Determinants of Child Stunting across Indian States: A Weighted Logistic-Regression Study of NFHS-5 Data

What this study does. Stunting (low height for age) is the most widely used indicator of long-term undernutrition in young children. The NFHS-5 national fact sheet reports that roughly one in three Indian children under five (35.5 %) is stunted, with large differences between states. This project asks a focused statistical question: after adjusting for other factors, which household and maternal characteristics are associated with higher odds of a child being stunted, and how much of the state-to-state gap remains?

We use the unit-level NFHS-5 (2019–21) children's recode file, obtained through a free registration with The DHS Program. The outcome is binary: a child is stunted when the height-for-age z-score is below −2 standard deviations of the WHO Child Growth Standards median. Covariates are mother's education, household wealth quintile, type of sanitation facility, birth order, child's age and sex, place of residence (urban/rural) and state.

The analysis respects the survey design: sampling weights, primary sampling units (clusters) and strata are declared with the R survey package before anything is estimated. We produce weighted descriptive tables, design-adjusted chi-square tests, a survey-weighted logistic regression reporting odds ratios with 95 % confidence intervals, model diagnostics (VIF, a goodness-of-fit check and ROC/AUC) and a state-level choropleth map. Key tables are cross-checked in SPSS and Excel so that the two-member team can defend every number in the viva.

Abstract

This project studies the determinants of stunting among children aged 0–59 months using the unit-level National Family Health Survey (NFHS-5, 2019–21) children's recode. Stunting is defined as height-for-age z-score (HAZ) < −2 SD of the WHO Child Growth Standards. Using R's survey package to account for weights, clustering and stratification, we estimate weighted prevalence by background characteristics, test associations with design-adjusted (Rao–Scott) chi-square tests, and fit a survey-weighted logistic regression. Results are reported as adjusted odds ratios with 95 % confidence intervals, followed by diagnostics and a state-level map.

Introduction

Stunting reflects chronic undernutrition and repeated infection in early life. It is linked to poorer schooling and adult earnings, which is why it appears in national nutrition programmes such as POSHAN Abhiyaan. India's NFHS series gives nationally and state-wise representative anthropometric data, which makes it an ideal real dataset for an undergraduate statistics project that goes beyond textbook examples.

Existing work and the gap

Official NFHS reports publish one-way tables — stunting by wealth, by education, by state — but a one-way table cannot separate overlapping effects. Richer households also tend to have better-educated mothers and improved toilets, so a raw gap by education partly reflects wealth. Many student reports also analyse NFHS data without weights or design variables, which biases estimates and understates standard errors. Our project fills this gap at undergraduate level by (a) building a multivariable model and (b) doing every estimate under the correct complex-survey design.

Proposed study

1. Obtain NFHS-5 KR data from The DHS Program, restrict to living, de jure children with a valid HAZ.
2. Recode covariates following the Guide to DHS Statistics definitions.
3. Weighted descriptive analysis and Rao–Scott chi-square tests.
4. Survey-weighted logistic regression (quasibinomial) with state fixed effects.
5. Diagnostics: VIF, goodness of fit, ROC/AUC; sensitivity analysis.
6. A choropleth map of weighted state prevalence and adjusted state effects.

Feasibility

• Technical: R, RStudio and all packages are free; SPSS is available in most DU college labs. The KR file is large but fits comfortably in 8 GB RAM once unnecessary variables are dropped.
• Economic: zero cost — the data is free for registered research use.
• Operational: a two-member team can finish the work in about 14 weeks alongside Semester 6 courses; data access takes a few working days, so it is requested in Week 1.

Although stunting has declined slowly over successive NFHS rounds, the level remains high and uneven across Indian states. Policy discussions often quote single-factor gaps — "children of uneducated mothers are more often stunted" — but these factors overlap strongly with household wealth, sanitation and rural residence. Without a multivariable model it is impossible to say which factors remain important after the others are held constant, or whether the differences between states are explained by their socioeconomic profile.

A second problem is methodological. NFHS is a stratified two-stage cluster sample, and treating it as a simple random sample produces biased prevalence estimates and artificially narrow confidence intervals, which can make weak associations look significant.

This project therefore sets out to estimate, from unit-level NFHS-5 data and under the correct survey design, the adjusted association between child stunting (HAZ < −2 SD) and mother's education, wealth quintile, sanitation, birth order, residence and state, and to report these associations with honest uncertainty and model diagnostics.

In scope

• Children aged 0–59 months in the NFHS-5 children's recode (KR) with a valid height-for-age z-score; all states and union territories covered by the survey.
• One binary outcome (stunted vs not stunted) and a fixed set of covariates: mother's education, wealth quintile, sanitation facility (improved / not improved / open defecation), birth order, child's age group and sex, residence and state.
• Weighted descriptive statistics, Rao–Scott chi-square tests, survey-weighted logistic regression and diagnostics.
• A state-level choropleth map and a short comparison of R output with SPSS/Excel output.

Out of scope

• Causal claims: the cross-sectional design supports associations, not cause and effect.
• District-level modelling, multilevel (random-effects) models and trend analysis across NFHS rounds — listed under future scope.
• Severe stunting (HAZ < −3 SD), wasting and underweight as separate outcomes.

Research design: secondary analysis of a nationally representative cross-sectional household survey, following a quantitative, hypothesis-testing approach.

Hypotheses (expected directions, to be tested — not results):

• H1: the odds of stunting decrease as mother's years of schooling increase.
• H2: the odds of stunting decrease across wealth quintiles from poorest to richest.
• H3: children in households without improved sanitation have higher odds of stunting.
• H4: higher birth order is associated with higher odds of stunting.
• H5: state differences remain significant after adjustment for the above.

Model specification

logit P(stunted = 1) = β0 + β1·MotherEdu + β2·Wealth + β3·Sanitation + β4·BirthOrder + β5·AgeGroup + β6·Sex + β7·Residence + Σ γs·State_s

Reference categories: no education, poorest quintile, improved sanitation, birth order 1, age 0–11 months, male, urban, and a large state chosen as the reference (justified in the report). Estimation uses svyglm() with family = quasibinomial(); adjusted odds ratios are exp(β) with Wald 95 % CIs.

Timeline (14 weeks, team of two)

Quality control: every recode is checked with a frequency table before and after; the unweighted count of children in the analytic sample is logged at each filtering step so the sample flow chart in the report can be reproduced exactly.

• Multilevel logistic models (children nested in clusters, districts and states) to separate household and area-level variation.
• District-level analysis using NFHS-5 district identifiers and small-area mapping.
• Trend decomposition between NFHS-4 and NFHS-5 to see which factors explain the change in stunting.
• Additional outcomes: severe stunting, wasting, underweight and anaemia, modelled jointly.
• Interaction terms such as mother's education × wealth, or sanitation × residence.
• Sharing a reproducible R Markdown / Quarto report so juniors can rerun the pipeline on the next NFHS round.