What Drives MGNREGA Work Demand? A District-Level Regression Study of Uttar Pradesh and Bihar (R)

The Mahatma Gandhi National Rural Employment Guarantee Act (MGNREGA) promises every rural household up to 100 days of wage employment a year on demand. Yet the number of days households actually get varies enormously from one district to the next — even between neighbouring districts in the same state, with the same notified wage rate. This dissertation asks which district characteristics explain that variation.

The study builds a cross-section of 113 districts — 75 in Uttar Pradesh and 38 in Bihar — for financial year 2022-23 entirely from public secondary sources: the MGNREGA MIS (person-days, households worked, households completing 100 days), Census 2011 district tables (agricultural labourers, SC/ST share, female literacy), NITI Aayog's National Multidimensional Poverty Index 2023 (district headcount ratios) and IMD district monsoon rainfall departures.

Two outcomes are modelled with ordinary least squares in R: average person-days per working household and the share of working households that completed 100 days. The analysis reports heteroskedasticity-robust standard errors, checks multicollinearity (VIF), functional form (RESET) and influential districts (Cook's distance), and compares the two states with a state dummy and an interaction.

The bundle gives you the full research design, a variable dictionary with sources, an R workflow you can reproduce from scratch, a chapter plan for the UGCF Semester 7–8 dissertation, a presentation outline and a viva bank. The data assembly is the real work here — and the part your supervisor will check most carefully.

Abstract

MGNREGA is India's largest public-works programme, but its reach is uneven: some districts provide 60+ days of work to the average participating household while others struggle to cross 35. Using a cross-section of 113 districts of Uttar Pradesh and Bihar for FY 2022-23, this dissertation estimates the association between MGNREGA work outcomes and district-level distress, labour-market and social indicators. Outcomes are taken from the MGNREGA MIS; predictors from Census 2011, NITI Aayog's National MPI 2023 and IMD rainfall data. OLS models with robust standard errors are estimated in R. The findings are discussed in the light of the "self-targeting" argument for employment guarantees and the literature on administrative capacity and rationing.

Introduction

An employment guarantee is designed to be self-targeting: because the work is manual and the wage modest, only households that need it will ask for it. If the programme works as designed, districts with greater distress — more landless labourers, higher poverty, a failed monsoon — should generate more person-days. If instead outcomes track administrative capacity or social inequality in who gets heard, the pattern will be different. District data let an undergraduate test this idea with public, verifiable numbers.

Literature and gap

Earlier studies of MGNREGA have examined participation using household surveys (NSS, IHDS), documented rationing — households wanting work but not getting it — and analysed wage payment delays and the role of state capacity. Much of this work is either national in scope or based on survey rounds that are now a decade old. District-level, post-pandemic analyses that combine the MIS with the new NITI Aayog MPI are rare, particularly comparisons between two high-poverty states with very different administrative histories. This dissertation fills that gap on a modest scale.

Research questions

1. How much do MGNREGA person-days per household vary across districts of UP and Bihar in FY 2022-23?
2. Are districts with a higher MPI headcount and a larger share of agricultural labourers associated with more person-days?
3. Is a monsoon rainfall deficit associated with higher work demand in the same year?
4. Does the association differ between Uttar Pradesh and Bihar?

Feasibility

• Data: every variable is published by a government body and downloadable free of cost.
• Tools: R and RStudio are free; the econometrics needed (OLS, robust errors, diagnostics) is covered in Introductory Econometrics.
• Time: two semesters — Sem 7 for review, data assembly and a pilot regression; Sem 8 for final models, robustness checks and writing — fit the UGCF 6 + 6 credit structure.
• Risk: the main risk is inconsistent district names across sources; a concordance table is built in the first month to manage it.

MGNREGA is meant to provide work when and where it is needed, yet public data show large differences in the average number of days households receive across districts with similar poverty levels and identical wage rates. Policy discussions often attribute these gaps either to demand ("people there don't need the work") or to supply ("the administration doesn't open works"), usually without systematic evidence.

For Uttar Pradesh and Bihar — two states that together account for a large share of India's multidimensionally poor — there is little recent district-level analysis that links MGNREGA outcomes to measurable indicators of need. This dissertation assembles a 113-district dataset from official sources and uses regression analysis to test whether work outcomes are associated with poverty, the agricultural labour share, social composition and monsoon shocks, and whether these associations differ between the two states. The results indicate whether the programme's outcomes are consistent with its self-targeting design and where closer field study is warranted.

In scope

• All districts of Uttar Pradesh (75) and Bihar (38) as they existed in FY 2022-23.
• A single-year cross-section (FY 2022-23), with FY 2021-22 used only as a robustness check.
• Public secondary data; OLS regression with robust standard errors and standard diagnostics in R.
• Interpretation in terms of need-based (self-targeting) versus capacity-based explanations.

Out of scope

• Household-level or village-level analysis, and any primary survey.
• Causal identification strategies (instrumental variables, difference-in-differences) — noted as future work.
• Wage-payment delays, social audits and corruption, which need different data.
• Districts created after 2011 whose Census values cannot be separated cleanly (handled by merging with the parent district and flagging them).

Research design

A quantitative, cross-sectional, correlational design using secondary data at the district level. The unit of analysis is the district; the reference year is FY 2022-23.

Variables and sources

Model

PD_HH_i = b0 + b1 MPI_HC_i + b2 AGLAB_i + b3 SCST_i + b4 FLIT_i
          + b5 RAIN_DEV_i + b6 BIHAR_i + e_i

Model 2 replaces the outcome with HH100; Model 3 adds BIHAR × MPI_HC to test whether the poverty association differs by state.

Estimation and diagnostics

• OLS with HC1 heteroskedasticity-robust standard errors (Breusch–Pagan test reported).
• VIF for multicollinearity (MPI and female literacy are expected to correlate).
• Ramsey RESET for functional form; log specification as an alternative.
• Cook's distance to identify influential districts; results re-estimated without them.
• Robustness: FY 2021-22 outcomes; dropping merged post-2011 districts.

Timeline (two semesters, monthly progress reports)

• Panel data: extend to FY 2014-15 to FY 2023-24 and estimate district fixed-effects models to control for time-invariant administrative differences.
• Causal design: use rainfall shocks as an instrument for agricultural distress, or compare districts before and after a policy change.
• Spatial econometrics: test for spatial autocorrelation (Moran's I) and estimate a spatial lag model with the spdep package.
• Field validation: follow up two high- and two low-performing districts with interviews of gram rozgar sevaks and job-card holders.
• Gender lens: model women's share of person-days alongside the female labour-force indicators from PLFS.