Profile-Based Auditing of Domestic Violence Survey Selection Across South Asia
Authors: Mohammad Abu Obaida Mullick, Rezaur Rahman Ratul, Mymuna Nourin, Md. Motaharul Islam
Organizations: Institute of Advanced Research, United International University, Dhaka, Bangladesh · Dept of Computer Science & Engineering, United International University, Dhaka, Bangladesh
Domestic violence (DV) data in the Demographic and Health Surveys (DHS) are collected through a specialized module whose household-level sampling and privacy requirements create a distinct observation process. We ask whether demographic, socioeconomic, and household characteristics form reproducible population profiles and whether these profiles contain information about selection into the DV module across South Asia. Using DHS Individual Recode data for 796,374 women aged 15-49 across Afghanistan, India, Myanmar, Nepal, and Pakistan, we harmonize 61 numeric and 8 categorical features, reduce the profile space with PCA, and identify six interpretable profiles. We then evaluate 22 machine-learning models under primary-sampling-unit-aware holdout validation. The best model, HistGradientBoosting, achieves ROC-AUC 0.9167 and PR-AUC 0.9612, versus 0.8475 and 0.9203 for a design-only baseline. Adding woman-level characteristics yields the largest incremental gain, while explicit interaction terms add little out-of-sample performance. Country-specific analyses reveal substantial heterogeneity, and leave-one-country-out evaluation shows that transferability of the incremental profile signal is limited. These results demonstrate that interpretable ML can audit differential coverage in sensitive survey modules beyond the formal sampling design.
Figures & tables
Country
Women
Selected
Selected (%)
Interviewed
Interviewed/Selected (%)
Afghanistan
29,461
22,132
75.12
21,324
96.35
India
724,115
494,830
68.34
72,320
14.62
Myanmar
12,885
4,563
35.41
4,530
99.28
Nepal
14,845
5,192
34.97
5,178
99.73
Pakistan
15,068
4,216
27.98
4,085
96.89
Total
796,374
530,933
66.67
107,437
20.24
Table 1: Study population and realized DV-module selection in the five-country modeling universe.
Feature representation
HistGradientBoosting
Logistic
LightGBM
Design
0.8287
0.8264
0.8285
Design + country
0.8516
0.8479
0.8511
woman characteristics
0.9161
0.9013
0.9156
household characteristics
0.9166
0.9020
0.9144
agency
0.9167
0.9021
0.9139
norms
0.9167
0.9021
0.9139
Table 2: Incremental predictive value of feature blocks for DV-module selection.
Appendix figures & tables40 assets
Supplementary material from the paper’s appendix.
Appendix
Country
Survey
Recode
Primary unit
Analysis role
Afghanistan
2015
HR/PR/IR
Household / woman
Included
Bangladesh
2022
HR/IR
Household / woman
IR audit; excluded from V044 models
India
2019–21
PR/IR
Member / woman
Included
Myanmar
2015–16
PR/IR
Member / woman
Included
Nepal
2022
PR/IR
Member / woman
Included
Pakistan
2017–18
PR/IR
Member / woman
Included
Appendix
Table 3: DHS recode inventory and role in the analysis.
Figure 1: Weighted mean household size and residence composition across countries. The left panel reports weighted mean household size, while the right panel summarizes the weighted urban–rural residence composition using the household sampling weight.
Figure 2: Weighted age–sex composition of listed household members in the supplied PR files. The figure describes the household-member population represented by the available roster files and should not be interpreted as the distribution of completed women’s interviews.
Figure 3: Country-wise female study-frame size and availability of women aged 15–49 indicators. The figure documents the size of the available female study frame and the corresponding coverage of key indicators across the supplied surveys.
Figure 4: Women aged 15–49 by country and demographic or socioeconomic dimension. Country-specific weighted distributions are shown for the dimensions used in the descriptive audit and harmonized analysis.
Block
Principal variables and derived measures
Design
Eligible women, reciprocal eligible-women count, single-eligible-woman indicator, household size, children under five, crowding, capped eligibility categories.
Country
Country and survey context, including survey year representation.
Woman
Age, education, marital and union status, age at cohabitation, fertility, parity, pregnancy, child-marriage indicators, partner age and education, and related gaps.
Household
Urban residence, wealth, assets, water, sanitation, fuel, dwelling quality, household-head characteristics, and relationship to household head.
Agency
Employment, occupation, earnings control, household decision-making, asset ownership, healthcare barriers, insurance, mobile phone, financial accounts, internet, and media exposure.
Norms
Indicators and derived indices representing attitudes toward wife beating.
Appendix
Table 4: Feature blocks used in the pooled selection models.
