Measuring Crime and the Dark Figure

Nearly every public argument about crime rests on a number, and nearly every one of those numbers means something other than what it is taken to mean.

This lesson is 7.4.4 applied to the most politically charged statistic in the field. The crime rate is not a measure of crime. It is a measure of a long social process, of which the crime is the first step and the least reliably recorded — and understanding the gap between the two is the difference between reading a crime statistic and being deceived by one.

One assault, and the five gates it has to pass.

A man is punched in a pub. What has to happen for this to become a recorded crime?

Gate one: it has to be perceived as a crime. The man has to define being punched as a criminal assault rather than as a fight he was part of, a private matter, or something to be settled between them. A great deal of what the law calls crime is not experienced as crime by the people involved — which removes it from the statistics before anything else happens.

Gate two: it has to be reported. He has to decide to tell the police. And most crime is not reported. He may think it too trivial, may not want the trouble, may distrust the police, may fear the offender, may blame himself, may think nothing will come of it, or may — in the case of many offences — not want anyone to know. The decision to report is where the largest share of crime is lost, and it is not lost randomly.

Gate three: it has to be recorded. The police, told of it, have to write it down as a crime. And they do not always (see 7.4.4 on the UK statistics scandal — around one in five reported crimes going unrecorded). They may judge it not to meet the threshold, may reclassify it downward, may resolve it informally, may be under pressure to keep figures low. Recording is a discretionary act, and the discretion is patterned.

Gate four: it has to be detected — a suspect identified — for it to enter the statistics on offenders at all. Most recorded crimes are never cleared.

Gate five: it has to be prosecuted, convicted and sentenced for the person to enter the statistics on convictions and punishment.

At each gate, cases are lost, and they are not lost at random.

Now the point. A "crime rate" built on recorded crime is a measure of what survived gates one, two and three. A "conviction rate" is a measure of what survived all five. And each gate filters by the same things: the seriousness of the offence, the relationship between the parties, the trust between the public and the police, the pressures on the recording agency, and — repeatedly — the social position of everyone involved.

The crime rate is the output of that machine. It rises and falls with changes in behaviour, yes — and also with changes in perception, reporting, recording, detection and prosecution, none of which is behaviour. When the number moves, the first question is always: which gate moved?

The dark figure: the crime that never becomes a statistic, and it is most of it.

The dark figure of crime is the gap between crimes committed and crimes recorded (see 7.4.4). It is not a small correction. For most offence types it is the majority of the crime, and its size varies enormously by offence in ways that systematically distort the picture.

Vehicle theft has a small dark figure — reported at very high rates, because insurance requires it. So it is over-represented in the statistics relative to how common it is.

Sexual offences and domestic abuse have enormous dark figures — reported at low rates, for reasons of shame, fear, distrust and the relationship to the offender. So they are massively under-represented, and a picture of crime built from recorded statistics understates exactly the offences that do the most harm to the most vulnerable.

Which produces a systematic distortion: the recorded-crime picture over-weights the visible, the stranger-perpetrated, the insurance-driven, and under-weights the intimate, the shameful, and the offences of the powerful (see 10.2.3). The statistics are not a random sample of crime. They are a biased sample, biased in a known direction — which is 7.3.2's survivorship problem, in the crime figures.

So the founding question of criminal measurement is: how do you measure the crime that never entered the record? And the answer, which transformed the field, was to stop asking the agencies and start asking the population.

The two measures, and why they disagree

Police recorded crime and the victimisation survey — what each is and what each misses.

Police recorded crime counts offences recorded by the police. Its strengths: it covers the whole country, is available over long periods, and includes offences (like homicide) that surveys cannot measure. Its weakness is the five gates — it measures processing, not crime, and it changes when recording practice changes (see 7.4.4). A "crime wave" in recorded statistics is frequently a recording change — a new counting rule, a crackdown, a drive to record more of a previously ignored offence.

