Reading a crime number

Suppose your local paper runs the headline Violent crime jumps 20% in our area. You walk home from work after dark, or your teenage son does. Should either of you change anything?

You can't answer that yet, and neither can the paper, until you know what was counted. Lesson 1 left you with half of Americans telling a pollster that crime had risen, and a promise that this lesson would show you how to check. A crime number is never a plain count of crimes. It's what survives a series of filters: whether the victim told anyone, whether the police wrote it down, who the source could reach, how the offence was defined, which months it covers and what it was divided by. Change any one of those and the number moves, with no change in how many people were hurt. This lesson follows those filters one at a time and ends with five questions you can put to any crime figure, starting with that headline.

Two ways to count

There are two main ways to find out how much crime there is. You can count what the police recorded, or you can pick a sample of people and ask them what happened to them. Both the US and England and Wales do both, and nearly every crime figure you'll read comes from one side or the other.

In the US, the survey is the National Crime Victimization Survey (NCVS), run by the Bureau of Justice Statistics. In a typical year it interviews "about 240,000 persons in approximately 150,000 households", people aged 12 and over, and asks about nonfatal violence (rape or sexual assault, robbery, aggravated and simple assault) and household property crime, whether or not anyone told the police.¹ Because it works by interviewing victims, there is one crime it can never count. Its tables carry the footnote: "Excludes homicide because the National Crime Victimization Survey (NCVS) is based on interviews with victims."¹

The police side is the FBI's Uniform Crime Reporting programme, now built on the National Incident-Based Reporting System (NIBRS). For 2024 it drew on 16,675 police agencies covering 95.6% of the population, with estimation filling in the rest.² It counts offences known to the police, so it includes homicide, crimes against businesses and crimes against children under 12, none of which the survey reaches. The older Summary Reporting System (SRS) also had a counting rule that still trips people up: "The SRS Hierarchy Rule states only the one most serious offense within a criminal incident is counted. For example, if a robbery and a homicide occur in the same incident, SRS will only count the homicide." NIBRS records up to ten offences per incident instead.²

In England and Wales, the survey is the Crime Survey for England and Wales (CSEW), which asks people aged 16 and over, and the police side is police recorded crime. The Office for National Statistics (ONS) publishes both in one bulletin and says plainly what each is for. ONS says the CSEW is better for tracking long-term trends in violence, theft, criminal damage, fraud and computer misuse. For the rest, it writes: "Police recorded crime is not always a good indicator of general trends in crime, because of changes in how it is recorded over time. However, it can give more insight into lower-volume, higher-harm offences that are reported to the police".[7] The crimes it names for police data are homicide, offences involving weapons, and robbery.

The survey has its own blind spots, and ONS lists them: "The survey excludes crimes against commercial or public sector bodies, tourists, or those living in communal establishments like care homes, student halls of residence, and prisons."[7]

As lesson 1 said, Scotland and Northern Ireland count their crime separately and this course did not read their figures, so every crime figure in this course from outside the US is for England and Wales only.

Side by side, the four look like this. Each is correct about what it counts and silent about what it does not.

What kind of number is it?

Once you know who counted, the next question is what kind of number they produced. The NCVS bulletin defines three, and the difference between them is where most misreadings start.¹

• Victimizations: "the total number of times that persons or households were victims". Someone robbed twice is two victimizations.
• Incidents: "specific criminal acts involving one or more victims". One robbery of two friends is one incident and two victimizations.
• Prevalence: "the number or percentage of unique persons who were crime victims". Someone robbed twice counts once.

Watch what that does to one year's figures. For 2024 the NCVS put the rate of violent victimization at 23.3 per 1,000 people aged 12 and over. It also gave a prevalence: "In 2024, 1.45% of persons age 12 or older experienced at least one violent victimization".¹ Turn both into what statisticians call natural frequencies, meaning counts of people out of a round number, and the picture is plain. Out of every 1,000 people aged 12 and over, about 15 were victims of violence at least once that year, and between them they had about 23 victimizations. The gap is people who were victims more than once, and lesson 3 comes back to them.

