Chapter 2 — The Invention of the Crime Rate

Before anyone could investigate crime scientifically, somebody had to count it. That turned out to be much harder than investigating it, and two hundred years later it is still not solved.

The consequences run through everything that follows in this book. Every claim that a technique works, every claim that a country has a problem, every claim that anything has improved, is a claim about a number — and the numbers in this field are stranger than almost anyone using them realises.

The scaffold pays its bill

In the 1820s the French state began publishing something no state had published before. The Compte général de l'administration de la justice criminelle, first volume 1827, was the world's first continuous national criminal-justice statistics series: the accused, the convicted, the recidivists, counted annually across every penal jurisdiction in France. [ESTABLISHED] (Criminocorpus; Ministère de l'Intérieur, on measuring delinquency)

Two men read it and found something disturbing.

André-Michel Guerry presented his Essai sur la statistique morale de la France to the Académie des Sciences on 2 July 1832, publishing it in 1833. He mapped crime, suicide and literacy across the French départements — tables and shaded maps, a technique so novel it won the Prix Montyon. It is generally regarded as the founding document of empirical social science. [ESTABLISHED] (Friendly, A.-M. Guerry's Moral Statistics of France, arXiv:0801.4263)

Adolphe Quetelet, working the same data from 1826 to 1831, found the thing that unsettled the century. Crime totals did not fluctuate. They recurred, year after year, with a regularity you would expect from a physical process rather than from millions of separate human choices. He put it in a sentence that has never been improved on:

"There is an account paid with a terrifying regularity; that of the prisons, the galleys, and the scaffolds… there is a tribute man pays more regularly than those owed to nature or to the Treasury; the tribute paid to crime!"

[ESTABLISHED] — from Sur l'homme et le développement de ses facultés (1835), in the standard English rendering (MacTutor, Quetelet on crime).

Understand what this meant to people who read it. Crime was supposed to be the purest expression of individual moral failure — a free choice, made by a person, blameable. Quetelet had shown that society produced a stable annual quantity of it, as reliably as it produced wheat. The murders had not yet been committed and their number was already, approximately, known.

From that came the homme moyen, the average man, introduced in his 1831 crime work and elaborated in 1835: a fictitious person for whom everything happens at the societal average. [ESTABLISHED] And from that lineage came sociology itself. Durkheim's Le Suicide (1897) built directly on Quetelet's demonstration of statistical regularity as evidence that social facts exist independently of the individuals who compose them — while attacking the average man as a confusion between the collective character of a society and the mean of its members. [ESTABLISHED] (Porter, Statistical and Social Facts from Quetelet to Durkheim)

So the crime rate is not a bureaucratic convenience. It is one of the founding objects of the social sciences, and the argument about what it measures is the same age.

The crack in the foundation

There is an obvious objection, and Quetelet had already made it himself in the early 1830s: the figures count crimes the state found out about. Beneath them lies an unmeasured residue — offences committed and never reported, never detected, never recorded.

That residue acquired a name, the dark figure of crime, though I should be careful here: the concept is Quetelet's, while the fixed phrase and its first user I could not pin down. [ESTABLISHED] for the concept; [ABSENT] for the coinage. (dark figure of crime, with its citation to Quetelet 1831)

For a hundred and forty years nobody could measure it, because the only instrument available was the police record, and the whole problem was that the police record was the thing in doubt.

The solution, when it came, was to stop asking the police and start asking the public.

• The United States fielded the first National Crime Survey in July 1972 — 72,000 households,

15,000 businesses, 26 metropolitan areas — becoming an annual series in 1973 and renamed the National Crime Victimization Survey in 1992. [ESTABLISHED] (Bureau of Justice Statistics)

• The British Crime Survey followed in 1982, created explicitly in response to an acknowledged

crisis of confidence in Home Office police-recorded figures. It became the continuous Crime Survey for England and Wales in 2001. [ESTABLISHED] (Crime Survey for England and Wales)

Ask people directly whether anything happened to them, and compare the answer to what the police wrote down. Here is what fifty years of that has produced.

In England and Wales, year ending March 2026, an estimated 36% of headline crime — theft, robbery, criminal damage, violence — was reported to police. Nearly two-thirds was not. [ESTABLISHED] In the United States, NCVS 2023: 46% of rape and sexual assault, 42% of robbery, 57% of aggravated assault, 41% of simple assault were reported. [ESTABLISHED] (BJS, Criminal Victimization 2023)

Roughly half of ordinary crime, in two of the best-instrumented countries on earth, never enters the statistics at all.

Fraud is worse, and fraud is our subject

Now the number that should reorganise how you read every figure in this book.

