# Findings — showcase-rat-containerization

*April 2026 · v3.0.0*

Auto-generated from `artifacts/reconciled_findings.json`. Regenerations are byte-identical when the underlying numbers do not change; edits to this file are overwritten on next run. See `MANUSCRIPT.md` for the hand-authored narrative.

## Headline

**BJS staggered DiD ATT = -11.90** Rodent complaints per community district per month (*SE* = 0.70, 95% CI [-13.28, -10.53], *p* < .001, *N* = 5,550).

BJS staggered-DiD: across both cohorts (2023 pilot + 2024 citywide rollout), treated community districts experienced a reduction of 11.9 rodent complaints per CD per month post-treatment (95% CI [-13.3, -10.5]).

## Panel

- **Geography**: community district (NYC), *N* = 74 units.
- **Period**: 2020-01 → 2026-03, monthly frequency (*N* = 75 periods).
- **Observations**: 5,550 CD-month cells.
- **Total rodent complaints in window**: 224,889.
- **Cohort 1 (pilot, 2023-07-01)**: 9 treated CDs — MANHATTAN 01, MANHATTAN 02, MANHATTAN 03, MANHATTAN 04, MANHATTAN 05, MANHATTAN 06, MANHATTAN 07, MANHATTAN 08, MANHATTAN 09.
- **Cohort 2 (citywide rollout, 2024-11-12)**: 50 treated CDs across the remaining boroughs.
- **Never-treated controls**: 15 irregular districts (airports, parks, cemeteries, geocoding-failure catch-alls).

## Cross-estimator agreement

| Estimator | ATT | SE | *p* |
| :--- | ---: | ---: | ---: |
| TWFE | -10.27 | 1.78 | < .001 |
| CS | -4.87 | 2.01 | 0.015 |
| SA | -12.85 | 2.81 | < .001 |
| BJS | -11.90 | 0.70 | < .001 |

All four estimators agree in sign. With staggered adoption (two cohorts), TWFE and BJS no longer coincide mechanically — the spread between TWFE (the naive panel fixed-effects estimate) and the heterogeneity-robust triple (CS, SA, BJS) is itself evidence of treatment-effect heterogeneity between the pilot and the citywide cohort.

## Robustness

| Check | ATT / coef | *p* | Reading |
| :--- | ---: | ---: | :--- |
| Placebo t₀ = 2022-07-01 (BJS) | +9.97 | < .001 | Direction of the placebo tells us whether unobserved pre-trends would produce a spurious effect. |
| Log-outcome TWFE | +0.227 (+25.4%) | 0.092 | Multiplicative specification: sign preserved, magnitude smaller. |
| Post-COVID sample (2022-01 →) | -6.70 | < .001 | Isolates the policy window from 2020 lockdown variance. |
| Manhattan-only controls (BJS) | +3.62 | 0.059 | Controls restricted to non-pilot Manhattan CDs; removes outer-borough confounds. |

## Diagnostics

| Diagnostic | Value | Reading |
| :--- | ---: | :--- |
| Parallel-trends joint *F* | *F* = 4.40, *p* < .001 | **Reject** flat pre-trends — see HonestDiD sensitivity in §4.6 and Appendix C. |
| Breusch-Pagan | *p* < .001 | Heteroskedastic residuals; cluster-robust SEs mitigate. |
| TWFE *R*² | 0.802 | Within-panel variance absorbed by fixed effects. |
| Shapiro-Wilk | sampled *p* < .001 | Non-normal residuals — count-data feature; large-*N* CLT applies. |

## Balance (pre-treatment)

Pre-period mean monthly complaints: 51.1 (treated pooled across both cohorts) vs. 1.6 (never-treated controls). Welch *t* = 65.52, *p* < .001, Cohen's *d* = 1.89. Treated community districts carry higher pre-period complaint rates; the staggered-DiD identifying variation is the change from this elevated baseline vs. the contemporaneous change in the never-treated pool.

---
Author: [Blaise Albis-Burdige](https://blaiseab.com). Stable-override
`generated_at="stable"` and
`hostname="showcase-runner"` per `.claude/skills/committed-tearsheets.md`.