Eight Hundred Thousand People Stopped Showing Up

Between 2019 and 2021, about eight hundred thousand people stopped turning up in Manhattan. The Census Bureau counted the drop: New York County's commuter-adjusted population, the number of people actually present on an ordinary working day, fell by roughly that much in two years. Its resident population fell by about fifty thousand. Both numbers come from the same survey. Only one of them has ever appeared underneath a crime rate for Manhattan.

A crime rate is a fraction. Nearly all of the argument in this field happens on the top.

The half nobody audits

We have written a great deal about the numerator. How an incident gets a coordinate, in the geocoding problem. How much of a feed measures patrol rather than crime, in the enforcement confound. Why a count of three cannot support a sentence, in Three Burglaries Is Not a Trend. Those are all arguments about what goes in the top of the fraction.

The bottom usually arrives as a single number from a household survey, describing where people sleep. It comes with no method note, because it isn't treated as a method. It gets divided into everything.

Keith Harries put the objection plainly in 1991. Using residential population uncritically as the denominator for every offense category, he wrote, yields patterns that are “at best misleading and at worst bizarre.” Sarah Boggs had made a version of the same argument in 1965. Her point was that the population at risk depends on the offense. The people exposed to a commercial burglary are not the people who sleep on the block.

Sixty years later, the residential denominator is still the default in nearly every published rate, including the ones we quote. The 2024 US violent crime rate of 359 per 100,000 means 359 per hundred thousand people counted at their home address. At the national level that is close enough to harmless. Commuters mostly stay inside the country.

Nobody sleeps where the crime is

Zoom to a city and the assumption starts to bend. The Census Bureau publishes a commuter-adjusted population precisely because it knows this: residents, plus workers who come in, minus workers who leave. Its own chart of the top twenty cities breaks the daytime figure into four bars, one of which is workers commuting in and another of which is workers commuting out. For Washington, Boston, and San Francisco the incoming bar carries a large share of the whole. For Los Angeles the outgoing bar is the longest on the page.

The pandemic then ran a natural experiment on the whole arrangement. Manhattan's residents working from home went from about fifty-nine thousand in 2019 to roughly three hundred and twenty-five thousand in 2021, and the commuters who stopped arriving vastly outnumbered them. Across the East River, Kings County went the other way: eighty-one thousand more residents, three hundred thousand more people present on a workday.

Two boroughs of one city, moving in opposite directions, and the resident count registers almost none of it.

A crime rate computed for Manhattan in 2021 divided by the resident population produced a number that had not noticed the largest change in exposure the borough had seen in living memory. The arithmetic was fine. It answered a question nobody was asking.

Four ways to count who is actually there

The thing researchers want is called the ambient population, and Patrick Johnson, Martin Andresen, and Nick Malleson define it about as simply as it can be defined: the number of people in a given area engaged in their day-to-day activities. That includes the people whose day-to-day activities happen fifty feet from their own bed. Four measurement approaches have real literature behind them, and each one is biased in its own direction.

The first is a global population grid. Andresen and Greg Jenion used Oak Ridge National Laboratory's LandScan database as a crime-rate denominator for Vancouver in a 2010 Security Journal paper, and reported large differences in the resulting spatial patterns. LandScan is a modeled twenty-four-hour average, which is its strength and its limitation: it smooths across the entire day, so it cannot tell you about two in the morning.

Second, the Census workday population, the same commuter adjustment described above. Malleson and Andresen tested five candidate denominators for theft from the person across London in a 2016 paper in the Journal of Criminal Justice, including mobile telephone locations and social media, and found the workday population the most appropriate of the set. It is also the only one on this list that a small agency can obtain for free with a documented methodology.

The third is telecom infrastructure. Johnson, Andresen, and Malleson built a denominator from OpenCellID cell tower data in a 2021 study of five disaggregated property crime types in Vancouver, published in the European Journal on Criminal Policy and Research. Cell data has global coverage. It also counts phones, which is not the same as counting people, and it is denser where the network is denser.

