Perceived crime trends
A Socratic walk-through of perceived crime trends — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #Why are people sure crime is rising in the very years it is falling?
Gallup has asked Americans the same question nearly every year since the early 1990s: is there more crime in the country than there was a year ago? In most of those years a clear majority has said yes. Over the same span, the violent victimisation rate measured by the National Crime Victimization Survey fell from about 80 per thousand people in 1993 to under 20 by the 2010s, and the FBI's tally of crimes reported to police fell alongside it.
So a large majority was wrong, repeatedly, about a matter of public fact, in the same direction, for two decades. That is not the shape of random error. Random error would put roughly half the wrong answers on each side. What could make an error systematic?
Reasoning it through
REASONING #Begin with something easy to overlook: what would it take to perceive a crime rate directly? A rate is a count divided by a population over a period. Nobody has ever seen one. You can be robbed, and you can hear about a robbery, but the denominator — the millions of people who were not robbed today — produces no experience at all. Non-events are unobservable by construction. So whatever people are reporting when they answer that question, it cannot be a measurement. It has to be an inference from what comes to mind.
And what does the mind use as a stand-in for frequency? The ease with which examples are retrieved — the availability heuristic, in Tversky and Kahneman's terms. Usually a decent proxy: common things are easier to recall. Now ask the crucial question. Is the ease of recalling a crime story tied to the crime rate?
Only loosely, and here the coupling has been getting weaker. A newsroom fills a fixed amount of airtime; it does not run fewer crime stories in a safe year, it draws from a wider net — a killing two states away is now nationally available in a way it was not in 1990, and social platforms and doorbell-camera footage supply an effectively unlimited stream regardless of the base rate. So the supply of vivid instances stayed steady or grew while the rate halved. If your estimate is built from instances, the estimate does not fall.
There is a striking piece of evidence sitting inside the same polls. Ask people about crime in their own area and the answer is far more optimistic than the answer about the nation. Why would that gap exist? Because locally you have your own evidence — your street, your neighbours, the fact that nothing happened — and nationally you have only the mediated stream. Where direct experience is available it corrects the estimate; where it is not, the stream stands unopposed. That gap is close to a controlled experiment on the mechanism, and it is hard to explain any other way.
Two more contributions round it out. Negativity bias means threatening information is encoded more strongly and recalled more readily than reassuring information, so the two kinds of evidence were never weighted equally to begin with. And crime is a standing political subject, so the question is often being answered in a season when someone has an interest in raising its salience — which is a real effect, though I would be cautious about how large a share of the gap it explains.
One honesty check before moving on: the decline was not uniform. Homicide rose sharply in 2020, by around 30 percent, the largest single-year increase on record, before falling again over the following years. So "people are wrong" is a claim about the long trend, not about every year — and in 2020 the pessimists were right.
The analogy
THE ANALOGY #It is the same error as judging a lake's water level by counting the boats you can see. Boats are visible, salient, easy to count; the water is not, and its level is the thing you were asked about. If the fleet grows while the lake drains, your count rises and your conclusion inverts — not because you counted badly, but because you counted the wrong thing and had no way to see the right one.
a lake's level is at least in principle observable if you walk to the shore, whereas a national crime rate is never observable by any individual at all — it exists only as a statistic somebody else compiled, so the correction has to be imported rather than looked up in your own experience.
Clarifying the model
THE MODEL #A few refinements, because this is easy to turn into simple contempt for public opinion.
The first is that the public is not answering a badly-posed question stupidly; it is answering an unanswerable question reasonably. Given no access to the denominator, using recalled instances is a defensible strategy. It fails here for a specific structural reason — the supply of instances was decoupled from the rate — not because the reasoning is foolish.
The second is that the official numbers are themselves contested in ways that give the sceptic something real to hold. Reported-crime figures depend on how much people report and how police record it, and FBI collection changed substantially with the shift to the newer reporting system, leaving gaps in some years. The victimisation survey is independent of reporting behaviour but has its own sampling error. The two disagree on levels. What makes the long decline credible is that they agree on direction, from different measurement bases.
The third is that perception matters even when it is wrong. Fear of crime independently shapes where people walk, whether they take the bus, what housing they buy and how they vote — so a persistent gap between perceived and measured crime is a policy fact in its own right, not merely an error to be corrected.
A picture of it
THE PICTURE #How to readthe bars are the share of Americans telling Gallup that crime rose over the past year; the line is the measured violent victimisation rate per thousand people. The line falls by roughly three quarters while the bars stay high and then climb again — the two are not tracking each other, which is the whole point.
What became clearer
WHAT CLEARED #The mismatch is not a failure of attention or intelligence. It follows from a structural fact: the quantity being asked about is invisible to direct experience, so people substitute the one signal they do receive — the availability of vivid instances — and that signal is produced by newsrooms and feeds whose output does not scale down when crime does. When the proxy stops tracking the thing it proxies for, sincere and careful people converge on a confident, shared, wrong answer.
Where to go next
ONWARD #- Why the local-versus-national gap appears in so many other perception questions, from the economy to schools to Congress.
- What actually shifts perception when it does shift, and whether publishing better statistics ever has.
- How fear of crime, as distinct from crime, gets priced into housing, insurance and policing budgets.
Key terms
TERMS #| Term | What it means |
|---|---|
| Availability heuristic | estimating how common something is by how easily examples come to mind. |
| National Crime Victimization Survey | a large annual household survey that measures victimisation directly and therefore does not depend on whether crimes were reported to police. |
| Uniform Crime Reporting | the FBI's compilation of crimes recorded by police agencies, sensitive to both reporting rates and agency participation. |
| Salience | the prominence of a topic in public attention, which can rise or fall independently of the underlying rate. |
Every term the collection defines is gathered in the glossary.