What Do the Racial Violence Statistics Actually Measure?

Almost every argument in this area is conducted with statistics that answer a different question from the one being asked. There are three questions, and they need different data.

Was the violence racially motivated? That is hate crime. Measuring it requires a record that someone judged the motive, which means a police bias flag, a prosecution decision, or a victim saying so in a survey.

Did the violence cross racial lines? That is interracial crime. It needs the recorded race of a victim and an offender and nothing else. Motive does not enter into it. A robbery, a bar fight and a domestic homicide all count if the two people were recorded as different races.

Do rates differ between groups? That is a disparity. It needs a numerator, a population denominator, and an argument that the two describe the same people.

The numbers that circulate are overwhelmingly from the second and third categories. The claims made with them are overwhelmingly about the first. That single substitution accounts for more of the confusion in this subject than every data quality problem in this article put together, and it is committed in both directions.

Four systems produce these figures and they do not measure the same thing. Arrest data records who police apprehend. Police-recorded crime records incidents that were reported and then logged under a counting rule. Victimisation surveys record what people say happened to them. Conviction data records court outcomes. A claim that holds in one and reverses in another is not a finding about violence; it is a finding about the instrument.

Two consequences follow, and both are load-bearing later in this article.

A change in police-recorded figures can come from reporting, recording, investigation, clearance or classification. None of those is offending. When a recorded count rises, the rise has to be attributed before it means anything, and the agencies that publish these series usually say so themselves in text that does not travel with the screenshot.

Missing data is not neutral. Where a large share of records have no offender race attached, the cases that are missing are not a random sample of the cases that are present. Any claim built on the known share is a claim about solved, recorded, classified crime, which is a different population from all crime.

The image is almost always the same one. FBI Uniform Crime Reporting, Crime in the United States 2019, Expanded Homicide Data Table 6, "Race, Sex, and Ethnicity of Victim by Race, Sex, and Ethnicity of Offender." Its figures are real and they are these.

Four things about this table are not visible in a screenshot of it.

It covers under half of the homicides in the FBI's own file. The four rows sum to 6,578. The FBI received supplementary homicide data for 13,927 murder victims in 2019. Table 6 is restricted to incidents with a single victim and a single offender, which the FBI states were 50.6% of the murders it received supplemental data for. The screenshot is a subset of a subset.

The offender was never identified in nearly half of all cases. This is the figure that decides how much the table can carry, and it is published one page away from it: in 2019 the offenders were not known for 6,808 murder victims, 48.9% of the total. Across the wider file the race of the offender was unknown for 4,752 offenders. A table of victim race by offender race can only contain solved cases. Clearance rates are not equal across places or victim groups, so the missing half is not missing at random.

They are counts. Nothing in the table is a rate. Converting 566 and 246 into anything comparable requires dividing each by the population of the group it came from, and the table supplies no denominators. The arithmetic people perform on these two numbers is their own construction, not an FBI figure.

Race and ethnicity are separate fields, and one of them is mostly empty. In the same table, of the 3,299 white-victim cases, 738 offenders were recorded Hispanic or Latino, 1,542 not Hispanic or Latino, and 1,019 unknown. Nearly a third of the ethnicity field is blank, and Hispanic offenders appear inside the race column as white. "White offender" in this table is not a count of non-Hispanic white offenders, and no adjustment in the table lets you produce one.

None of the four rows contains a motive field. Whatever the table shows, it is not hate crime.

The same source that produces the offender table also produces victim figures, and those are rarely screenshotted alongside it.

The Bureau of Justice Statistics reports that in 2023 the homicide victimisation rate was 21.3 per 100,000 for black Americans and 3.2 per 100,000 for white Americans. That is a ratio of about 6.7 to 1, and unlike the offender figures it is a rate with a stated population base, drawn from records where the victim's race is documented rather than inferred from a solved case.

An independent compilation using the FBI's own Supplementary Homicide Reports for 2022 puts the black victimisation rate at 29.0 per 100,000 against 4.2 for white Americans. The two sets of figures differ because they use different years and different source files. They point the same way.

Both numbers come out of the same data system as Table 6. A reading of that system which reports the offender column and omits the victim rate is not being defeated by a lack of data.

