Measuring crime: police statistics, victim surveys and the dark figure

Linh Tran arrives at the second class with a printout from the Port Ardent crime analyst's office and a puzzle. The city's police-recorded break-ins fell by 9% last year, but in the community survey the service ran, more residents said they had been victims of a break-in than the year before. Which number is true? Professor Desjardins says both may be, and that the gap between them is one of the oldest problems in criminology. Every measure of crime captures some of it and misses some of it. A good analyst knows what each source can and cannot show, and never treats a single number as the whole truth.

The crime funnel and the dark figure

For a crime to appear in police statistics, a long chain of events has to occur. Someone has to notice that a crime happened, which is not always obvious: many frauds and thefts go undetected. Someone, usually the victim, has to decide to tell the police. The police have to record the report as a crime, which requires them to accept that an offense occurred and classify it. Only then does it enter the statistics. At each link some incidents drop out. The crime that never enters official statistics is called the dark figure of crime, a phrase that goes back to nineteenth-century European statisticians.

The size of the dark figure varies enormously by offense. Almost every homicide becomes known to police, because a body is hard to hide and the death must be explained. Most motor vehicle thefts are reported, because insurers require a police report. At the other extreme, the great majority of sexual assaults, much shoplifting, most minor assaults between acquaintances and a large share of online fraud never reach police. Victims stay silent because they think the incident was too minor, they believe police could do nothing, they fear the offender or the process, they handled it another way, or they distrust the police. For some groups, including many Indigenous, Black and newcomer communities, distrust is a significant reason.

Police-reported statistics

The main source of police-reported crime data in the United States is the FBI's Uniform Crime Reporting program, which began in 1930. For decades it collected summary counts of a few serious offenses, but since 2021 it has relied on the National Incident-Based Reporting System, NIBRS, which records details of each incident, including all offenses committed, victims, offenders, weapons, property and arrests. Participation is voluntary, and the FBI estimates national figures from the agencies that report.

Police data have real strengths. They are collected continuously, cover the whole country, include details on accused persons and outcomes, and can be broken down to small areas such as neighborhoods and street segments, which is essential for police deployment. Their weaknesses follow from the funnel. They count only what is reported and recorded. They can change because of police activity rather than crime: a crackdown on impaired driving or drug dealing produces more recorded offenses, because these crimes are mostly discovered by police rather than reported by victims. Changes in recording rules or computer systems can create jumps that have nothing to do with behavior.

A similar problem affects any crime rate that counts every offense equally, so analysts look closely at serious violent crime separately. The FBI's national figures for 2025, released in August 2026, showed murder down 18.1% from 2024 to roughly 4.1 per 100,000 people, the lowest rate in decades, with violent crime down 9.3% and property crime down 12.4%. Because NIBRS records every offense in an incident, rather than only the most serious, its counts are not directly comparable with older summary statistics, and analysts must be careful when comparing years across the changeover.

Victimization surveys

Victimization surveys ask a random sample of people whether they have been victims of crime in a recent period, whether they reported it and why or why not. Because they ask victims directly, they capture crime that never reached police, and they allow researchers to estimate the dark figure. The Bureau of Justice Statistics' National Crime Victimization Survey, the NCVS, interviews about 240,000 people aged 12 and older in about 150,000 households every year. In 2024, almost half of violent victimizations were reported to police, from 24% of rapes and sexual assaults to about 70% of robberies and aggravated assaults. Preliminary results for 2025 from the redesigned survey, released in October 2026, found serious violent victimization down 25% and property victimization down 19% from 2024.

Victimization surveys have their own weaknesses. They depend on people remembering incidents accurately and placing them in the right time period. Respondents may not recognize an incident as a crime, or may not wish to disclose it, especially if the offender is a family member. Surveys cannot ask homicide victims, and they usually exclude young children, people in institutions and people experiencing homelessness, some of whom face very high risks. They do not cover crimes without an individual victim, such as drug trafficking, and they are expensive, so samples are often too small to give reliable figures for individual cities or neighborhoods. Finally, a survey result is an estimate with a margin of error.

Self-report studies

A third method asks people about crimes they have committed. Self-report studies, usually anonymous questionnaires given to students or young adults, began in the 1940s and 1950s. They produced a surprising finding: most young people admit to some offending, such as minor theft, fighting, vandalism or underage drinking, but only a small share are ever caught. Offending was far more widespread across social classes than police records suggested, while serious and frequent offending was concentrated in a much smaller group. Self-report data also allow researchers to study the causes of offending, because the same survey can ask about family, school, friends and attitudes.

Self-report surveys work best for common, less serious offenses and for groups that can be surveyed easily, such as students. The most serious and frequent offenders may be absent from school on survey day, may refuse to answer or may exaggerate or minimize. Researchers check validity by comparing answers with official records, and generally find that self-reports are reasonably accurate for most respondents, although accuracy varies across groups and offenses.