Then there is the denominator, the "per how many". Each source picks its own:

• the NCVS: per 1,000 persons aged 12 or older (or per 1,000 households for property crime);
• the FBI: per 100,000 inhabitants;
• ONS homicide figures: per million people;
• CSEW prevalence: a percentage of adults aged 16 and over.

None of these is wrong. They just can't be laid side by side until you convert them, and even then only if they count the same thing.

The last label is the period, and it's slipperier than it looks. The NCVS for 2024 does not cover calendar 2024. People interviewed during 2024 were asked about the six months before the month of their interview, so the 2024 survey covers crimes from 1 July 2023 to 30 November 2024, and BJS files each crime under the year of the survey, not the year it happened.¹ The CSEW works in years ending in March, and it too asks about the 12 months before each interview.[7] So "2024" from one source and "the year ending March 2024" from another are not the same stretch of time.

Three correct numbers for one year

The clearest example comes from a short brief BJS published in March 2026, The Nation's Two Crime Measures, 2015–2024, which sets the survey and the police data side by side.³ To line the two up, the brief leaves simple assault out of the survey's violent crime and uses three offences, rape or sexual assault, robbery and aggravated assault, while the police figure covers homicide, rape, robbery and aggravated assault.

The police rate here is 3.7 per 1,000, where the FBI figure you converted earlier came to about 3.6. The brief uses the NIBRS Estimation Program's summary estimate, a different estimate from the FBI's headline, built from the same police records.³ The FBI publishes more than one estimate from its own data, and says of them: "Both collections are estimates and should be viewed as analogous measures, rather than as one being more accurate than the other."²

Take them apart one filter at a time.

The three US rates of violent crime excluding simple assault, 2024, on one scale. The red lines are the 95% confidence intervals BJS publishes with each figure, the kind you met in Habits and Self-Discipline lesson 5; the police figure's is too narrow to draw. Source: BJS, The Nation's Two Crime Measures, 2015–2024, appendix table 1.³

From 8.9 to 5.3: the reporting filter. Both numbers come from the same interviews, the same people and the same offences. The only difference is that 5.3 keeps just the crimes the victim said had reached the police. That's about 60% of 8.9, and the NCVS bulletin gives the same share for this group of crimes in 2024: 60.2% of violent crime excluding simple assault was reported.¹ So the first filter is people deciding, for all sorts of reasons, whether to tell the police.

From 5.3 to 3.7: the source filter. Now you're comparing a survey with police data, and several things change at once. Start with the denominator. The survey divides by people aged 12 and over. The police rate divides by all US residents.

That's not the whole story, because the scope pushes the other way. BJS notes that the police figure includes crimes against businesses and against people aged 11 or younger, homeless people and people in institutions, none of which the survey covers, and it includes homicide.³ Those push the police rate up. So the denominator accounts for some of the gap, the wider scope works against it, and the brief doesn't say how the two net out.

The rest of the gap could come from several places. BJS itself says the two collections use different methods and measure offences that are similar but not identical.³ A crime a victim says was reported still has to be recorded by the police before it's counted, and in England and Wales an inspection found that often it was not (two sections on); the sources here do not say how often that happens in the US. And the survey's figure is an estimate from a sample: the interval around 5.3 runs from about 4.1 to 6.6.³

The survey's intervals in the chart are wide because it interviews a sample of people. The police figure's is tiny; it is an estimate too, but built from agencies' records rather than a sample. Narrow does not mean closer to the truth about crime: it means a precise count of what reached the police and was written down. The FBI says of the survey and its own data that "each program has unique strengths. Together, they can provide a greater understanding of crime trends in the United States."²

Most crime never reaches the police

The first filter is the biggest, and it's the one people forget. In the US survey for 2024, "Approximately 48% of violent victimizations were reported to police in 2024", and for property crime the share was 30%.¹ As natural frequencies: of every ten violent crimes the survey found, about five reached the police; of every ten property crimes, three.