On 1 January 2021 the FBI retired the Summary Reporting System and accepted only the National Incident-Based Reporting System. Agencies that had not converted stopped appearing. Coverage fell from roughly 95% of the population in a normal year to about two thirds, and the departments that dropped out included some of the largest in the country.

Anything that splices a pre-2021 count to a post-2021 count and reports the difference as a change in violence is measuring the transition. The FBI now publishes modelled estimates with confidence intervals for these years, which is a different kind of number from a count, and the distinction has to be carried with the figure.

FBI hate crime reporting is voluntary. Agencies may submit, may submit a zero, or may submit nothing. In 2023, 16,009 agencies participated, covering about 95% of the population; in 2024, 16,419 agencies covering about 95%. Well over two thousand agencies submit nothing at all, and a large share of those that do participate report zero incidents for the year.

This makes the headline total move with participation. The FBI publishes a separate like-for-like trend using only agencies that reported at least six comparable months in both years, and on that basis incidents fell slightly from 2022 to 2023 and again from 2023 to 2024, in the opposite direction from the "record high" framing attached to the all-agency total in the same release.

For scale, the National Crime Victimization Survey estimated an average of roughly 247,000 hate crime victimisations a year between 2005 and 2019, against FBI totals in the low tens of thousands. The gap is the reporting gap, and it means the FBI series describes what reaches a participating police department, not how much bias-motivated crime occurs.

Britain is the exception in Europe. Under section 95 of the Criminal Justice Act 1991 the Ministry of Justice has a statutory duty to publish statistics on differences between ethnic groups across the justice system, and the data exists at every stage from stop and search to prison.

The department attaches a warning to it that is more important than most of the figures. Its comparisons carry no controls for income, geography, offence mix or offender history, so the differences they show cannot be read as effects of ethnicity and are not evidence of bias. That caveat is in the publication. It is almost never in the post.

Stop and search is where the denominator does the most work. The standard disparity ratio divides a group's search rate by the white search rate using resident population from the census. Searches are heavily concentrated in a small number of urban areas whose ethnic composition differs sharply from the national average, so a national resident denominator assumes an evenness that does not exist. Substituting an available population, meaning people actually present where and when searches happen, reduces the ratio substantially. Neither denominator settles whether the disparity reflects bias, because a search is a police action and the statistic measures who gets searched rather than who was carrying anything.

This is the clearest case in the article of a number that failed and a public argument that did not update when it did.

The claim is that 84% of group-based child sexual exploitation offenders are Asian. It originates in a 2017 report by the Quilliam Foundation, built from cases collected out of media coverage rather than from any official record. As a sample it cannot support a national proportion, and it was criticised on exactly that ground.

What the 2020 Home Office paper on group-based CSE actually says is narrower than either side usually reports. Its words: some studies have indicated an over-representation of Asian and black offenders, however it is difficult to draw conclusions about the ethnicity of offenders as existing research is limited and data collection is poor. It also records that research has found group-based CSE offenders are most commonly white, and immediately adds that it is not possible to conclude this is representative of all such offending, because of data quality, sample selection and the way ethnicity is recorded.

That paper is frequently quoted as having found no credible evidence that any ethnic group is over-represented. That sentence is not in it. It belongs to one of the academics whose work the paper cites, and the paper's own position is the weaker one above: the research cannot answer the question either way.

The position moved in 2025. Baroness Casey's National Audit on Group-Based Child Sexual Exploitation and Abuse, published 16 June 2025, found that ethnicity is not recorded for two thirds of perpetrators and that the national data is not good enough to support any statement about the ethnicity of group-based CSE offenders at national level. In local data from three police force areas, Greater Manchester, West Yorkshire and South Yorkshire, the audit found clear evidence of over-representation among suspects of Asian and Pakistani-heritage men. Casey also wrote that claims most offenders were white were at best misleading, and that flawed data had been used to dismiss concerns as sensationalised.

Both of the positions people argue from are now out of date. The 84% figure has no source that can carry it. So does any national claim in the other direction. What exists is documented local over-representation in three force areas and an admitted national data vacuum, and the recording changes made after the audit mean the national question may become answerable in a few years rather than remaining permanently open.