Figure 5: Weighted education and marital-status composition among women aged 15–49. The panels document the cross-country distributions of two principal harmonized demographic dimensions.
Figure 6: Single-year age distribution of women aged 15–49 by country. Weighted age profiles provide a direct check on the comparability of the principal female study population across surveys.
V044
Meaning
0
Not selected for the DV module
1
Selected and interviewed
2
Selected but privacy was not possible
3
Selected but not interviewed for another reason
Appendix
Table 5: Coding of the DHS V044 domestic-violence module status variable used in the analysis.
Figure 7: DV-module selection rates across demographic and socioeconomic characteristics. The figure provides the descriptive selection landscape before multivariate profile learning and predictive modeling.
Figure 8: Outcomes among women selected for the DV module and conditional completion patterns. The panels show interview completion, privacy failure, and other non-interview outcomes, together with completion patterns across household crowding quintiles and the relevant household-role characteristic.
Figure 9: Domestic-violence item response coverage and question-level ever-experience distributions by country. The figure separates item availability and response completeness from the primary DV-module selection analysis.
Figure 10: Distribution of V044 and comparison of observed selection with the design expectation. The left panel reports the unweighted composition of V044 states by country; the right compares observed selection with 100E[1/V138] .
Country
Women
Selected
Selected (%)
Interviewed among selected (%)
Afghanistan
29,461
21,323–21,325
75.12%
96.35%
India
724,115
494,824–494,896
68.34%
14.6%
Myanmar
12,885
4,562–4,564
35.41%
99.3%
Nepal
14,845
5,191–5,192
34.97%
99.7%
Pakistan
15,068
4,216–4,217
27.98%
96.9%
Total
796,374
530,933
66.67%
–
Appendix
Table 6: Principal selection and completion outcomes in the five-country modeling universe.
Target
Model
ROC-AUC
PR-AUC
Completed, all women
Logistic regression
0.7350
0.4340
Completed, all women
Random forest
0.7377
0.4367
Completed, all women
LightGBM
0.7332
0.4317
Interviewed ∣ selected
Logistic regression
0.6607
0.5307
Interviewed ∣ selected
Random forest
0.6600
0.5272
Interviewed ∣ selected
LightGBM
0.6638
0.5330
Appendix
Table 7: Prediction of interview completion under unconditional and selected-only targets.
Figure 11: PCA scree plot and cumulative explained variance. Nineteen principal components are retained because they explain at least 80% of the variance in the standardized profile space.
Figure 12: Cluster-number diagnostics for the PCA-reduced profile representation. K-means inertia, silhouette score, and Gaussian-mixture BIC are shown across candidate values of K .
Profile
Interpretation
Key cross-country pattern
P0
Middle rural, no/low education, high parity
Most prevalent in Nepal; also substantial in Pakistan and India.
P1
Richer urban, higher education
Common in India, Nepal, and Pakistan; uncommon in Afghanistan.
P2
Middle rural, secondary education
Broadly distributed and particularly prevalent in India and Nepal.
P3
Poorer rural, secondary education, never-married
Especially represented in Myanmar and India; nearly absent in Afghanistan and Pakistan.
P4
Poorer rural, no/low education, high parity
Dominant in Afghanistan and strongly represented in Pakistan and Myanmar.
P5
Richer urban, higher education, never-married
More represented in Myanmar, Nepal, and India; nearly absent in Afghanistan and Pakistan.
Appendix
Table 8: Interpretation of the six reported demographic–socioeconomic profiles.
Figure 13: Weighted composition of the six demographic–socioeconomic profiles within each country. The left panel reports the country-by-profile composition and the right panel presents the same values as a heatmap.
Figure 14: Women in the first two principal components and country centroids in the profile space. The figure visualizes the low-dimensional representation used for profile construction and overlays the centroid of each modeling country.
Figure 15: Weighted distributions of continuous Individual Recode measures by country. Density curves provide additional descriptive context for the demographic and socioeconomic heterogeneity underlying the profile solution.
Figure 16: Profile-level DV-module selection and cross-country heterogeneity. The panels show pooled selection rates by profile, deviations from the design expectation in percentage points, and the profile-by-country selection-rate matrix.
Country
P0
P1
P2
P3
P4
P5
Afghanistan
11.7%
4.3%
5.4%
0.1%
78.1%
0.3%
India
21.0%
20.9%
22.5%
12.5%
11.4%
11.8%
Myanmar
9.9%
12.9%
17.9%
19.5%
25.4%
14.3%
Nepal
28.7%
19.0%
21.6%
10.1%
8.5%
12.2%
Pakistan
22.6%
21.5%
21.5%
0.3%
33.7%
0.4%
Appendix
Table 9: Weighted profile composition by country (%).
