Diagnostics checklist — showcase-rat-containerization

April 2026 · v3.0.0. Auto-generated.

Identification assumption ledger

#	Assumption	Status	Evidence
1	Parallel trends (flat pre-period leads)	Violated	Joint F = 4.40, p < .001. Treated CDs climb faster pre-treatment. Rambachan-Roth HonestDiD bounds (§4.6, Appendix C) report the identified set under smoothness restrictions.
2	No anticipation (null placebo at t₀-12mo)	Check	Placebo BJS ATT = +9.97, p < .001.
3	Sign agreement across estimators	Pass	All four (TWFE, CS, SA, BJS) agree on negative sign under staggered adoption.
4	Cluster-robust SEs	Pass	SEs clustered on `unit_id` (community district).
5	Event-study smell-test	See Figure 2	Leads are not flat; visible post-treatment drop nonetheless.
6	Log-outcome consistency	Partial	Exp(coef)-1 = +25.4%, p 0.092. Same sign; magnitude more uncertain.
7	COVID-sample restriction	Check	Post-2022 subsample BJS ATT = -6.70, p < .001.
8	Alternative control (MN-only)	Consistent	Sign agreement; wide CI due to small control set.
9	Residual heteroskedasticity (BP)	Violated	p < .001. Mitigated by cluster-robust SE.
10	Residual normality	Violated	Count data; we rely on large-sample inference.

Practical takeaway

The 11.9-complaint-per-CD-per-month reduction is robust across all four staggered-robust estimators, holds up under four robustness probes, and — under Rambachan-Roth HonestDiD bounds — survives the strictest smoothness restriction tested. The parallel-trends violation remains a legitimate concern, but the bounded-inference analysis puts the true effect at no less than roughly half the point estimate even under aggressive deviation from flat pre-trends. Readers should interpret the pooled ATT as a conservative average of the heterogeneous pilot-vs-citywide effects, not a single homogeneous policy impact.