England and Wales measure it differently: across what ONS calls CSEW headline crime, its name for the main set of crimes the survey counts, rather than violence alone. ONS says "only an estimated 36% of CSEW headline crime incidents (excluding fraud and computer misuse) were reported to the police in the CSEW YE March 2026."[7] Roughly one in three. The two figures cover different crimes, so they do not compare the countries; what they share is that in both, much crime never reaches the police.

Fraud is further out still. Lesson 1 met the Federal Trade Commission's 3 million US fraud reports for 2025. Reports are the small part: the FTC's own testimony cites a study finding "only 4.8 percent of victims of mass-market consumer fraud complained to a Better Business Bureau or a government entity".[10] A Better Business Bureau is a private US complaints body. So about one victim in twenty told anyone official. In England and Wales, the survey estimated that 7.8% of adults were victims of fraud in the year to March 2026, and fraud is one of the crimes for which ONS calls the survey its preferred source.[7]

The US bulletin names some of the reasons victims may not report, among them fear of reprisal, not wanting to get the offender in trouble, believing the police could not or would not help, and thinking the matter too personal or too trivial.¹ None of those is a mistake by the victim. They are the reason a police count cannot tell you how much crime there is, only how much reached the police. So if your local headline is a police count, part of what it measures is how many people in your area chose to tell the police.

When a count rises and crime does not

The second filter is recording. A crime reported to the police still has to be written down as a crime before it appears in any figure, and whether it is depends on rules and on how closely forces follow them.

In 2014 Her Majesty's Inspectorate of Constabulary (HMIC, as it was then called) audited police recording in England and Wales and found this: "Over 800,000 crimes reported to the police have gone unrecorded each year. This represents an under-recording of 19 percent. The problem is greatest for victims of violence against the person and sexual offences, where the under-recording rates are 33 percent and 26 percent respectively."[8] As natural frequencies, about one reported violent crime in three was not making it into the count. Earlier that year, before the report came out, police recorded crime had already lost its accreditation as official statistics. ONS writes: "Their accreditation was removed by the Office for Statistics Regulation (OSR) in January 2014 following an assessment that found evidence that the quality and consistency of the underlying data may not be reliable." It has not been restored; the figures are still published as official statistics, without that mark of quality.[7]

Now put that finding to work.

Walk through why the two can move apart. If forces start recording more of what's reported to them, the police count rises even if the number of attacks stays the same, because the filter has opened. ONS says this of the police count directly: "Increases in volume over the last 10 years have been largely influenced by improvements in recording standards."[7] A survey that asks people directly is not touched by what the police write down. So the honest version of the headline is something like: Police recorded more violent crime this year. The crime survey, which the statistics office prefers for violence, shows no rise. The statistics office says changes in recording have made police figures a better measure of police activity than of crime, and a 2014 inspection found a third of reported violence going unrecorded. Your local 20% could be this: the same attacks, more of them written down.

The survey is not beyond question either, and its weaknesses are public. The regulator suspended the CSEW's accredited status in July 2022, citing quality concerns "including a drop in the survey response rate", and after a review wrote: "In 2024, we reinstated the accredited official statistics status of the CSEW estimates."[9] In April 2026 it answered claims it had seen about the survey. As OSR puts them, critics had said the national estimates don't match what people experience in their own lives, that the sample undercounts cities, and that "Police recorded crime figures are inherently more accurate than CSEW estimates". OSR answered that a national estimate cannot mirror every person's experience, which does not make it wrong; that the sample is drawn at random from households across England and Wales; and that the third claim is a misunderstanding, because the two measure different things.[9] Each source is good at something, and neither is the real number for everything.

