SafetyAtlas
2026 · Vol. I
Safety Atlas Research · Procurement Guide

How to evaluate crime risk data

Direct answerGood crime-risk data is traceable, coverage-aware, geographically defensible, normalized with a disclosed denominator, and reproducible from versioned source inputs. A single score without those fields is not enough.

The ten-check buyer audit

#CheckEvidence to request
1Source provenancePrimary source URL, file or endpoint, access date, and source year
2Agency coverageParticipating ORIs, mapping rules, and explicit missing-data flags
3Geographic fitBoundary or service-area definition for every place
4Offense definitionCodes, labels, inclusions, exclusions, and system used
5Zero handlingRule separating a reported zero from null or unavailable
6NormalizationPopulation source, year, geography, and per-capita formula
7RecencyLatest complete period plus update schedule
8Score constructionWeights, caps, shrinkage, and eligibility rules
9LicensingPermitted use, redistribution, attribution, and retention terms
10ReproducibilityVersion identifier, QA sample, and correction process

Red flags

Match evidence to the decision

Newsrooms may prioritize transparent sources and reproducible charts. Insurance and investment teams may need consistent licensing, batch delivery, and audit logs. Policy researchers may require longitudinal fields and sensitivity tests. In every case, coverage should be a first-class field rather than a footnote.

Test a sample before accepting a dataset

  1. Ask for one populated record, one reported zero and one unavailable record. Require a source reference and explanation for each.
  2. Recalculate one per-capita rate from the supplied count and matched population: count divided by population, multiplied by 100,000. Treat rounding separately from a denominator mismatch.
  3. Compare two years for the same agency and offense. Check reporting months, definitions and participation before calling the difference a trend.
  4. Trace one city mapping to its agency service area. Ask for a visible limitation if the geographic units do not match.
  5. Retain the vendor version, retrieved date and correction contact with your review.

Worked arithmetic example

In a hypothetical record, 20 reported offenses and a matching covered population of 10,000 give a rate of 200 per 100,000. This is an arithmetic example, not a claim about any real city. An unavailable count cannot be turned into zero, and an annual population cannot automatically explain every monthly rate.

Record an acceptance decision

Accept a record only when the source, period, geography, definition and denominator can be reconciled. Mark incomplete evidence for review and keep it out of rankings until resolved. This distinguishes a missing-data problem from a place with a low reported count.

Use the reporting-system guide for terminology and the product comparison guide to choose a delivery format that fits your question.