Machine-learning classification of high parity among women in India: A survey-aware NFHS-5 analysis with external transport assessment in Nepal

Existing machine-learning studies of fertility often ignore complex survey clustering, permit preprocessing leakage, or describe cross-sectional parity as prospective risk. We developed a survey-aware classifier for the observed state of having three or more children ever born among women aged 15–49 in India and evaluated locked-model transport to Nepal. The analysis used 724,115 NFHS-5 India respondents. The binary outcome was children ever born ≥3. Twenty-six corrected demographic and reproductive predictors were processed inside scikit-learn pipelines using training-fitted numeric imputation with missingness indicators, categorical imputation, one-hot encoding, and scaling. All preprocessing and model-development pipelines were implemented using the scikit-learn framework [Pedregosa et al.: Scikit-learn: machine learning in python. Journal of Machine Learning Research. 2011;12:2825–2830] An 80:20 split was disjoint by primary sampling unit (PSU). Six classifiers were compared using five-fold stratified -validation; final models were fitted to all 579,400 training observations and evaluated once on 144,715 held-out observations. Uncertainty was quantified by 300 PSU-within-stratum bootstrap replicates. The threshold was fixed from training out-of-fold predictions. Nested 5×3 grouped cross-validation, learning curves, feature-stability resampling, survey-weight sensitivity, and preprocessing/feature-selection/imbalance analyses were prespecified. A common-feature India model was applied to all 14,845 Nepal DHS 2022 women without retraining or Nepal-based preprocessing. LightGBM was selected from training-only cross-validation (mean AUC 0.9385, SD 0.0019). On the untouched India holdout, AUC was 0.9433 (95% cluster-bootstrap CI 0.9424–0.9443), accuracy 0.8470, F1 0.7717, sensitivity 0.8938, specificity 0.8280, negative predictive value 0.9504, balanced accuracy 0.8609, Matthews correlation coefficient 0.6739, and Brier score 0.1005. The locked threshold was 0.4890. Fully nested grouped validation produced AUC 0.9414 (SD 0.0007). The calibration slope was 1.0149, but the intercept was −0.8931 and ECE was 0.0786, indicating systematic probability overestimation under class weighting. Nepal transport AUC was 0.9357; the India-minus-Nepal AUC difference was 0.00494 (95% CI 0.00083–0.00887). The revised pipeline provides strong discrimination and stable grouped-validation performance for cross-sectional high parity. It does not establish causal determinants or prospective fertility risk. The Nepal result supports limited transportability of common constructs, not equivalence or immediate deployment. Local recalibration and prospective evaluation are required before predicted probabilities are used operationally.

Authors

Institutions

Publication Details

Journal
Statistics
Published
2026-11-12
DOI
https://doi.org/10.1080/02331888.2026.2762186
Primary Topic
Reproductive Health and Contraception
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Machine-learning classification of high parity among women in India: A survey-aware NFHS-5 analysis with external transport assessment in Nepal

Poonam Singh, Ankur Kumar Jaiswal
Statistics
Reproductive Health and Contraception
article

Machine-learning classification of high parity among women in India: A survey-aware NFHS-5 analysis with external transport assessment in Nepal

Poonam Singh, Ankur Kumar Jaiswal
article en

Abstract

Existing machine-learning studies of fertility often ignore complex survey clustering, permit preprocessing leakage, or describe cross-sectional parity as prospective risk. We developed a survey-aware classifier for the observed state of having three or more children ever born among women aged 15–49 in India and evaluated locked-model transport to Nepal. The analysis used 724,115 NFHS-5 India respondents. The binary outcome was children ever born ≥3. Twenty-six corrected demographic and reproductive predictors were processed inside scikit-learn pipelines using training-fitted numeric imputation with missingness indicators, categorical imputation, one-hot encoding, and scaling. All preprocessing and model-development pipelines were implemented using the scikit-learn framework [Pedregosa et al.: Scikit-learn: machine learning in python. Journal of Machine Learning Research. 2011;12:2825–2830] An 80:20 split was disjoint by primary sampling unit (PSU). Six classifiers were compared using five-fold stratified -validation; final models were fitted to all 579,400 training observations and evaluated once on 144,715 held-out observations. Uncertainty was quantified by 300 PSU-within-stratum bootstrap replicates. The threshold was fixed from training out-of-fold predictions. Nested 5×3 grouped cross-validation, learning curves, feature-stability resampling, survey-weight sensitivity, and preprocessing/feature-selection/imbalance analyses were prespecified. A common-feature India model was applied to all 14,845 Nepal DHS 2022 women without retraining or Nepal-based preprocessing. LightGBM was selected from training-only cross-validation (mean AUC 0.9385, SD 0.0019). On the untouched India holdout, AUC was 0.9433 (95% cluster-bootstrap CI 0.9424–0.9443), accuracy 0.8470, F1 0.7717, sensitivity 0.8938, specificity 0.8280, negative predictive value 0.9504, balanced accuracy 0.8609, Matthews correlation coefficient 0.6739, and Brier score 0.1005. The locked threshold was 0.4890. Fully nested grouped validation produced AUC 0.9414 (SD 0.0007). The calibration slope was 1.0149, but the intercept was −0.8931 and ECE was 0.0786, indicating systematic probability overestimation under class weighting. Nepal transport AUC was 0.9357; the India-minus-Nepal AUC difference was 0.00494 (95% CI 0.00083–0.00887). The revised pipeline provides strong discrimination and stable grouped-validation performance for cross-sectional high parity. It does not establish causal determinants or prospective fertility risk. The Nepal result supports limited transportability of common constructs, not equivalence or immediate deployment. Local recalibration and prospective evaluation are required before predicted probabilities are used operationally.

Statistics
Central University of Haryana (IN), Banaras Hindu University (IN)
Gender equality, Reduced inequalities
Reproductive Health and Contraception
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon