective 
public health interventions. This study applies a rigorous causal machine learning framework to estimate the causal effect of smoking on diabetes risk using the 2015 Behavioral Risk Factor Surveillance System (BRFSS) dataset (N = 253,680 U.S. adults). We constructed a directed acyclic graph (DAG), identified the average treatment effect (ATE) via the backdoor criterion, and estimated it using propensity score matching (PSM) within the DoWhy library. Smoking was associated with a statistically significant 1.6 percentage point increase in diabetes probability (ATE = 0.016, 95% CI: 0.010–0.022). Robustness was confirmed through multiple 
refutation tests, including subset refutation (new ATE = 0.0086, p = 0.52) and random common cause addition (new ATE = 0.0071, p = 0.20). These findings provide policy-relevant causal evidence supporting intensified anti-smoking interventions as part of diabetes prevention strategies and demonstrate the value of integrating causal inference with machine learning in observational public health research.