Other

Spatial multilevel and decomposition analysis of minimum dietary diversity among Indian children aged 6 to 23 months in NFHS-5 with an evaluation of the Aspirational Districts Programme

Bhuyan H, Barman B, VK M, Talukdar K, Ojah M et al. · 2 Jul 2026

Abstract

Abstract Despite a decade of national nutrition programmes, the proportion of Indian children aged 6–23 months meeting the WHO/UNICEF (2021) Minimum Dietary Diversity (MDD) criterion remains low, and existing district-level evidence is fragmented across studies using different definitions, weighting choices, and analytical approaches. This study integrates spatial, multilevel, decomposition, machine-learning, and quasi-experimental methods in a single framework to identify where dietary-diversity failure clusters, why neighbouring districts diverge, and whether the Aspirational Districts Programme (ADP) has altered MDD between NFHS-4 and NFHS-5. Children aged 6–23 months (n = 64,123) from the National Family Health Survey-5 (NFHS-5; 2019–21) were analysed using sampling weights (v005), strata (v022), and PSU (v021). MDD was defined as consumption of ≥ 5 of 8 food groups in the previous 24 hours (WHO/UNICEF 2021), with the WHO 2008 indicator (≥ 4 of 7) reported as a sensitivity analysis. A three-level random-intercept logistic regression (child–district–state) yielded variance partition coefficients (VPC) and median odds ratios (MOR). Wagstaff-normalised concentration indices with Erreygers correction decomposed wealth-related inequality. Blinder–Oaxaca and Fairlie decompositions partitioned the rural–urban and tribal–non-tribal gaps. District-level MDD was tested for spatial autocorrelation using Moran’s I and Local Indicators of Spatial Association (LISA), with a bivariate LISA against district stunting. An XGBoost classifier with SHAP values ranked non-linear determinants. The ADP impact on MDD was estimated through a district-level difference-in-differences design comparing NFHS-4 to NFHS-5 with state fixed effects. National MDD prevalence was 22.9% (95% CI 22.4–23.5); only 11.3% met the Minimum Acceptable Diet criterion. State prevalence ranged from 14.9% (Rajasthan) to 53.5% (Meghalaya). Multilevel partitioning showed VPCₘ = 17.8% (district) and 6.1% (state), with district MOR = 1.86, indicating substantial contextual effects. The Erreygers-corrected concentration index was + 0.144 (pro-rich); maternal education (41.2%) and household wealth (22.7%) contributed the largest shares of inequality, with antenatal-care attendance third (12.4%). Bivariate LISA identified 84 High–High co-cluster districts spanning Bihar, Uttar Pradesh, Madhya Pradesh, Jharkhand, and parts of Rajasthan and Gujarat. XGBoost-SHAP ranked maternal education, household wealth, child age, and antenatal-care attendance as the top predictors (AUC = 0.83). The ADP × NFHS-5 DiD coefficient was + 0.7 pp (95% CI − 0.6 to + 2.1; p = 0.31), indicating no statistically significant ADP effect on MDD net of pre-trends and state fixed effects. Approximately three in four Indian children aged 6–23 months still fail the WHO/UNICEF MDD criterion. The deficit is spatially clustered, co-located with stunting, and substantially shaped by district- and state-level contextual factors. Region-specific, multi-sectoral interventions that combine maternal-education investments, antenatal counselling on complementary feeding, and food-system strengthening in the identified co-cluster districts are required, alongside an explicit dietary-diversity target in the ADP, POSHAN 2.0, and Saksham Anganwadi frameworks.