Small area variations in four measures of poverty among Indian households: Econometric analysis of National Family Health Survey 2019-2021

The study investigates within-district and between-community (cluster) variations in poverty across India to inform effective targeting of poverty-eradication policies. Despite substantial national reductions in poverty, progress has been uneven geographically, with 22% of people living on less than USD 1.90 per day as per the 2011 census. The authors assess four measures of household poverty—bottom 10th wealth percentile, bottom 20th wealth percentile, below-poverty-line (BPL) card status, and multidimensional poverty index (MPI)—using NFHS-5 (2019–2021) to quantify small area variation and identify where inequalities are most pronounced, thereby guiding targeted policy interventions.

The paper situates its analysis within a broad literature documenting rising income and wealth inequality in India and spatial disparities across states and districts (e.g., Chancel and Piketty, 2019; Mishra and Bhardwaj, 2021; Subramanian and Jayaraj, 2013; Alkire and Seth, 2015; Mohanty et al., 2016; Menon et al., 2018). Prior multilevel studies highlight the importance of small-area geographies (clusters) in shaping socioeconomic and health outcomes, including poverty and child undernutrition (Kim et al., 2016; Jain et al., 2021; Rajpal et al., 2021). Explanations for persistent and widening inequalities include between-caste disparities, regional variations in agriculture and climate, and differential infrastructure investments (Chauhan et al., 2016; Ghosh and De, 1998; Palmer-Jones and Sen, 2003, 2006; Zacharias and Vakulabharanam, 2011). The study also draws on social epidemiology literature emphasizing challenges in measuring wealth as socioeconomic position and the role of area-level indicators for health outcomes (Braveman et al., 2001; Howe et al., 2012; Kawachi et al., 2010; Wilkinson and Pickett, 2006).

Data source and sample: The study used the National Family Health Survey (NFHS-5, 2019–2021), which employed a two-stage cluster sampling design. Primary sampling units (PSUs) are clusters (groups of adjacent households). Rural and urban clusters were first selected; clusters >300 households were segmented and then sampled. Up to 22 households were selected per PSU. The analytic dataset included 2,795,894 de jure individuals nested in 30,170 clusters across 707 districts and 36 states/UTs. Outcomes: Four binary outcomes were constructed: - Bottom 10th wealth percentile (vs. above) based on the DHS wealth index distribution at the national level. - Bottom 20th wealth percentile (vs. above). - Below-poverty-line (BPL) status, defined by household possession of a state-issued BPL card entitling subsidized grain; all individuals in BPL households classified as BPL. - Multidimensional poverty (MDP/MPI) following the OPHI/UNDP methodology. Deprivations span three dimensions: health (nutrition, child mortality), education (years of schooling, current attendance), and standard of living (cooking fuel, sanitation, drinking water, electricity, housing quality, assets). Households with MPI score > 0.33 classified as multidimensionally poor; individuals in such households classified as MDP. Statistical analysis: Individuals (level 1) were nested within clusters/PSUs (level 2), districts (level 3), and states (level 4). For each binary outcome, a four-level variance-components logistic model with random intercepts at cluster, district, and state levels was estimated to decompose geographic variance into between-cluster, between-district, and between-state components. Level-1 residual variance is not estimable for logistic models. Variance partition coefficients (VPCs) were computed as each level’s variance divided by total geographic variance. Models were estimated in MLwiN 3.05 using MCMC with a burn-in of 500 and 5,000 monitoring iterations. Precision-weighted predictions: The authors derived precision-weighted cluster-level predicted prevalences and summarized within-district between-cluster heterogeneity using the standard deviation (SD) of cluster predictions within each district. District-level precision-weighted predicted prevalences were also generated. Correlations were computed between district-level prevalences across measures and between district prevalences and cluster SDs to assess associations among measures and the relationship between overall level and within-district heterogeneity.

