Beaman et al. (2012) Replication

Python replication and robustness analysis of:

Beaman, L., Duflo, E., Pande, R., & Topalova, P. (2012). "Female Leadership Raises Aspirations and Educational Attainment for Girls: A Policy Experiment in India." Science, 335(6068), 582-586. doi:10.1126/science.1212382

Overview

This repository replicates the main tables from Beaman et al. (2012) and conducts inference robustness checks. The paper examines whether exposure to female political leaders affects adolescent girls' aspirations in rural India, exploiting random assignment of village council seats to women.

Key Findings

Baseline Replication

Baseline means in never-reserved villages match the published values within rounding tolerance.

Outcome	Boys (paper)	Boys (reproduced)	Girls (paper)	Girls (reproduced)
no_housewife	0.998	0.998	0.600	0.585
wish_graduate	0.296	0.290	0.195	0.190
marry_after18	0.980	0.980	0.660	0.662
wish_pradhan	0.499	0.504	0.485	0.487

Inference Summary

P-values for the twice-reserved × female interaction across all seven outcomes:

Table	Outcome	Coefficient	p (raw)	p (wild)	p (Bonferroni)	p (BH)
2	no_housewife	0.089	0.036	0.051	0.252	0.160
2	marry_after18	0.089	0.069	0.131	0.480	0.160
3	can_read_write	0.062	0.069	0.101	0.479	0.160
2	wish_graduate	0.018	0.698	0.667	1.000	0.801
2	wish_pradhan	-0.014	0.801	0.768	1.000	0.801
3	attends_school	0.023	0.754	0.879	1.000	0.801
3	grade_completed	0.153	0.485	0.434	1.000	0.801

Notes on inference corrections:

• Wild cluster bootstrap addresses the small number of clusters (20 GPs in the twice-reserved cell) using Cameron, Gelbach & Miller (2008).
• Bonferroni controls family-wise error rate across seven tests.
• BH (Benjamini-Hochberg) controls false discovery rate.

One outcome (no_housewife) shows p < 0.05 with cluster-robust standard errors. After wild bootstrap correction, p = 0.051. After Bonferroni correction across all seven outcomes, no outcome reaches p < 0.05.

Data

Sources:

• Harvard Dataverse (J-PAL)
• Stanford Redivis

Required files:

File	Rows	Description
powerful_women_in_india_teenager_survey.dta	3,680	Adolescent aspirations
powerful_women_in_india_pradhan_seats_reserved_for_women.dta	165	Treatment assignment
powerful_women_in_india_household_roster.dta	37,263	Demographics
powerful_women_in_india_household_survey.dta	7,425	Location data
powerful_women_in_india_adult_survey.dta	13,508	Adult attitudes

Treatment cells:

• Twice-reserved: 20 GPs (both cycles)
• Once-reserved: 70 GPs (one cycle)
• Never-reserved: 72 GPs (control)

Methods

Robustness Checks

Method	Purpose	Reference
Wild cluster bootstrap	Small-sample cluster inference	Cameron, Gelbach & Miller (2008)
Bonferroni / Holm	Family-wise error rate control
Benjamini-Hochberg	False discovery rate control
Randomization inference	Permutation-based p-values
Alternative clustering	GP, household, village levels

Sensitivity Analysis

Analysis	Purpose
Leave-one-out	Influence of individual GPs
Ceiling effects	Boys near 100% on some outcomes
Age placebo	Adults socialized pre-1993
Unused survey items	E_1 to E_10 gender attitudes
Time use	School attendance, domestic chores

Running the Code

pip install -r requirements.txt

python scripts/01_replicate.py --data-dir data
python scripts/02_robustness.py --data-dir data
python scripts/03_mechanisms.py --data-dir data
python scripts/04_sensitivity.py --data-dir data

Output Files

File	Description
tabs/inference_summary.md	Combined p-values across corrections
tabs/table2_baselines.md	Aspiration baselines
tabs/table2_coefficients.md	Aspiration treatment effects
tabs/table3_baselines.md	Education baselines
tabs/table3_education.md	Education treatment effects
tabs/robustness_wild_bootstrap_all.md	Wild bootstrap for all outcomes
tabs/robustness_multiple_testing_all.md	Multiple testing corrections
tabs/robustness_clustering.md	Alternative clustering
tabs/robustness_randomization.md	Randomization inference

Codebook

Variable definitions are in scripts/codebook.py.

Key variables:

Variable	Source	Coding
no_housewife	B1_3	1 if occupation ≠ housewife
wish_graduate	B1_1	1 if B1_1 == 13
marry_after18	B1_2	1 if B1_2 >= 19
wish_pradhan	B1_4	1 if B1_4 == 1
twice_res	Treatment	1 if reserved both cycles

Caveats

• 162 of 165 GPs match between treatment and household files
• Wild bootstrap uses Rademacher weights imposing the null
• Bootstrap p-values shift slightly across runs; use --bootstrap-reps 4999 for stability

License

MIT