England and Wales, year ending March 2025. The Crime Survey estimated 4.2 million fraud incidents. Action Fraud, the police-facing recording body, recorded 299,046 fraud offences in the same period.

That is a capture rate of about seven per cent. Ninety-three per cent of fraud, in a wealthy country with a mature statistical apparatus and a dedicated national reporting centre, does not appear in the recorded-crime series. The Office for National Statistics says so itself: the recorded fraud series "is not a good indicator of volume or trends over time because a large proportion of fraud victims do not report incidents to Action Fraud." [ESTABLISHED] (ONS, Fraud and computer misuse in England and Wales, YE March 2025)

Computer misuse in the same period: 692,000 incidents estimated, 55,576 recorded. About eight per cent. [ESTABLISHED]

And the Home Office's estimate of the total economic and social cost of fraud, year ending March 2024: £14.4 billion — £9.2bn falling on individuals, £5.2bn on businesses. [ESTABLISHED] (Home Office)

Sit with the seven per cent. It means that a national fraud statistic is not a measurement of fraud. It is a measurement of reporting behaviour, lightly correlated with fraud. If reporting rises ten per cent because a campaign told people where to complain, the statistics record a crime wave that did not happen. If reporting falls because victims conclude nothing will come of it, the statistics record a triumph.

Every figure in this book that comes from a police or complaint system — including all of them — sits on top of that.

Why one country's number will not compare to another's

This is the part that answers the question people actually ask: is it worse here than there?

Three independent things break the comparison, and each alone is enough.

Legal definition. What counts as the offence differs. UNODC states it about the offence usually considered the most comparable in the world: "Varying legal definitions of homicide among countries, together with differences in capacity, criteria and approaches to case recording, mean that a comparative analysis of homicide statistics must be conducted cautiously." [ESTABLISHED] (UNODC, Methodological Annex, Global Study on Homicide 2019)

If a body with a bullet in it is not cleanly comparable, nothing softer is.

Counting rules. Does a series of offences against one victim count as one crime or many? Does an incident with several offences count once, under the most serious? Countries answer differently, and the answer multiplies or divides the total.

Reporting propensity. Compare the English fraud figure above with the American violent-crime figures and you are comparing two publics with structurally different willingness to report, shaped by trust in police, by stigma, and by insurance — a claim usually requires a crime number, which inflates recorded property crime in countries with high insurance penetration and does nothing for the uninsured.

The instrument built to escape all this was the International Crime Victims Survey, first fielded in 1989 in thirteen countries, run six times — 1989, 1992, 1996, 2000, 2004/05, 2010 — across eighty countries over its life. [ESTABLISHED] (UNICRI)

Note the last round. 2010. The best mechanism anyone built for comparing crime between nations stopped running sixteen years ago.

What exists instead is a classification standard: the International Classification of Crime for Statistical Purposes, adopted by the UN Statistical Commission in March 2015, precisely because national legal categories and pipeline stages could not be reconciled. [ESTABLISHED] (ICCS final text) A common vocabulary is real progress. It is not a common measurement.

The example that teaches the lesson

Sweden recorded 64 reported rapes per 100,000 residents in 2016; Germany recorded 10. On the face of it, a factor of six.

Recalculate the Swedish incidents under Germany's narrower legal definition and counting rules and the Swedish figure falls to about 15 per 100,000 — still higher, no longer a different planet. Three mechanisms drive the gap: Sweden's broad definition, extended in 2018 to a consent standard; a counting rule that records each incident in a series of repeat victimisation as a separate offence; and the fact that around 40% of Swedish reported rapes in 2016 involved no violence, a category some other countries would not have classified as rape at all. [ESTABLISHED] for the recalculation (Brå, Reported and cleared rapes in Europe, 2020)

And the other side, which must be stated because it is strong: a broad, consent-based definition is arguably the more accurate legal standard, and the narrower ones elsewhere are the failures. Higher willingness to report — itself a sign of functioning institutions — raises the count independently. Disentangling artifact from behaviour is not settled. [CONTESTED]

What the evidence supports is only this, and it is enough: the raw cross-country number will not carry either conclusion. Anyone who quotes it to prove a country is dangerous, or safe, is doing something the data cannot do.

Four ways to say who has the worst fraud problem

Here is the scale question, put concretely. Same phenomenon, four normalisations, and the ranking does not survive the trip.

[ESTABLISHED] for each figure within its own stated scope. Sources in the dossier and in Appendix B for the Thai line.

By absolute loss, the United States and the United Kingdom have the world's worst fraud problem.