The victimisation survey asks a large representative sample of the population what has actually happened to them, whether or not they reported it (see 7.4.4). This is the instrument built to measure the dark figure. Its strengths: it captures unreported crime, it is not affected by recording practice, and — because the methodology is stable — it gives a more reliable measure of trends for the offences it covers.

Its weaknesses are equally specific. It misses crimes with no individual victim to survey (drug offences, much fraud, environmental and corporate crime). It misses the most serious crimes by definition — you cannot survey homicide victims. It misses crimes against people not in the sampling frame (the homeless, those in institutions, children below the survey age — which historically excluded much crime against children). And it is subject to the recall and social-desirability problems of all surveys (see 7.4.1) — people forget, telescope, or will not disclose the most sensitive victimisation even to a survey.

And here is the finding that makes both essential: they can move in opposite directions.

Recorded crime can rise while survey crime falls — because a rise in reporting or recording (more people willing to report domestic abuse, a police drive to record it properly) increases the recorded figure even as the underlying behaviour declines. This has happened repeatedly for sexual and violent offences, and it produces headlines about "soaring crime" that describe soaring recording.

And recorded crime can fall while survey crime is stable — because of a recording change in the other direction, or a collapse in public willingness to report.

Which is why no serious analysis relies on one measure. The two are triangulated (see 7.8.3): where they agree, confidence is high; where they diverge, the divergence itself is the finding — it tells you that reporting or recording changed, which is worth knowing.

The third source, and why it is the anchor.

Homicide is the measure that compares across countries and over time better than any other (see 7.4.4), for a specific reason: a body is difficult to conceal, and almost every country investigates unexplained deaths. So the dark figure for homicide is small, and the recording is relatively consistent.

Which makes homicide the anchor of comparative crime research. When you want to know whether one country is more violent than another, or whether violence has risen or fallen over a century, homicide is the series to use — because it is the one least distorted by the gates. Almost every other cross-national "crime rate" is substantially a comparison of legal definitions, reporting cultures and recording practices rather than of crime (see 7.2.2 on measurement invariance).

And the homicide series carries one of the most important findings in the field, which the next lesson develops: that violence over the very long run has fallen dramatically — a finding that would be invisible in any measure more distorted than homicide, and that contradicts almost everyone's intuition (see 10.2.2).

Goodhart's law, applied to crime, which is where the manipulation lives.

When a crime figure becomes a target, it stops measuring crime (see 7.2.2). And crime figures are targets everywhere — for governments, for police forces, for individual officers.

The documented manipulations are extensive. No-criming — recording a reported crime as no crime. Downgrading — recording a serious offence as a lesser one. Cuffing — not recording at all. Reclassification — moving offences between categories to hit a target. The UK statistics scandal (see 7.4.4) was precisely this: pressure to show falling crime produced systematic under-recording, and the headline crime rate was, for a period, substantially a measure of recording practice under target pressure.

And detection and clearance rates are even more manipulable, because they are ratios (see 7.6.1 on denominators). A clearance rate can be improved by clearing easy cases, by getting offenders to admit to crimes they did not commit in exchange for leniency, by writing off cases as "cleared" administratively, or by not recording the crimes least likely to be solved. A rising clearance rate can mean better policing or better bookkeeping, and the statistic alone cannot tell you which.

The general lesson: any crime statistic attached to a consequence — a target, a budget, a career, an election — should be read as a measure that someone has an interest in moving, and that can be moved without changing any behaviour. This is not cynicism; it is the standing condition of administrative statistics (see 7.4.4), and crime figures are the most politically loaded administrative statistics there are.

Because "crime is rising" and "crime is falling" are among the most consequential claims in politics, and both are usually statements about a gate rather than about behaviour.

Four questions for any crime statistic.

Recorded or survey? Recorded crime measures processing and changes with recording practice; survey crime measures victimisation and misses the victimless and the most serious. Neither alone is the crime rate, and a claim resting on one should be checked against the other.

Which offence, and what is its dark figure? A change in recorded sexual offences is far more likely to be a change in reporting than a change in behaviour, because the dark figure is so large and so responsive. A change in recorded vehicle theft is more likely to be real, because the dark figure is small.

