Background Conventional logistic regression estimates a single population-average odds ratio (OR), masking heterogeneity in how social determinants affect health outcomes across population subgroups. We applied machine learning (ML)-based stratified g-computation to estimate individual-specific wealth-associated counterfactual ORs of household wealth on modern contraceptive use among currently married Bangladeshi women in the two extreme wealth quintiles. Methods We analyzed 11,510 currently married women in the poorest and richest wealth quintiles from the Bangladesh Demographic and Health Survey (BDHS) 2022. Six model classes (Logistic Regression, Random Forest, Gradient Boosting, XGBoost, LightGBM, Stacking Ensemble) were compared with five-fold cross-validated area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) analysis, primary sampling unit (PSU)-level bootstrap confidence intervals (CIs), doubly-robust augmented inverse probability weighted (AIPW) sensitivity analysis, Benjamini-Hochberg false discovery rate (BH-FDR) correction, and E-value analysis were conducted. Findings apply to the two extreme wealth quintiles and should not be extrapolated to the middle three quintiles. Results Modern contraceptive prevalence was 37.7% in the poorest quintile versus 32.6% in the richest. The Stacking Ensemble achieved the highest CV-AUC (0.7108). All six AIPW estimates were near-null (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(p = 0.09\)</EquationSource></InlineEquation>–0.83), revealing that the aggregate wealth effect is zero because opposing subgroup effects cancel: a <InlineEquation ID="IEq2"><EquationSource Format="TEX">\(+17.6\)</EquationSource></InlineEquation> percentage-point absolute risk difference (ARD) in the 35–49/higher-education subgroup is precisely offset by a <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(-11.3\)</EquationSource></InlineEquation> percentage-point ARD in the 15–24/primary-education subgroup. Individual ORs exhibited a Gini coefficient of 0.4525 (95% bootstrap CI: 0.4475–0.4568); interaction tests confirmed heterogeneity by education (<InlineEquation ID="IEq4"><EquationSource Format="TEX">\(p = 0.039\)</EquationSource></InlineEquation>) and age (<InlineEquation ID="IEq5"><EquationSource Format="TEX">\(p = 0.005\)</EquationSource></InlineEquation>). Median ORs ranged 6.4-fold: from 0.552 (15–24/primary; FDR <InlineEquation ID="IEq6"><EquationSource Format="TEX">\(q = 0.008\)</EquationSource></InlineEquation>) to 3.557 (35–49/higher; E-value = 6.57). Parity was the strongest SHAP predictive contributor, followed by employment and age. Conclusions Wealth-associated counterfactual OR differences in contraceptive use are patterned within demographic subgroups; population-average estimates are zero because opposing effects cancel, and only individual-level ML-based g-computation reveals this structure. The programmatic targeting framework derived from this approach offers a reproducible prioritisation tool for health equity analyses in any DHS-based survey.

Heterogeneous wealth-related differences in modern contraceptive use in Bangladesh: a machine learning-based g-computation analysis of the DHS 2022 cross-sectional survey

期刊论文
出版模式：

开放获取
发表日期：

2026年7月8日

查看全文

Sudarshan Saha,
Sohel Rana

摘要

Background

Conventional logistic regression estimates a single population-average odds ratio (OR), masking heterogeneity in how social determinants affect health outcomes across population subgroups. We applied machine learning (ML)-based stratified g-computation to estimate individual-specific wealth-associated counterfactual ORs of household wealth on modern contraceptive use among currently married Bangladeshi women in the two extreme wealth quintiles.

Methods

We analyzed 11,510 currently married women in the poorest and richest wealth quintiles from the Bangladesh Demographic and Health Survey (BDHS) 2022. Six model classes (Logistic Regression, Random Forest, Gradient Boosting, XGBoost, LightGBM, Stacking Ensemble) were compared with five-fold cross-validated area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) analysis, primary sampling unit (PSU)-level bootstrap confidence intervals (CIs), doubly-robust augmented inverse probability weighted (AIPW) sensitivity analysis, Benjamini-Hochberg false discovery rate (BH-FDR) correction, and E-value analysis were conducted. Findings apply to the two extreme wealth quintiles and should not be extrapolated to the middle three quintiles.

Results

Modern contraceptive prevalence was 37.7% in the poorest quintile versus 32.6% in the richest. The Stacking Ensemble achieved the highest CV-AUC (0.7108). All six AIPW estimates were near-null (\(p = 0.09\)–0.83), revealing that the aggregate wealth effect is zero because opposing subgroup effects cancel: a \(+17.6\) percentage-point absolute risk difference (ARD) in the 35–49/higher-education subgroup is precisely offset by a \(-11.3\) percentage-point ARD in the 15–24/primary-education subgroup. Individual ORs exhibited a Gini coefficient of 0.4525 (95% bootstrap CI: 0.4475–0.4568); interaction tests confirmed heterogeneity by education (\(p = 0.039\)) and age (\(p = 0.005\)). Median ORs ranged 6.4-fold: from 0.552 (15–24/primary; FDR \(q = 0.008\)) to 3.557 (35–49/higher; E-value = 6.57). Parity was the strongest SHAP predictive contributor, followed by employment and age.

Conclusions

Wealth-associated counterfactual OR differences in contraceptive use are patterned within demographic subgroups; population-average estimates are zero because opposing effects cancel, and only individual-level ML-based g-computation reveals this structure. The programmatic targeting framework derived from this approach offers a reproducible prioritisation tool for health equity analyses in any DHS-based survey.