# Methodological revision: scope and issue resolution

24 September 2026. This document responds to methodological concerns about the initial release. Revised numerical results and the independently frozen sensitivity plan accompany it. A change in statistical significance is not a target of the revision. The original release remains available for comparison.

| Feedback topic | Action | Remaining boundary |
|---|---|---|
| 1. Outcome described as crime | Use arrest-record groups consistently in the study title, abstract, estimates and downloads; distinguish the broader research question from the measured outcome. | Arrest records cannot be transformed into a census of crime, victims, calls or people. |
| 2–3. Small estimate, confidence interval and significance | Report the amended reference estimate, interval, raw p-value and revision-family adjusted p-value together. Preserve the original estimate and all revised specifications. | Nonsignificance is neither proof of no association nor an error to be repaired. |
| 4. Enforcement/reporting pathway | Describe the outcome as recorded administrative activity. Keep causal language out of conclusions. | Reporting, police response and enforcement can affect counts independently of offending. |
| 5. Provenance | Recheck original hashes/headers and Agency fields; compare with official operational-report definitions. Publish a source-specific clarification draft. | No matching agency response letter, extraction query, database crosswalk or completeness certification was found. |
| 6. Changing identifiers and metadata conflicts | Correct the feedback's header error: 2024 retains Incident Number; only 2025 uses Arrest Ref Nbr. Retain conflict flags and unique-date/ZIP eligibility. | Exact source-ID, offense-date and location meanings remain unverified. |
| 7. Geographic exposure | Retain representative-point and grid metadata; explicitly distinguish ZIP-associated modeled weather from an incident-site reading. | Indoor/personal exposure and precise offense locations are unavailable. |
| 8. Weather measurement error | Add paired-sample NOAA/ERA5-Land sensitivity models using a fixed nearest-station/distance/elevation rule. Retain NOAA quality controls and all source-agreement results. | Station exposure is not ground truth at every ZIP; station observation windows differ. Do not apply an unsupported six-station bias correction. |
| 9. Weather dates and temporal ordering | Rebuild daily weather from UTC hourly data with America/Los_Angeles civil dates, including 23/25-hour days; fit previous-day maximum and elapsed-day-duration sensitivities. | An arrest-record date is not a verified offense date. Same-day maximum can occur later than a record; lagging cannot establish a causal exposure window. |
| 10. Confounding | Add observed federal holidays, precipitation, humidity and ZIP-specific annual seasonality, separately and together. Fit a fixed nonlinear temperature curve and dependence-window checks. | Tourism, local events, staffing, reporting, individual behavior and other daily factors remain unmeasured. Some proposed controls could be mediators; indiscriminate adjustment is not a causal design. |
| 11. Population offset | Explain the count-per-ZIP-day estimand and algebraic absorption of any fixed ZIP population offset by ZIP effects. Do not manufacture a population risk denominator. | Changing visitors/residents or agency coverage cannot be corrected by a constant Census denominator. A total-arrest offset would ask a different, conditional-composition question and may condition on temperature-related activity. |
| 12. Charge categories | Retain exact observed-code rules, overlapping-category warnings and source-label terminology. | These are not FBI crime categories, validated DV incidence or mutually exclusive events. |
| 13. DV and warrants | Audit all co-charges; add a strict screen removing 67 DV groups and 5,781 general groups from the original eligible sample, with both baseline and strict results visible. | Excluding explicit warrants does not establish that remaining violence occurred on the recorded date. |
| 14–16. Reproducibility, validity and confidence | Independently review the revised estimator, reproduce coefficients/covariances, reconcile inputs and publish runnable aggregate-only code. Label computational audits accurately. | Agreement between implementations does not authenticate records, verify completeness, establish external validity or constitute external human peer review. |
| 17. Newsroom wording | Provide wording tied to the measured outcome and uncertainty: supplied arrest records, observational association, and the reported interval. | Neither a nominally significant sensitivity nor a positive point estimate alone supports a causal headline. |
| 18. Agency questions | Prepare a targeted request for exact extraction/query coverage, unit definitions, date/geography meanings, schema changes, 2025 gaps and same-query control totals. | The draft has not been sent; agency-dependent questions remain open until documented answers arrive. |

## Why a population offset does not automatically solve the problem

For fixed ZIP population P_z, a model with log(P_z) as an offset and an unrestricted ZIP intercept a_z has the same fitted values as a model with intercept a_z + log(P_z). Adding that fixed offset therefore cannot change the within-ZIP temperature coefficient. It can change how the intercept is expressed, not create missing daily exposure-population data. Resident population also does not measure commuters, visitors or the population covered by an incomplete administrative export. The study continues to estimate record counts per ZIP-day.

## Revision inference

The fixed plan contains eleven specifications for each of two outcomes, including the paired legacy-weather comparator. Ten linear temperature coefficients per outcome plus spline joint and nonlinearity tests form 24 tests, corrected together with Holm's method. Individual confidence intervals and spline pointwise bands remain nominal, not simultaneous. HAC14/28 comparisons are declared covariance diagnostics. The model battery and quota-driven sample amendment were frozen before computing revised estimates, after the original results and external feedback were known; this remains retrospective work. A separately documented numerical amendment computes the extended-MLE boundary for exactly certified seasonal zero rows after a diagnosed sparse-DV separation failure; it preserves positive events and the planned temperature hypotheses.

Original source files, published aggregates and prior numerical outputs are preserved. The revised interactive weather explorer uses civil-day values for the 64 fully acquired ZIPs; the full record explorer's totals and category definitions remain the same. The provider's daily acquisition limit prevented completion of 48 additional ZIPs. The available subset contains 98.5% of general and 98.0% of DV groups in the original eligible sample, but availability is not random. Original-versus-corrected weather comparisons use exactly the same acquired ZIPs. The full original reference remains distinct. A formal list of release files, hashes, sources and executable code accompanies the amended publication.

## Documentation

- [Source and warrant review](PROVENANCE_RESOLUTION.md): exact headers, provenance distinctions, official report comparisons and statute links.
- [Frozen model plan](PLAN.md) and [independent methods plan](FROZEN_SENSITIVITY_PLAN.md).
- [Open-Meteo historical data documentation](https://open-meteo.com/en/docs/historical-weather-api); [ERA5-Land](https://doi.org/10.24381/cds.e2161bac).
- [NOAA daily data and quality-flag documentation](https://www.ncei.noaa.gov/pub/data/ghcn/daily/readme.txt).
- [Newey and West, original working paper](https://www.nber.org/papers/t0055); [Driscoll and Kraay](https://doi.org/10.1162/003465398557825); [Holm](https://www.jstor.org/stable/4615733).