Do Violent Offending Rates Differ by Race?

This is about offending. Not who gets victimised, not who offends against whom, and not whether violence was racially motivated. Those are three separate questions answered by three separate datasets, and swapping one for another is how most arguments in this area go wrong. A companion piece at the foot of this page deals with that problem. This one deals with rates.

Two things have to be said before any number appears. Rates of registered offending are not rates of offending: they record what was reported, investigated, cleared and classified. And a difference between groups, before or after statistical adjustment, is not a property of the groups. Every figure below carries the population it describes and whether it has been adjusted.

The Bureau of Justice Statistics has published an actual offending rate, derived from the FBI's Supplementary Homicide Reports and divided by resident population. Averaged over 1980 to 2008, the homicide offending rate was 34.4 per 100,000 for black Americans and 4.5 for white Americans. BJS describes that as almost eight times higher. Over the same period the victimisation rates were 27.8 and 4.5.

That is the headline figure people are usually reaching for, and it is real, published, and comes with its denominator attached. It also stops in 2008. No equivalent series has been published since. Anyone citing a current American homicide offending rate by race is either using this report, which is now seventeen years old, or has built the number themselves.

The series has a second problem that limits what any version of it can carry. In the FBI's 2019 expanded homicide data, of 16,245 murder offenders the race was unknown for 29.3%, and offenders were not identified at all for 48.9% of murder victims. Clearance is lower where victims are black, so the cases missing from an offender table are not missing at random, and the direction of that bias is to undercount rather than inflate.

The obvious objection to any arrest-based figure is that it measures policing. There is one study designed specifically to test that, and it is the most useful document in this whole subject.

BJS compared the race of offenders as described by victims in the National Crime Victimization Survey, who have no knowledge of whether anyone was arrested, against actual UCR arrests for the same year. If arrest disparities were being generated by police rather than by offending, the two distributions would diverge.

BJS states the result in its own words: there were no statistically significant differences by race between offenders identified in the NCVS and persons arrested per the UCR, and white and black people were arrested proportionate to their involvement in serious nonfatal violent crime.

This is the single finding in the article that most annoys both sides. It means the black share of serious violent arrests is not an artefact of differential arrest, because victims who cannot see the arrest data describe the same distribution. It also means the disparity in that arrest figure is a disparity in involvement, which is the thing the rest of this article then tries to explain.

Three limits on it. Hispanic arrestees were 21.4% against 12.0% of offenders in reported incidents, which is a genuine gap in the other direction; BJS notes victims often could not determine Hispanic origin, which may account for some of it. Offender race was not reported in 14.2% of incidents overall, and 18.0% where the victim was black. And the finding covers nonfatal violence only. It says nothing about homicide, or about any stage of the system after arrest.

American crime statistics record race and Hispanic ethnicity in separate fields, and roughly 93% of Hispanics are coded as white in the race field. That inflates white offending figures and deflates black ones.

Steffensmeier and colleagues re-estimated the national series using California and New York arrest data, which flag Hispanic ethnicity, to separate the two. The black share of homicide offending rose from 57% to 65%; the black share of assault rose from 42% to 44%; the black share of robbery fell slightly, from 57% to 54%.

Their wider conclusion is the interesting one. The apparent decline in the black share of violent offending over three decades largely disappears once Hispanics are separated out. What looked like a falling share was substantially a growing Hispanic population being counted as white.

Two studies do the work here and they reach different places, which is itself the finding.

At the individual level. Sampson, Morenoff and Raudenbush followed 2,974 people aged 8 to 25 across 180 Chicago neighbourhoods, using self-reported violence rather than police records. Raw, the odds of perpetrating violence were 85% higher for black than white participants. Controls for family and neighbourhood context explained over 60% of that gap, with parental marital status, immigrant generation and neighbourhood social context doing most of the work. A reduced gap remained after controls.

The same study found Latino violence 10% lower than white in the raw data, and the entire Latino-white difference explained by the controls. Any account that arranges groups into a simple hierarchy has to deal with that, and most do not.

At the neighbourhood level. Peterson and Krivo analysed roughly 8,900 census tracts across 87 cities. Raw violent crime rates ran 10.0 per 1,000 in black neighbourhoods against 2.0 in white ones, a ratio of about five to one. Controlling for concentrated disadvantage cut the black-white ratio from 4.27 to 1.65. Adding the disadvantage and crime of surrounding neighbourhoods cut it to 1.13, which is not statistically significant. The Latino-white ratio fell to 1.02, also not significant.

That looks like the gap disappearing, and the authors' own data explains why it should not simply be read that way. In their sample, 88.9% of white neighbourhoods have none of the extreme forms of disadvantage they measure, while 56.4% of black neighbourhoods have four or more. The distributions barely overlap. A model that controls for disadvantage is therefore comparing places that mostly do not exist against each other, and the result should be read as what it is: a statement that if the neighbourhoods were comparable the crime rates would be too, not a demonstration that they are comparable.

So the residual depends on where you measure. At the individual level a reduced but real gap remains after the strongest available controls. At the neighbourhood level it goes to non-significance, on an extrapolation the authors are open about. Neither model contains individual criminal history, exposure to prior violence, or several family and neighbourhood processes nobody has measured. "Not explained by this model" is not the same as "unexplainable", and it is not the same as "explained by race".

