How we count

The rules every page follows. How race was recorded across the sources is on Race and ethnicity.

Age

Age at the offense, in the same bands on every page.

Under 12 and 12 to 17 make under 18. Arrest records band ages 10 to 12 together, so their youngest band is under 13.
A missing age is age unknown, never guessed.

Biological sex: female or male.

A missing sex is sex unknown.
Sexual identity sits on top of sex, never in place of it. No source we use records it yet.

Six groups, the same on every page. Hispanic is of any race; the other five are not Hispanic, so no one is in two.

White+MENA includes Middle Eastern and North African people. AIAN is American Indian or Alaska Native. NHPI is Native Hawaiian or Other Pacific Islander.
Race unknown means no race was recorded. A bar's lighter part is people police gave a race but no ethnicity, placed as Hispanic or not by CrimeDB.
Where a source never crossed race with Hispanic origin, or recorded it for too few people (arrests before 2021, death row before 1997, the homicide reports before 2017), each race keeps its Hispanic people: White+MENA, Hispanic.
Traffic deaths and federal sentences count everyone recorded Hispanic as Hispanic, whatever the race; people with a race and Hispanic origin unknown keep their race, and only their Hispanic origin is placed by CrimeDB: the lighter part.
Death row and death sentences from 1997 until BJS crossed the two are split into Hispanic and not by CrimeDB, scaled to BJS's own count of Hispanic people: the lighter part.

The victim's relationship to the offender, as police recorded it.

Family violence is a spouse or partner, current or former, or a family or household member. It filters relationship; it is not its own group.
Stranger: the victim did not know the offender. Relationship unknown: an offender was described, but not how the victim knew them. No offender identified: the record links the victim to no offender.
A victim who was also an offender, as in a mutual fight, is its own group.
The homicide reports record the first victim's relationship only; other victims in the same record are relationship unknown.

Each person counts once in each row they fit, so rows can add up to more than the total.

A victim of rape and robbery in one incident counts once in rape, once in robbery, and once in violent crime.
A victim with two offenders counts once in each row the offenders fit: once in White men if both are White men, once in each row if one is White and one is Black.
A victim's relationship is the closest of their offenders' (an intimate partner, then family, then someone known, then a stranger), so relationship rows add up to the total.
The Risk model's tables of who hurts whom still count pairs (one victim and one offender each) until they are rebuilt once per row.

Unknown is always its own row, and every share reads against everyone.

A share divides by the full count, unknowns included, so it never reads as if every person's trait were recorded.
A rate for a group divides by the people in that group, and nobody is counted in an unknown group. So when a trait is often unrecorded, every group's rate is a floor: the true rate is at least that high.
Pages say "at least" when 5% or more of the people have the trait unrecorded.
CrimeDB never fills in an unknown. An estimate, when we make one, is labeled and kept beside the recorded value.

Every count is shown, however small.

No count is hidden for being small, for adults or children, at any level: the nation, states, counties, cities, towns, and police agencies. Counts carry no names.
Zero is never hidden.
A rate on a count under 20 is still shown, in grey, with a note that it rests on few cases and may swing a lot.
A number a source withheld or coded as not reported is never shown, filled in, or worked out. A sum that leaves it out reads "at least" or "Not reported", with no rate.

When a figure is missing, the page says why.

Withheld, named for the publisher ("FBI withheld", "CDC hidden"): the publisher printed a mark instead of a number, such as an FBI "X" or a school's reserve code. We keep the mark as printed. CrimeDB hides nothing itself, and shows nothing the publisher withheld.
Not recorded: the record exists, but this field was left blank or coded unknown.
Grey rate: the count under it is under 20, so the rate may swing a lot from year to year or place to place.

Pages call it the rate: how many for every 100,000 people, or every million for rare events. It always sits beside its count, never in place of it.

"1 in 4,100" is the same rate turned around: one person in so many. Big numbers are shortened, so 1 in 1.6M is one in 1.6 million.
Crime reported to police divides by the people living where the reporting agencies police. An agency that did not report adds neither offenses nor people.
Rates for victims and offenders by age, sex, or race divide by the Census Bureau's count of that group, scaled to the share of people the reporting agencies cover. They assume those agencies serve the nation's mix of people.
Other topics divide by every U.S. resident on July 1 of the year.
A topic that counts everyone in the nation, such as traffic deaths, divides each race group by the Census Bureau's count of that group in the 50 states and DC. A group a source splits between its own code and a joined code has no rate, since its count misses people its Census group holds.
Every rate carries a 95% range from the size of its count. A rate on a small count shows grey, with its range.
Some counts have no fitting population, such as fraud reports, which count the people who chose to file. Those show no rate, and the page says why.

Police figures count only the agencies that sent records to the FBI.

Agencies choose whether to report, and which years. A count covers the reporting agencies only, and the page says how many people they police.
Victims, offenders, places, and hours come from incident records (NIBRS), which fewer agencies send than offense counts. Those breakdowns never claim to split the main count.
2021 stands alone: the FBI took incident records only that year, and many large agencies sent none. It is never joined to other years in a trend.