Police-recorded hate crime in England and Wales roughly tripled between 2013 and 2022. The Home Office attributes that rise substantially to improvements in how forces record crime and better identification of what counts as a hate crime, and states plainly that its figures are not a good measure of prevalence and do not provide reliable trends over that period.

The Crime Survey for England and Wales, which asks victims directly and is unaffected by recording practice, shows a long-run decline over much of the same period. The two series are not in conflict. They answer different questions, and the divergence between them is the size of the recording change.

The reverse inference fails too. When a recorded figure rises because a force started logging something it had been missing, the crimes now appearing were happening before. Whether they are new and whether they are newly visible are separate questions, and recorded data can only answer the second.

The first thing to establish about any EU-wide claim comparing ethnic groups is that the official data it would need has not been collected.

France restricts the collection and processing of data revealing racial or ethnic origin by law, and its crime statistics are not broken down that way. German police statistics record nationality, not ethnicity. Sweden publishes crime figures by foreign background rather than by ethnic group. Across most member states the variable simply is not in the system.

So any statement of the form "in Europe, group X commits Y times more crime" is not coming from official European data, because official European data does not contain the category. It is coming from a proxy, and the three proxies in use are different from each other and different from ethnicity.

Nationality is legal citizenship, and it changes on naturalisation. Country of birth misses everyone born in the country to migrant parents. Migration background usually means foreign-born or having foreign-born parents, which captures a second generation but groups a Polish plumber with a Syrian refugee. Relabelling any of the three as race changes the claim into one the data was never built to support.

Where Nordic statistical offices do publish crime by migration background, they publish two versions of it, and the gap between them is the finding. The raw figures show over-representation. Adjusting for the composition of the groups, which differ sharply in age and sex from the general population and skew young where offending is concentrated, reduces the difference substantially without eliminating it. Quoting the raw figure as though it were the adjusted one, or the adjusted one as though the raw disparity were an artefact, are both misreadings of the same publication.

The European Union Agency for Fundamental Rights runs the large surveys people reach for at this point, including its work on people of African descent. Those surveys measure self-reported discrimination and victimisation. They contain no information about who commits offences. Using them as evidence about offending by minority groups is a category error rather than a disagreement about numbers.

Hate crime reporting through the OSCE covers most participating states but their legal definitions, recording thresholds and levels of police participation vary enormously. A country reporting a higher figure may have a broader definition and a better recording system. The table supports comparison of national recording outputs. It does not support a ranking of how much hate crime occurs.

The failure modes are symmetrical, and pointing at one of them is not an argument against the other.

Claims that inflate take Table 6's raw counts, present them as rates or as evidence of racial motive, and omit that most violence is within-group, that the table covers half the single-offender homicides in a file that itself covers a fraction of homicides, and that black Americans are killed at several times the white rate in the same dataset.

Claims that deny treat every disparity as a policing artefact without supplying the quantified alternative the argument requires, or extend an "insufficient data" finding into a positive finding of no difference. The Home Office CSE paper did not find there was no over-representation; it found it could not tell. Casey then found over-representation in the local data she could examine. Those are different results and neither is the other's refutation.

Claims that read recording as offending, in either direction, are the most common of all, and they are the easiest to check: look for whether a survey series moves with the recorded one.

1. The FBI table at the centre of this genre is genuine, and it is a count of single-victim, single-offender solved homicides with no motive field and no denominator. The offender was never identified in 48.9% of 2019 murder victimisations.

2. In the same American data, black Americans are killed at roughly six to seven times the white rate. That figure is a rate with a stated population base, which the offender counts are not.

3. England and Wales records ethnicity throughout its justice system and warns that its own uncontrolled comparisons cannot identify ethnicity as a cause or demonstrate bias.

4. On group-based child sexual exploitation, the 84% figure has no source that supports it, the 2020 Home Office paper concluded the research could not answer the question, and the 2025 Casey audit found the national data still cannot while local data in three force areas shows clear over-representation of Asian and Pakistani-heritage suspects.

5. No EU-wide comparison of offending by ethnicity is possible from official data, because most member states do not collect the variable. Every such claim is running on a proxy.

6. Hate crime series on both sides of the Atlantic move with participation and recording practice, by the publishing agencies' own statements.