- Sample: Of 2,795,894 individuals, 258,808 (bottom 10th percentile) and 532,760 (bottom 20th percentile) were classified as poor by wealth cutoffs. Of 2,791,372 with BPL data, 1,366,554 were BPL. Of 2,795,894 with MDP data, 438,955 were multidimensionally poor. - All-India prevalence (2021): bottom 10th wealth percentile 9.3%; bottom 20th 19.1%; BPL 48.9%; MDP 15.7%. - Correlations among district-level prevalences (Pearson r): • Bottom 10th vs. bottom 20th: r = 0.93 (p < 0.001). • Bottom 10th vs. MDP: r = 0.81 (p < 0.001). • Bottom 20th vs. MDP: r = 0.88 (p < 0.001). • Bottom 10th vs. BPL: r = 0.29 (p < 0.001). • Bottom 20th vs. BPL: r = 0.34 (p < 0.001). • BPL vs. MDP: r = 0.27 (p < 0.001). - Variance partitioning (share of total geographic variance): • Bottom 10th wealth percentile: states 66% largest; districts smallest share. • Bottom 20th wealth percentile: states 63% largest; districts smallest share. • BPL: states 54% largest; districts smallest share. • MDP: clusters 44% largest; districts smallest share. - Within-district between-cluster heterogeneity (SD of cluster prevalences within districts): • Bottom 10th wealth percentile SD range 0.0004 to 32.9; median 6.9. • Bottom 20th wealth percentile SD range 0.0001 to 33.6; median 14.2. • MDP SD range 0.04 to 31.0; median 12.1. • BPL SD range 2.6 to 31.2; median 17.6. - Correlation between district prevalence and cluster-level SD (within-district inequality): • Bottom 10th: r = 0.75 (p < 0.001). • Bottom 20th: r = 0.75 (p < 0.001). • MDP: r = 0.79 (p < 0.001). • BPL: r = −0.17 (p < 0.001). - Spatial patterns: Poverty generally clusters in north, central, and parts of eastern India, though patterns vary by measure; maps highlight districts with high between-cluster inequality.

The study addressed its central question by decomposing geographic variance and quantifying within-district, between-cluster heterogeneity across four poverty measures. Findings show that the dominant scale of variation differs by measure: states account for most variation in wealth-based and BPL measures, whereas clusters account for the largest share for multidimensional poverty. Districts contribute the least to total variance across all outcomes, underscoring the importance of sub-district targeting. Significant positive associations between district prevalence and within-district inequality for bottom wealth and MDP imply that districts with higher poverty also exhibit greater small-area disparities. In contrast, BPL’s weak negative association with within-district inequality suggests it captures different aspects of deprivation and targeting. These results have policy relevance: anti-poverty programs should consider smaller geographies (clusters/PSUs) when prioritizing interventions, complementing state-level strategies. The heterogeneity across measures highlights measurement challenges in using wealth as a socioeconomic indicator and supports incorporating area-level indicators. The discussion links to prior literature on India’s rising inequality and posits potential mechanisms—caste stratification, regional agriculture/climate differences, and infrastructure disparities—calling for tailored policies sensitive to within-district conditions. Given established links between poverty, inequality, and health outcomes, pinpointing districts with high between-cluster inequality can guide targeted public health and social welfare interventions.

Poverty in India varies not only between states and districts but also within districts across small areas (clusters). The degree of geographic inequality depends on both the administrative level and the poverty measure considered. Recognizing and addressing small-area variations can improve the design and targeting of anti-poverty programs and has potential to enhance health, social, and economic outcomes. Future research should investigate determinants of small-area inequality (e.g., caste, agriculture/climate, infrastructure) and explore how to tailor programs to varying within-district and between-cluster conditions rather than adopting uniform approaches.

Two main limitations are noted: (1) Some household wealth measures in NFHS are self-reported, introducing potential measurement error, though NFHS data quality is generally high. (2) Precision-weighted estimates were not adjusted for sociodemographic correlates (e.g., caste, education of household head), which could bias estimates and interpretation of small-area variation.