The questions under a series can change too. In the year to March 2026 the survey found 40.9% of adults had experienced or witnessed anti-social behaviour, a significant rise that ONS says was partly real but "inflated because of changes to how questions in the CSEW were asked from April 2025."[7] In the US, the NCVS ran its old and new questionnaires side by side in 2024, each on about half the sample, and the 2024 bulletin reports only the old half, so its estimates are less precise than usual. BJS says the split lets it keep comparing years across the change. Even so, if a US figure for 2025 or later looks like a sudden jump, check which questionnaire it came from and whether BJS itself calls the change significant.¹

The hardest numbers to count

Rarer and more private harms have the figures most often misread, in two ways.

Rough numbers for rare crimes. Two rows from the NCVS bulletin's table of crimes reported to police show the first trap. For rape or sexual assault, the share reported was 46% in 2023 and 24% in 2024. For robbery it went the other way, from 42% to 73%.¹ BJS marks both changes as statistically significant, the kind of test First Aid and CPR lesson 3 described: each change is bigger than the survey's sampling error alone would usually produce, so neither is just noise. But each year's figure rests on a small number of interviews, and BJS's own margin around the 2024 rape figure runs many points either way.¹ So the fall is BJS's finding, how big it really was is very uncertain, and 24% on its own is not a reporting rate anyone should quote.

A panel of the National Research Council, reporting on the NCVS in 2014, explained why rape figures are especially hard for this survey. It concluded that the survey "does not measure the low incidence events of rape and sexual assault with the precision needed for policy and research purposes. Comparisons across subgroups and years are particularly problematic."⁵ It also found that the way the survey handles people who report many similar incidents lets a few such reports move the national estimates of rape a long way from one year to the next.⁵ The panel was writing about how much rape occurs, not the share reported, but its warning about comparing years carries across.

So the figure to hold on to is the one across all violence: in the 2024 survey, about half of US violent victimizations reached the police, and under a third of property crime.¹

Lifetime against annual. The US Centers for Disease Control and Prevention (CDC) publishes this on its intimate partner violence page: "More than 1 in 3 women (nearly 43.5 million) and more than 1 in 6 men (20.7 million) experienced contact sexual violence, physical violence, and/or stalking by an intimate partner during their lifetimes."[6] The NCVS, for 2024, put intimate partner violence at 2.7 victimizations per 1,000 people aged 12 and over.¹

Those can look like a contradiction, and they are not. The CDC figure is a lifetime prevalence: the share of people who have ever experienced any of a broad set of harms, including stalking, over their whole lives so far. The NCVS figure is a one-year rate of victimizations, for a narrower set of crimes, from a survey that asks about crime. You cannot subtract one from the other, and you can't use the pair to say whether anything is rising or falling.

The US argument since 2020

Whether US crime rose or fell after 2020 became a political argument, and the two measures you've met are at the centre of it. This course teaches how to read them and takes no side on whose account was right.

The two BJS sources show this. The NCVS rate of all violent victimization was 16.4 per 1,000 in 2020 and 23.3 in 2024, higher than 2020 and 2021 and similar to 2022 and 2023.¹ In the Two Crime Measures brief, the survey's narrower measure rose from 5.6 per 1,000 in 2021 to 9.8 in 2022 and has not changed significantly since, while the police-data rate stayed between 3.7 and 4.0 per 1,000 across the whole decade.³ The FBI's figures also showed murder falling an estimated 14.9% from 2023 to 2024.²

The Council on Criminal Justice, which describes itself as independent and nonpartisan, examined the gap in its report When Crime Statistics Diverge. Its key finding is about the long run: "The nation’s two major tools for measuring crime, the UCR and the NCVS, show similar long-term trends for rates of serious violent crime, motor vehicle theft, and burglary." It adds that "Much of the difference between the two sources stems from variations in methodology, rather than inaccuracies in the data." It also records that the gap between the two sources' estimates for 2021 to 2022, a 2% fall in the FBI's figures against a 75% rise in the survey's, "became a heated point of political contention", and that the argument intensified in autumn 2024, when the FBI revised earlier figures "without explanation".⁴

The honest summary: over the long run the two sources agree; over a few recent years they have pointed different ways, and that gap became a political argument; and both agencies describe the two as different tools with different strengths, not as a true figure and a false one.²³⁴ As lesson 1 showed, how people read the national trend also follows their politics, for supporters of both parties.[11]

So if one measure next year shows violence up and the other flat, and someone says the flat one is rigged, rule out the ordinary things first: different offences, periods or denominators, a change that isn't significant, a new questionnaire, or a change in recording. That shows what evidence would settle the claim, not that it is wrong.

