Methods

Every report follows the same approach. Each report lists its own sources, definitions and caveats; this page explains what they have in common.

Rates, not counts

Counts show where most reports come from, but they mostly reflect how many people are in a group. Rates divide by that population, so groups of different sizes can be compared.

A rate here is reports per 100,000 residents a year: the number of victims in a group, divided by the group's resident population, divided by the number of years covered, times 100,000. A ratio compares two rates. "3 times" means one group's rate is three times the other's.

Sources

• Police records: incident data from city open-data portals and the FBI's National Incident-Based Reporting System (NIBRS).
• Population: American Community Survey 5-year estimates by race, sex and age, for each city and its census tracts, through Census Reporter.
• Other measures, where a report tests what police data captures: the National Crime Victimization Survey, hospital emergency department records and 911 call data.

Who is in each group

• Victim race is recorded by officers, as they see it. The Census counts residents by their own answers. Because the two can differ, each report shows how far its main comparison could move if multiracial victims were recorded under a single race.
• Most police data records Hispanic ethnicity apart from race, and it is often missing. Reports say how often, and how the comparison would change if victims with unknown ethnicity were Hispanic.
• Groups with very small counts are included in totals but not compared, because their rates are unstable. Comparisons that shift a lot between periods are kept out of headlines.
• Victims with unknown race or sex are left out of rates. Each report says how many there are and, where the data allows, where they are concentrated.

Testing explanations

A gap between rates can have many sources. Each report tests the ones its data allows:

• Age: rates standardized to one age mix.
• Place: rates compared inside police districts or divisions, and neighborhoods compared by census tract.
• Type and severity: aggravated and simple assault separately, and injuries and weapons where recorded.
• Time: each year separately, to see whether a gap holds.
• Relationship: with and without intimate partner assaults, where the data records them.
• Model: a tract-level model that adds age, year, district and neighborhood income, poverty, unemployment and housing one step at a time.
• Replication: the same rates in a second source, such as a new records system or a later period.

Neighborhood conditions are themselves shaped by segregation and disinvestment. When a gap narrows after these controls, the report says where the gap sits. It does not treat the gap as explained away.

What police data cannot show

• Only crimes that reach the police are counted. Willingness to report, and how reports are recorded, can differ by group, place and time.
• Rates count reports, not people. Someone assaulted twice is counted twice.
• Rates divide by where people live, not where they spend their time.
• Police data reflects where officers patrol and who calls them. It cannot separate more policing or more reporting from more crime.
• The data describes victims. It does not measure causes, offenders or circumstances, and the reports do not suggest any.

What the assault-victim gap measures tests how much of the gap in police records comes from reporting, recording and counting the same people more than once.

Wording

• Findings say what the data shows and stop there. They name the groups being compared and say what the analysis did not explain.
• A rate from police records is a rate of reported crimes. It measures what reached the police, not every assault, and summaries on this site say "reported" for that reason.
• Before publication, every report that compares racial groups is checked for bias in its data and its wording: how race is coded, where race is missing, unstable comparisons, language that implies causes or offenders, and missing caveats.

Reproducing a report

Each report's numbers are generated by scripts in its public repository, linked from the report. The shared tools are open source in disparity-kit. Each repository explains how to rebuild its report from the raw data.

Corrections

If you find an error, email martin@spectahq.com. Corrections are made on the report and noted with the date.