An integrated analytical framework for gender-based violence research: A simulation study combining machine learning and causal inference

BackgroundCurrent research on Gender-Based Violence (GBV) typically separates predictive machine learning and causal inference into distinct analytical silos. Yet, grasping the multi-level determinants of violence requires an approach that can both identify high-value predictors and disentangle their causal mechanisms. MethodsThis study develops and demonstrates an integrated five-phase analytical framework for GBV research, applying sequential methods to a single synthetic dataset (N = 3,000) parameterized to reflect the prevalence patterns and risk factor distributions of the 2022 Kenya Demographic and Health Survey (KDHS). The framework was applied across five stages: descriptive epidemiology, Random Forest variable selection, logistic regression for adjusted associations, mediation analysis, and evidence synthesis. ResultsThe Random Forest model achieved 74.6% accuracy (AUC = 0.711; sensitivity = 37.3%; specificity = 89.8%), recovering partner alcohol use, childhood trauma, and marital conflict as the top-ranked predictors. Importantly, this accuracy offers only a modest gain over the null classifier baseline of 73.1%. Multivariable logistic regression yielded stable adjusted effect estimates for partner alcohol use (aOR = 6.60) and childhood trauma (aOR = 1.99). Mediation analysis tested three theoretically informed indirect pathways; however, none of the hypothesized indirect effects reached statistical significance (ACME p > 0.05 for all pathways), indicating that the specified risk factors operated primarily through direct rather than mediated routes in this simulation. ConclusionsThis proof-of-concept demonstration shows that the five-phase framework can be coherently applied to synthetic GBV data parameterized from KDHS estimates. The frameworks multi-level logistic regression approach produced meaningfully better model fit than single-level alternatives ({Delta}AIC = 67.5). Future research should apply this framework to real-world longitudinal data to rigorously evaluate its performance characteristics and causal inference capacity.