Figure 17: Weighted Individual Recode profile of women aged 15–49 across the South Asian modeling countries. The figure summarizes the principal demographic and socioeconomic characteristics entering the profile analysis.
Figure 18: Agreement with reason-specific indicators and number of reasons endorsed by country. The figure summarizes country-level variation in the normative indicators included in the broader feature representation.
Model
ROC-AUC
PR-AUC
F1
Majority class
0.5000
0.6670
0.800
Stratified random
0.5016
0.6677
0.668
Design-only baseline
0.8475
0.9203
0.784
Logistic regression
0.9022
0.9547
0.840
Linear SVM
0.9024
0.9547
0.835
Linear discriminant analysis
0.9031
0.9550
0.860
Appendix
Table 10: Complete pooled learner comparison for DV-module selection.
Specification
Added block
ROC-AUC
PR-AUC
Δ ROC-AUC
Δ PR-AUC
Interpretation
Design-only
Design
0.8287
0.8986
–
–
Sampling mechanism
Design + country
Country
0.8516
0.9250
+0.0229
+0.0264
Country variation
Design + demographics
Woman
0.9161
0.9609
+0.0874
+0.0623
Large additional signal
* household
Household
0.9166
0.9611
+0.0879
+0.0625
Marginal additional signal
agency + norms
Agency + norms
0.9167
0.9612
+0.0880
+0.0626
Performance plateau
interactions
Explicit interactions
0.9165
0.9611
+0.0878
+0.0625
Negligible incremental gain
Appendix
Table 11: Incremental predictive evidence from demographic, socioeconomic, agency, normative, and interaction feature blocks. Models are evaluated on the same held-out PSUs; Δ values are relative to the design-only baseline.
Figure 19: Marginal ROC-AUC contribution of successive feature blocks. The figure shows the incremental discrimination associated with each additional feature block in the nested model ladder.
Figure 20: Distribution of predicted selection probabilities across evaluated models. The figure provides an additional view of score separation beyond aggregate AUC statistics.
Figure 21: Confusion matrices for selected high-performing DV-module selection models on held-out PSUs.
Figure 22: ROC and precision–recall curves on held-out PSUs. Curves complement the learner-level AUC comparison by showing the ranking behavior of selected models across classification thresholds.
Figure 23: ROC-AUC and precision–recall AUC across the complete learner set. HistGradientBoosting provides the strongest pooled discrimination in the study.
Interaction
OR
95% CI lower
95% CI upper
Age × union status
0.241
0.226
0.258
Eligible women × education
1.244
1.188
1.303
Eligible women × wealth
0.806
0.757
0.859
Wealth × education
1.235
1.170
1.304
Urban × wealth
0.900
0.850
0.954
Eligible women × urban
1.084
1.041
1.128
Appendix
Table 12: Largest estimated interaction associations with DV-module selection. Odds ratios are expressed per standard deviation of the corresponding product term.
Figure 24: Feature importance and allocation of importance across feature blocks. The left panel ranks the top individual features; the right panel aggregates importance by model block.
Figure 25: Two-way partial dependence for the predicted probability of DV-module selection. Joint surfaces illustrate nonlinear relationships between selected feature pairs and the modelled probability of selection.
Figure 26: One-way partial dependence for the predicted probability of DV-module selection. Curves summarize the modelled marginal relationship between individual feature dimensions and predicted selection probability.
Country
Profile model ROC-AUC
Design ROC-AUC
Gain
India
0.9196
0.8486
+0.0710
Afghanistan
0.8713
0.8746
-0.0033
Pakistan
0.6102
0.6268
-0.0166
Nepal
0.6331
0.6642
-0.0311
Myanmar
0.6427
0.6771
-0.0344
Appendix
Table 13: Within-country profile-model performance relative to design-only baselines.
Figure 27: Within-country ROC-AUC, profile gain over the design baseline, and the relationship between profile gain and sample size.
Held-out country
Transfer ROC-AUC
Afghanistan
0.8719
India
0.8374
Myanmar
0.6668
Nepal
0.6484
Pakistan
0.6219
Mean transfer penalty
-0.0061
Appendix
Table 14: Leave-one-country-out transfer performance.
Figure 28: Within-country versus leave-one-country-out transfer performance. The panels report absolute transfer performance, transfer penalties, and whether the transferred profile model exceeds the target-country design baseline.
Selection bias studies how nonrandom sampling or observation distorts statistical inference, model training, and evaluation relative to a target population.