Public Information Is an Incentive for Politicians: Experimental Evidence from Delhi Elections

Abhijit Banerjee; Nils Enevoldsen; Rohini Pande; Michael Walton (2024) · American Economic Journal: Applied Economics

Published

October 10, 2026

import sys
from pathlib import Path

ROOT = next(
    p for p in [Path.cwd(), *Path.cwd().parents]
    if (p / "scripts/tranche50_report.py").exists()
)
sys.path.insert(0, str(ROOT / "scripts"))
from tranche50_report import report
report("aeraej-145", tranche="tranche100")

Abhijit Banerjee; Nils Enevoldsen; Rohini Pande; Michael Walton (2024). American Economic Journal: Applied Economics. Table 5.

Collection index · Results JSON · QMD · Reproduction code

Source specification

Ward effects and ward clustering are retained. The source pools survey rounds into its pre/post indicator; slum observations remain uncollapsed.

Source coefficient Template on source sample Source N Template source-sample N
-0.048032 -0.048032 932 932

Source-to-template coefficient difference: 1.42e-15, in original outcome units.

Role Column
Outcome toilet_index
Treatment policy
Unit ward_rd2
Time post
Cluster ward_rd2

No additional covariates beyond fixed effects.

Estimates below divide the outcome by its source-sample standard deviation, 0.809565. Static TWFE is the treatment coefficient. Dynamic TWFE, BJS and Sun–Abraham average supported post-treatment cells by observation count. CS-DID additionally reports an equal-event-time average.

Analysis sample and treatment timing

Observations Units Treated Control units Cohorts Clusters
932 108 51 57 1 108
Sample restriction Count
additional rows removed 4

Adoption is the first observed treated period. Initially treated units and units without an observed minus-one baseline are excluded. Post-adoption untreated rows are excluded for switching policies. The plot shows at most 100 seeded-sample units; estimation uses the full eligible sample. Empty cells remain missing.

y__d: 100 of 108 units shown (seed 20261008). All eligible units enter estimation. Purple: untreated; yellow: treated; white: unobserved.

Estimates

Estimator Estimate (SD) SE 95% lower 95% upper Target Status
Static TWFE -0.059331 0.037769 -0.1342 0.015541 Coefficient estimated
Dynamic TWFE -0.059331 0.037769 -0.1342 0.015541 Treated cells estimated
BJS / did2s -0.058038 0.03733 -0.13204 0.015965 Treated cells estimated
Sun–Abraham -0.059331 0.037769 -0.1342 0.015541 Treated cells estimated
CS-DID (equal event-time) -0.052798 0.039061 -0.12936 0.02376 Equal post-event times estimated
CS-DID (treated cells) -0.052798 0.039061 -0.12936 0.02376 Treated cells estimated
SDID -0.090006 nan nan nan Treated cells estimated
Matrix completion -0.090592 nan nan nan Treated cells estimated
Horizontal ridge 0.0067324 nan nan nan Treated cells estimated
Multisynth (cohort-averaged) nan nan nan nan Treated cells unsupported
Multisynth (individual) nan nan nan nan Treated cells unsupported

Intervals use the configured clusters and the estimator’s available inference. Hollow points have no computed interval; no new refit-bootstrap batch is run.

Estimator Reason
Multisynth (cohort-averaged) multisynth-cohort: requires one observation per unit-period; no observations collapsed.
Multisynth (individual) multisynth-individual: requires one observation per unit-period; no observations collapsed.

Event studies

Cohort homogeneity

The imported TestingInEventStudies deviations design tests positive-event-time cohort deviations jointly. This is the homoskedastic nested-model F-test; the estimator confidence intervals above use clustered inference.

**y__d: homogeneity test not identified.** Only one treated cohort remains on the common supported sample.

Fitting settings and checks

Specification Estimator Settings Pooling
y__d CS-DID {“est_method”: “dr”, “covariate_basis”: “intercept”, “control_group”: “notyettreated”, “xformla”: “y ~ 1”, “panel”: false, “effective_panel”: false, “allow_unbalanced_panel”: true, “weights_name”: “weight”, “fix_weights”: null, “base_period”: “universal”, “clustervar”: “cluster”, “aggregate”: “dynamic”, “faster_mode”: true} {}
y__d SDID {} {“parameter”: null, “value”: null, “status”: “not_applicable”, “fit_scope”: “Separate adoption-cohort fits”, “within_cohort_target”: “Mean treated outcome history”, “cross_cohort_partial_pooling”: false, “shared_structure”: null, “note”: “Regularization penalties are not partial-pooling parameters.”}
y__d Matrix completion {“lambda_fraction”: 0.25, “max_iterations”: 1000, “tolerance”: 1e-12} {“parameter”: null, “value”: null, “status”: “not_applicable”, “fit_scope”: “Joint untreated-cell fit”, “within_cohort_target”: null, “cross_cohort_partial_pooling”: false, “shared_structure”: “Low-rank matrix and unit/time effects”, “note”: “Regularization penalties are not partial-pooling parameters.”}
y__d Horizontal ridge {“penalty”: 1.0} {“parameter”: null, “value”: null, “status”: “not_applicable”, “fit_scope”: “Separate adoption-cohort fits”, “within_cohort_target”: “Mean treated outcome history”, “cross_cohort_partial_pooling”: false, “shared_structure”: null, “note”: “Regularization penalties are not partial-pooling parameters.”}
y__d Multisynth (cohort-averaged) {} {}
y__d Multisynth (individual) {} {} 
{
  "did2s_vs_direct_imputation_max_post_difference": 6.938893903907228e-18,
  "package_aggregate_vs_estimable_cell_weights": 0.0
}