By loss per head, that inverts. Singapore, with about six million people and roughly S$0.9bn in net losses, sits near $150 per person. The United States, $16.6bn across roughly 335 million, sits near $50. Thailand, roughly $1.7bn across about 72 million, sits near $24. The smallest absolute total produces the largest per-capita burden. [ESTABLISHED] as arithmetic on the figures above; the population denominators are approximate.

By share of GDP and by loss relative to median income, the ranking would shift again — and I did not compute either, because doing it properly requires GDP and income figures sourced fresh rather than carried. [ABSENT], declared rather than estimated.

None of the four is wrong. Each answers a different question. Absolute loss tells you where the money is and therefore where enforcement resources rationally go. Per-capita loss tells you the chance that this happens to you. Share of GDP tells you what the economy is losing. Loss against median income tells you what it does to a household — and that last one is the only one that describes suffering, because 50,000 baht taken from a household is a different event depending on what that household earns in a month.

Before the numbers, though, sits the plainer version of the same problem. When a Frenchman says there is a lot of crime and a Thai says there is a lot of crime, the two sentences are not comparable, and the reason is not that one of them is mistaken. They are describing different quantities, through different reporting systems, against different baselines of what feels normal. The reason this chapter exists is that the numbers, which look like they should settle it, turn out to be exactly as local as the sentences.

And the number everyone asks for

"What proportion of cases get solved?" is the natural question, and it has no international answer.

In the United States, the Uniform Crime Reporting definition of cleared includes both clearance by arrest — someone arrested, charged, and turned over for prosecution — and exceptional clearance, where something outside police control prevents that: the offender has died, the victim will not cooperate, extradition is refused. The traditional published aggregate does not distinguish between the two, which means a headline clearance rate can be substantially better than the arrest count would support. [ESTABLISHED] (FBI UCR, Clearances)

In England and Wales, since April 2014, every recorded crime is assigned one of around twenty defined outcome types — including categories like "evidential difficulties: suspect identified, victim does not support action," which has no American equivalent at all. Before 2014, only about 30% of crimes received any outcome. [ESTABLISHED] (Home Office crime outcomes)

These are not two measurements of one thing. They are two different constructs, computed over denominators that are themselves incomparable for the three reasons above.

For fraud specifically, the record thins to almost nothing. The clearest figure I found is British and dated: in the year ending September 2021, the Crown Prosecution Service prosecuted 7,609 defendants with fraud or forgery as the principal offence — around 0.75% of fraud crimes reported that year — securing convictions in 84.9% of those prosecutions. [CONTESTED] in the sense that I reached it through a summary of parliamentary evidence rather than the underlying document, and it should be re-checked before anyone quotes it. (UK Parliament Justice Committee, Fraud and the Justice System)

And the United States — the country publishing the most granular fraud-loss figures in the world — publishes no fraud clearance rate at all. IC3 is a complaint clearinghouse, not a clearance reporting system. [ABSENT]

Read those two together. Where the number exists it is under one per cent, and mostly it does not exist. Anybody who tells you what proportion of fraud gets solved, in any country, is either quoting something that has been re-checked or making it up.

What this chapter is for

Everything after this rests on measurements, and you now know what kind of objects they are.

A recorded-crime figure is a measurement of reporting behaviour. A cross-country comparison is almost always invalid and the invalidity is structural, not sloppy. A clearance rate means something different in every jurisdiction that publishes one, and for fraud most do not publish one.

This is not a reason for despair, and it is certainly not a reason to stop counting — the alternative to bad measurement is not good judgement, it is anecdote, and Quetelet's whole achievement was proving that anecdote had been lying to everyone for centuries. It is a reason to hold every number in this book by its handle: the period it covers, the body that issued it, the definition it used, and what it is actually a count of.

The volume that arrives out of all this — hundreds of thousands of reports a year, in every country with an internet connection — is the problem the second part of this book is about. But the first thing to know about that volume is that it is the visible seven per cent.

The Move — Ask for the denominator. Before proposing any improvement to any investigative process, ask the owning organisation for the number you intend to improve, as it stands today, with its definition and its window. Do not accept a proxy and do not accept a rate without being told what sits underneath it. If the number does not exist, that is your finding and it is your first piece of work: not the tool, but the measurement that would let anyone tell whether the tool helped. An organisation that cannot show you its denominator is one where your success cannot be demonstrated — which means it also cannot be defended when someone senior decides it should stop.

What would change my mind: if a body published cross-nationally harmonised fraud statistics with comparable definitions, counting rules and reporting-rate adjustments — the thing the ICVS was attempting when it stopped in 2010 — then most of this chapter's caution would become obsolete and comparison would become a legitimate move. The ICCS is a step toward it. It is a vocabulary, not yet a measurement, and until it is, the caution stands.