Last, geotagged social media, which Malleson and Andresen examined in a 2015 Crime Science paper on spatio-temporal hotspots. It has the finest time resolution of anything here, hour by hour, and the worst sample: it measures people who post, in places worth posting from.

The map moves when the denominator moves

This would be an academic quarrel if the two denominators produced similar maps. They don't. The 2021 Vancouver study reports striking differences in the spatial patterns of the crime rates built from each measure, and the difference is not confined to description. In their multivariate models, switching the denominator changed which variables were retained and which relationships came out significant. Their conclusion about what follows is unusually blunt for a journal article: policies designed around residential-based crime rates risk “having no effect, or even of increasing crime.”

Think about what that means for a product rather than a policy. Rank the neighborhoods of a city by violent crime per resident and the top of the list will be its downtown, its transit hubs, its stadium district, its hospital campus. Those places have real incidents. They also have almost nobody in the denominator. A rate is being asked to describe risk to a person, and the person it describes is a resident, and the incident happened to a commuter.

The distortion runs the other way too. A dense residential tract with a large population and a moderate incident count will rate as safe against a downtown that half a million people visit daily. We made a related point about state-level figures in the geography of the crime decline, where the national number hides a seven-to-one spread. The denominator problem is the same failure at a different scale, and it does not average out.

Where the denominator is zero

At block level the two problems compound. A small numerator sits over a small denominator, and both are noisy in different ways.

Some blocks have no denominator at all. A stadium parking lot, an industrial strip, a park, a rail yard, a block that is entirely a shopping mall. The resident population is zero or close enough that any division produces nonsense. Products handle this by suppressing the tile, or by falling back to a larger geography, or by quietly emitting a rate computed against eleven people. The third option is the common one, and it is the one that reaches users.

The kernel density and hex-grid methods we walked through in crime hotspot mapping inherit this directly. A density surface built on counts shows you where incidents are. A density surface divided by residents shows you where incidents are relative to bedrooms. Neither is the map most people think they are reading, which is risk to a person standing there.

Six things to do about it

1. Name the denominator in the response body. If an endpoint returns a rate, return the population figure, its vintage, and its type in the same object. A consumer who sees denominator: residential_acs_2024 can reason about the number. A consumer who sees a float cannot.
2. Match the denominator to the offense where you can.Residential burglary against housing units. Commercial burglary against business establishments. Motor vehicle theft against registered vehicles or parking capacity. This is Boggs' sixty-year-old advice and it is still mostly unimplemented, because the offense-specific denominators are harder to get than the population one.
3. Never rank places by per-resident rate across mixed land use. A league table that puts a central business district against a subdivision is comparing a rate to a different rate with the same name. If a ranking has to exist, restrict it to comparable land use, or rank on counts and say so.
4. Use counts where counts are the honest answer.“Fourteen incidents within a quarter mile in the last ninety days” is checkable, transparent, and does not smuggle in a population model. Most user-facing questions are answered better by a count and a radius than by a rate.
5. Treat the denominator as a time series. Manhattan between 2019 and 2021 is a structural break in the bottom of the fraction, and a rate series computed against a static population will read that break as a change in crime. Any post-2020 comparison of downtown rates carries this artifact whether or not anyone has looked for it.
6. Set a floor and refuse below it. Pick a minimum denominator and return null rather than a rate when a geography falls under it. Our own aggregation and normalization choices, including how small geographies are handled, are documented in the SpotScore methodology.

The number that does not have this problem

One useful check. Comparisons of the same place to itself over time are much less exposed to the denominator, because whatever population model is wrong is wrong in both periods and mostly divides out. That is part of why the headline figures from the Real-Time Crime Index are percent changes rather than levels. Its current sample of 590 agencies covering 119.4 million people shows violent crime down 5.9 percent and murder down 17.3 percent for January through June 2026 against the same months of 2025. Those figures survive a bad denominator. A cross-sectional ranking of neighborhoods does not.

The rule of thumb is short. Change over time in one place is fairly robust. Comparison across places at one time is where the denominator decides the answer.

Somewhere tonight a dashboard is computing violent crime per capita for New York County. It will divide by the number of people who sleep there. It will come out clean and confident, to one decimal place.