Did a gate move? When the number changes, ask whether behaviour changed or whether perception, reporting, recording, detection or prosecution changed. The divergence between the two measures is the tool for answering this — if they disagree, a gate moved.

And is the figure a target? If a consequence is attached to it, assume it has been optimised, and look for the manipulation the incentive predicts.

One closing observation that frames the whole Topic. Public debate treats the crime rate as a thermometer — a direct reading of how much crime there is. It is nothing of the kind. It is the output of a five-gate machine, each gate operated by people with discretion and incentives, filtering by seriousness, relationship, trust, pressure and social position.

Which means that the crime statistics are, among other things, a measure of the criminal justice system itself — of what it perceives, what it records, whom it detects, and whom it processes. A change in the figures is always ambiguous between a change in the world and a change in the machine — and telling them apart, which requires the triangulation this lesson has described, is the first thing a citizen needs to be able to do before any argument about crime can even begin.

A crime has to pass five gates to become a recorded offence — perception as a crime, reporting, recording, detection, and prosecution — and cases are lost at each, not at random. A "crime rate" measures what survived the first three; a conviction rate, all five.

The dark figure — the gap between crimes committed and recorded — is the majority of crime for most offences, and its size varies systematically: small for vehicle theft (insurance-driven reporting), enormous for sexual offences and domestic abuse (shame, fear, distrust). So recorded statistics over-weight the visible and stranger-perpetrated and under-weight the intimate and the shameful — a biased sample, biased in a known direction.

Police recorded crime covers the whole country over long periods and includes homicide, but measures processing and changes with recording practice — a recorded "crime wave" is frequently a recording change. The victimisation survey captures unreported crime and is stable over time, but misses victimless crime, the most serious crime, people outside the sampling frame, and the most sensitive disclosures. The two can move in opposite directions — recorded crime rising while survey crime falls, because reporting or recording increased — which is why serious analysis triangulates them and treats divergence as a finding.

Homicide is the anchor — small dark figure, consistent recording, comparable across countries and time — and it carries the long-run finding that violence has fallen dramatically (10.2.2).

Goodhart's law applies with full force: crime figures are targets, and no-criming, downgrading, cuffing and reclassification move them without changing behaviour — with clearance rates, being ratios, even more manipulable. Any crime statistic attached to a consequence should be read as one someone has an interest in moving.

And the upshot: the crime rate is not a thermometer but the output of a five-gate machine, so it is partly a measure of the criminal justice system itself, and any change is ambiguous between the world and the machine.

The five gates — perception, reporting, recording, detection, prosecution; each filters non-randomly.

Dark figure of crime — the gap between crimes committed and recorded; the majority of crime for most offences.

Police recorded crime — offences recorded by police; measures processing and changes with recording practice.

Victimisation survey — a population survey of what happened, capturing unreported crime; stable but blind to victimless and the most serious crime.

Triangulation of measures — using recorded and survey crime together, treating divergence as evidence that a gate moved.

Homicide as anchor — the offence with the smallest dark figure and most consistent recording; the basis of comparative research.

No-criming / cuffing / downgrading / reclassification — the manipulations of recorded crime under target pressure.

Clearance / detection rate — a manipulable ratio; a rise can mean better policing or better bookkeeping.

Reporting rate — the share of crime told to police; where most crime is lost, and not lost randomly.

One — walk a crime through the gates. Take one offence and describe, at each of the five gates, why it might not proceed. The exercise shows how much is lost before the statistic.

Two — check the two measures. For your country, find recorded crime and survey crime for the same offence over the same period. Where they diverge, a gate moved — work out which.

Three — rank by dark figure. List five offences from smallest to largest dark figure. Then note which are over- and under-represented in recorded statistics.

Four — find the target. Identify a crime statistic in your country attached to a government target. Then look for evidence of the manipulation the incentive predicts.

Five — use the anchor. For any claim that a country is unusually violent, find its homicide rate rather than its overall crime rate. The overall rate is mostly a comparison of systems; the homicide rate is closer to a comparison of violence.