Checking a headline yourself

Everything in this lesson comes down to five questions. Ask them of any crime figure before you let it change what you do.

1. Who counted? A police count or a victim survey, and which one.
2. What did they count? Which offences, defined how, and whether it's victimizations, incidents, offences or people.
3. Whose harm, and where? Which population and which geography: people 12 and over in US households, adults 16 and over in England and Wales, all residents, a named city.
4. When? Which period, and whether it is one year, a lifetime or a comparison between years.
5. Per how many? The denominator, and whether the figure is a count, a rate or a percentage.

If a figure cannot answer all five, it can't carry much weight. If it compares two years, ask the extra question from the last two sections: did the recording, the questions or the survey change in between?

You can check a claim yourself with the sources' own tools. For the US, BJS runs the NCVS Dashboard, which gives rates by crime, year and group, and the FBI runs the Crime Data Explorer. For England and Wales, ONS publishes Crime in England and Wales every quarter. BJS and the FBI publish each autumn, so a number in this lesson may already have a newer edition.

Now go back to the headline you started with, Violent crime jumps 20% in our area, and put the five questions to it. Who counted? If it's a police count, did the force change how it records anything? What was counted: which offences, and are they the ones you'd meet on that walk home? Whose harm, and where: the whole area, or the streets you use? When: one year against the last, or a run of years? Per how many: is 20% a rise from a small number, say from 50 recorded offences to 60, where one busy weekend could do it?

If the headline can't answer those, it isn't on its own a reason to change your walk home, or your son's. If it's a police count rising over several years in the offences you'd actually meet on that walk, it's worth reading lessons 4 to 7 with it in mind. That reading is sensible, untested, in this course's labels: it tells you how much weight a number can bear, not what to do. What you might sensibly change, and what the evidence says about each option, is the subject of those lessons.

What people get wrong

"Police figures are the real numbers." They are a precise count of crime that was reported and recorded, which misses most crime. The UK's statistics regulator lists a version of this as a misunderstanding.[9]

"Survey figures are just opinion." A victim survey asks people what happened to them, not what they think about crime, and it catches crime nobody reported. Its weaknesses are sampling error and response rates, which are measured and published.¹[9]

"Recorded crime went up, so crime went up." A change in recording can move the count with no change in crime. HMIC's 2014 audit found a third of reported violence going unrecorded, and ONS now calls police figures a better indicator of police activity than of crime trends.[7][8]

"One year's figure is the trend." A single year can mislead. A change can come from a new question or questionnaire (the anti-social behaviour figure, the NCVS redesign), and even a change that passes a significance test can rest on so few interviews that its size is very uncertain (the rape and robbery reporting rows).¹⁵[7]

"A lifetime figure and an annual figure are the same kind of number." One is the share of people who have ever experienced something; the other is how often it happened in one year. They cannot be compared to find a trend.¹[6]

"The two US measures contradict each other, so neither can be trusted." They measure different things, agree over the long run, and each is the better source for some questions.²³⁴

Practice

Connections

Lesson 1 showed the gap between how much crime people think there is and what the sources count; this lesson gave you the tools to check that gap for yourself.

From Logic and Argument, lesson 9 taught you to ask "share of what?", and every rate here is that question with a crime attached. Lesson 5 gave you the questions to put to any survey (how many, chosen how, asked what, compared with what), and the NCVS and the CSEW are surveys. Lesson 6's natural frequencies turned 23.3 per 1,000 and 1.45% into counts of people you can picture.

