All papers and comparison plots · Results JSON · Quarto source
1 of 1 registered specifications have estimates. Every registered specification appears below.
Bertrand et al. (2021), QJE. https://doi.org/10.1093/qje/qjab023
Data and estimation
Outcomes are divided by their unweighted source-sample standard deviation (ddof=1), before exclusions; they are not centered. Treatment adoption is the first observed treated period. Models share an eligible sample, fixed effects, controls and weights. Initially treated units, missing −1 baselines and unsupported untreated-outcome predictions are removed. For switching treatments, untreated rows after first adoption are removed. Exact counts and settings are in the JSON.
Static TWFE uses the treatment indicator. Dynamic TWFE omits event time −1. BJS/did2s fits untreated outcomes, then averages imputation residuals. Sun–Abraham estimates cohort-by-event-time effects. The three dynamic estimates share post-period aggregation weights. Event-study tails are unbinned. These are common-sample reruns, not Peter’s original estimator-specific sample and tail-binning rules.
The homogeneity test restricts identified cohort deviations at positive event times to zero; leads and event time 0 remain unrestricted. It is a homoskedastic nested-model F-test. Estimate intervals are separately cluster-robust, with the did2s two-stage correction. Large panels use algebraically equivalent sparse regressions; backend and validation results are recorded in JSON.
Adoption timing
Hall of Mirrors: 100 sampled trajectory patterns, not individual pairs. Exact frequency weights represent the underlying pair panel.
| Specification | Status | F_pvalue | Static TWFE | Dynamic TWFE | BJS / did2s | Sun–Abraham |
|---|---|---|---|---|---|---|
| S11-P001 | estimated | 7.706e-89 | 0.2471 | 0.1472 | 0.2604 | 0.1606 |
Specifications
Open a specification for its estimates, event studies and homogeneity test. Plots display −10 through 10; JSON contains the full paths. Large-panel intervals are calculated over the displayed window.
S11-P001 · HALL-MIRRORS-F1-COCOMMENT · homogeneity rejected
Specification JSON · Configuration
HALL-MIRRORS-F1-COCOMMENT
Source rows: 38,593; fitted rows: 38,528; units: 3,253; clusters: 3,253; treated cohorts: 4. Source outcome SD: 0.041438236.
Rows and units above count compressed trajectories. The frequency-expanded F-test uses 121,814,681 observations.
| method | estimate | se | lower | upper |
|---|---|---|---|---|
| Static TWFE | 0.24712 | 0.023894 | 0.20029 | 0.29396 |
| Dynamic TWFE | 0.14724 | 0.032343 | 0.083853 | 0.21064 |
| BJS / did2s | 0.26045 | 0.023866 | 0.21367 | 0.30722 |
| Sun–Abraham | 0.16055 | 0.031515 | 0.098785 | 0.22232 |
Post-period weighting: equal_periods. Homogeneity F(15, 112091782) = 30.77; p = 7.706e-89.
Reproduce
Code, configurations and environments · Generic Quarto template · Editable Jot template. Source datasets remain local. Run python templates/full_archive/run_archive.py, then python templates/full_archive/build_collection.py.