Estimating the dark figure

If we know how many incidents of a type police recorded and what share of such incidents victims report, we can estimate how many actually occurred. The estimate is rough: it assumes that the survey's reporting rate applies locally and that police record every incident reported to them. Linh uses the reporting rates from Port Ardent's 2024 community survey, which asked 1,200 residents about their experiences, together with police counts for 2025. The code below estimates the true number of incidents and the share of each type that never reaches police records.

# Estimating the dark figure: recorded incidents divided by the share reported
recorded = {"break-ins": 470, "motor vehicle theft": 260, "theft": 1840,
            "assault": 880, "sexual assault": 95, "fraud": 610}
reported_share = {"break-ins": 0.50, "motor vehicle theft": 0.85, "theft": 0.32,
                  "assault": 0.38, "sexual assault": 0.09, "fraud": 0.20}
total_rec = total_est = 0
for kind in recorded:
    est = recorded[kind] / reported_share[kind]
    total_rec += recorded[kind]; total_est += est
    print(f"{kind:20s} recorded {recorded[kind]:5d}  estimated {est:7.0f}  unrecorded {1 - recorded[kind] / est:5.0%}")
print(f"all six types: recorded {total_rec}, estimated {total_est:,.0f}, recorded share {total_rec / total_est:.0%}")

break-ins            recorded   470  estimated     940  unrecorded   50%
motor vehicle theft  recorded   260  estimated     306  unrecorded   15%
theft                recorded  1840  estimated    5750  unrecorded   68%
assault              recorded   880  estimated    2316  unrecorded   62%
sexual assault       recorded    95  estimated    1056  unrecorded   91%
fraud                recorded   610  estimated    3050  unrecorded   80%
all six types: recorded 4155, estimated 13,417, recorded share 31%

The estimates show why police data alone mislead. Vehicle theft is mostly reported, so police records come close to the true count. Sexual assault is mostly unreported, so the roughly 95 recorded incidents probably represent more than 1,000. Across these six offense types, police records capture only about three in ten incidents. Linh notes a second implication: a change in the reporting rate can move police statistics even if the true amount of crime stays the same. When a police service makes it easier to report sexual assault, or when public attitudes change, recorded sexual assaults may rise while actual victimization stays level or falls.

Other sources of crime data

Analysts increasingly combine the three classic sources with others. Calls for service show what the public asks police to deal with, including disorder and incidents never recorded as crimes. Hospital emergency department records capture assault injuries that victims did not report to police; a well-known program in Cardiff, Wales, shared anonymous hospital data on where and when assaults happened with police and licensing officials, and violent injuries fell compared with similar cities. Insurance claims, retail loss surveys and bank fraud records show property and financial crime that rarely reaches police, which Priya knows well from her years at the bank. Court and corrections data show what happens after charges are laid. Each source has its own gaps, but where several point the same way an analyst can be far more confident.

Data on people accused of crime need particular care. Police-reported data on accused persons describe only incidents where someone was identified, which is a minority of property crimes. They also reflect where police look. If one neighborhood is patrolled far more heavily than another, more of its residents will be stopped and charged for offenses, such as drug possession, that are equally common elsewhere. This is why criminologists check patterns in accused-person data against victimization and self-report data before drawing conclusions about who offends.

Reading crime statistics with care

Professor Desjardins gives the class five rules for reading any crime statistic. First, ask what source it comes from and what that source can capture. Second, prefer rates to counts when comparing places or years, because populations change. Third, look at trends over several years rather than a single year's change, since small numbers bounce around: a city that goes from 4 homicides to 6 has not had a 50% crime wave. Fourth, check whether definitions or recording practices changed. Fifth, look at more than one source; when police data and victimization surveys move in the same direction, the conclusion is much stronger.

These rules resolve Linh's puzzle. Police-recorded break-ins fell 9%, but the survey suggested a rise. The survey sample was small, so its estimate had a wide margin of error. At the same time, Port Ardent had launched an online reporting system for minor property crime in 2024, and some residents who reported online were not counted in the break-in category because the system filed them as thefts. The apparent contradiction came partly from chance and partly from a change in recording. Linh's recommendation to the analyst is to check how online reports were classified before anyone draws conclusions about break-ins.

Small numbers deserve special care. If a neighborhood averages 10 robberies a year, chance alone will produce years with 6 and years with 15. A useful rule of thumb for counts of rare events is that random variation is roughly the square root of the average, so a count of about 10 can easily vary by 3 or more either way without any real change. Analysts compare a year's count with this expected range before declaring a trend. The same applies to clearance rates and to rates for small groups, where one or two incidents can swing a rate dramatically.