Lesson 3 uses every label from this lesson on the question most people care about: who is harmed, by whom, where and when. You'll see that "most violence is by strangers" and "most violence is by people you know" each come out true or false depending on which source, which crime and whose harm. Later on the Core, Statistics for Citizens takes sampling error and confidence intervals further than this course needs to.

Go deeper

• BJS, The Nation's Two Crime Measures, 2015–2024 (March 2026). A short brief, read in full for this lesson. The side-by-side charts and appendix tables are the best single place to see the two US systems move over a decade.
• Council on Criminal Justice, When Crime Statistics Diverge. Read in full for this course. Why the US measures differ, and how the gap became a political fight.
• ONS, Crime in England and Wales: year ending March 2026. Read in full for this course. Its opening sections say which source ONS prefers for which crime, and why, in plain language.

Sources

1. Bureau of Justice Statistics, S. N. Tapp and E. J. Coen, Criminal Victimization, 2024, NCJ 310547, September 2025. Read: in full (37 pages). What the NCVS counts; the sample; the definitions of victimizations, incidents and prevalence; the 2024 rate and prevalence; the reference period; table 4 on reporting to police and appendix table 5 for its standard errors; the reasons for not reporting; intimate partner violence rate; the 2024 split-sample redesign.
2. Federal Bureau of Investigation, UCR Summary of Reported Crimes in the Nation, 2024, released August 2025, and the companion Reported Crimes in the Nation 2024 FAQs. Read: both in full. Coverage, the 2024 violent crime rate, the change in murder, the hierarchy rule, the statement that its own two estimate series are analogous measures, and its statement on how its data and the NCVS differ.
3. Bureau of Justice Statistics, L. Remrey and E. Tinney, The Nation's Two Crime Measures, 2015–2024, NCJ 310700, March 2026. Read: in full. The 8.9, 5.3 and 3.7 figures with their intervals (appendix table 1), the notes on denominators and scope, and the 2015 to 2024 series.
4. Council on Criminal Justice, When Crime Statistics Diverge,
   2025. Read: in full.
5. National Research Council, Estimating the Incidence of Rape and Sexual Assault, National Academies Press, 2014. Read: summary chapter in full. Conclusions 7-1 and 7-2.
6. Centers for Disease Control and Prevention, "About Intimate Partner Violence", Internet Archive capture 20260921054056 of cdc.gov/intimate-partner-violence/about. Read: page in full; the underlying survey data brief was not read. The lifetime figures.
7. Office for National Statistics, Crime in England and Wales: year ending March 2026, 23 July 2026. Read: in full. Which source ONS prefers for which crime; what the survey excludes; the 36% reporting figure; fraud and violence counts; the 1995 and 2016 comparisons; anti-social behaviour; the loss of accreditation; the note that recording changes make police figures a better indicator of police activity than of crime trends.
8. Her Majesty's Inspectorate of Constabulary, Crime-recording: making the victim count, November 2014. Read: summary and conclusions (pages 18 to 21) and the contents.
9. Office for Statistics Regulation, Statement on statistics from the Crime Survey for England and Wales, 2 April 2026. Read: in full. The three misunderstandings, the 2022 suspension and the 2024 reinstatement.
10. Federal Trade Commission, prepared statement to the Joint Economic Committee, 25 March 2026. Read: pages 1 to 3. The 2025 reports and losses, and the 4.8% complaint figure from footnote 6.
11. Gallup, "Crime in U.S. Seen as Less Serious for Second Straight Year", October 2025 poll. Read: in full. The partisan pattern in perceptions of crime.
12. Support lines in the content note, each read on its service's own page on 25 September 2026: Rape Crisis England and Wales, Victim Support, Rape Crisis Scotland, Victim Support Scotland, Rape Crisis Northern Ireland, Victim Support NI, RAINN and VictimConnect. Read: each page in full, 25 September 2026.