The Home Office publishes arrest rates per 1,000 population by ethnic group, which is more than the American sources manage.

Black people were arrested at about 2.2 times the white rate; among men, 38.2 against 16.0 per 1,000, about 2.4 times. Asian arrest rates sit below white ones, which again is awkward for a simple hierarchy. The broad categories hide wide internal ranges: within the Black group the rate runs from 13.1 for Black African to 52.4 for Black other, and within the Asian group from 2.8 for Chinese to 11.3 for Pakistani. Ethnicity was unknown for 12.7% of arrests.

For homicide, over the three years to March 2024 principal suspects convicted were 65% white, 20% black and 9% Asian, putting black suspects at roughly six times their share of the population. The same source puts black homicide victimisation at 39.8 per million against 8.5 for white, over four times higher.

The Ministry of Justice attaches this to its own statistics, and it governs every figure above: no causative links can be drawn from these summary statistics, no controls have been applied for other characteristics of ethnic groups such as average income, geography, offence mix or offender history, so it is not possible to determine what proportion of the differences are directly attributable to ethnicity, and the differences should not be taken as evidence of bias or as direct effects of ethnicity.

There is a second finding here, and it is an absence. No study for England and Wales does for offending what Sampson did for Chicago. The multivariate British work is about how the justice system processes people, not about population offending rates. So the question this article asks about the United States, how much of the gap survives controls for deprivation and neighbourhood, currently has no British answer at all. Only the raw comparison exists.

No European country in this section measures race. They measure country of birth, parents' country of birth, or citizenship. Those are different variables, they are defined differently in each country, and no figure below is comparable with any other figure below.

Sweden has the clearest published adjustment sequence. Among residents aged 15 and over, taking Swedish-born people with two Swedish-born parents as the baseline, the raw relative risk of being registered as a crime suspect over 2015 to 2018 was 2.5 for the foreign-born and 3.2 for Swedish-born people with two foreign-born parents. After standardising for age, sex, education, income and municipality type, those fell to about 1.8 and 1.7. Roughly half the excess is accounted for by the composition of the groups. The rest is not, in that model.

For serious violence the raw gaps are much larger and the adjusted ones still substantial. Two cautions on those figures: the category Brå labels attempted and completed homicide is predominantly attempted, and Brå itself identifies differential police detection as one contributor to registered-suspect differences.

Germany publishes a suspect rate per 100,000 residents. Excluding immigration offences, which only foreign nationals can commit, it ran 1,813 for Germans against 4,788 for non-Germans in the most recent year, about 2.6 times. For violent crime specifically the federal police put the non-German rate at roughly four times the German one. Their own caveat is that the statistic records police work rather than reality, and the denominator is a particular problem here, since suspects who do not live in Germany are counted in the numerator and absent from the population base.

The Netherlands recorded 61 suspects per 10,000 residents of Dutch background against 150 for those with a migration background. No matched adjusted figure is published alongside it.

Denmark publishes a crime index rather than a rate, with the male population average set at 100, and shows non-Western descendants well above that level. The reports disagree on how much adjustment the published index already contains, so the figure should not be quoted without checking which version it is.

Norway reports overrepresentation that falls substantially after adjusting for age and sex, and that has declined over time rather than grown.

Holds. Raw gaps are large and they appear in every country that measures anything. The American homicide offending gap is roughly eightfold on the last published series. The BJS test shows the serious-violence arrest gap tracks what victims describe, so it is not manufactured by policing. Nordic and German overrepresentation survives full socioeconomic adjustment at a reduced size.

Collapses. The neighbourhood-level black-white gap goes to non-significance under disadvantage and spatial controls. The entire Latino-white gap disappears under controls, and Latino violence starts below white in the raw data. Asian arrest rates in England and Wales sit below white ones. The apparent decline in the black share of American violent offending is mostly a Hispanic coding artefact. Any claim that ranks groups by a single ordering fails on at least one of these.

Neither. The residual at the individual level, after the best controls anyone has applied. It is smaller than the raw gap, it is real, and nothing in this evidence base identifies what it consists of. The models omit criminal history and much else. Reporting it as evidence of an inherent group difference is unsupported; reporting it as zero is also unsupported.

1. The last published US homicide offending rate by race is 34.4 per 100,000 for black Americans against 4.5 for white Americans, averaged over 1980 to 2008. There is no current equivalent, and nearly half of 2019 murder victimisations had no identified offender at all.

2. For serious nonfatal violence, the offender distribution described by victims matches the arrest distribution. The arrest gap reflects involvement, not differential arrest, with Hispanic arrestees the exception in the other direction.

3. Family and neighbourhood context explains over 60% of the individual-level black-white gap in the strongest US study, and the whole of the Latino-white gap, which runs the opposite way in the raw data.

4. At neighbourhood level the gap reaches non-significance under disadvantage controls, on an extrapolation between distributions that barely overlap.

5. England and Wales publishes a 2.2-fold black-white arrest gap and no study that adjusts it for anything. That absence is a finding.

6. Every European country that measures origin shows overrepresentation that roughly halves under adjustment without reaching zero. None of them measures ethnicity, and none is comparable with another.