Recorded Crime, Reported Crime: Measured Underreporting and the (Un)Identifiability of Peru States-of-Emergency Effect on Crime

Working paper (v2, corrected inference). Governments evaluate mano-dura security policies against recorded crime, but recorded crime is a censored, endogenous measurement: whether a victim files a police complaint depends on institutional trust and police presence, which the policy itself changes. Using Peru 2024-2026 states of emergency, we make this reporting endogeneity measurable. Combining a district-month panel of police complaints (SIDPOL), a verified inventory of emergency decrees, and the national victimization survey (ENAPRES, which asks about both victimization and reporting), three findings follow. First, the recorded effect on extortion complaints is estimator-fragile: a two-way fixed-effects Poisson yields +15 percent but a heterogeneity-robust staggered estimator collapses it to a sign-mixed null. Second, identified directly from the survey, the reporting rate of extortion rose from 7.3 percent in 2017-2021 to 23.7 percent in 2022-2024, a survey-design-adjusted increase of about +16 percentage points (95 percent CI +8.5 to +24.0 via PSU cluster bootstrap with weights, p<0.001; weighting-sensitive, unweighted +8.2 points), roughly five times the largest increase for any other crime (internal placebo); the eightfold rise in recorded extortion is therefore substantially a reporting phenomenon, not a pure crime increase. Third, the clean lethal-violence outcome that could adjudicate the emergency effect is not openly available at monthly district resolution. We formalize the resulting partial identification with Manski bounds disciplined by the measured reporting rate, and conclude that Peru administrative data cannot on their own verify the government headline security claims. This version reports survey-design-adjusted inference. Open, reproducible, CPU-only.