The NYPD’s numbers are exactly what regression to the mean predicts.
Every month, the NYPD releases the prior month’s crime numbers, and lately those numbers have been moving in the right direction. The sharpest drops are happening in the department’s deployment “zones,” small swathes of the city flooded with officers on foot. In these zones, shootings, violent crime and overall crime are all reported down well past the citywide declines. The department credits the deployment strategy.
The drops are real. Fewer people are being shot on those blocks than during comparable time periods in the prior year. But there’s a question lurking underneath the press releases that rarely gets asked out loud: Is the strategy doing the work, or is something else? A long-known statistical phenomenon called regression to the mean — a fancy way of saying that random variations, especially in small sample sizes, typically explain most increases or decreases — predicts that when you pick the blocks that spiked the highest last year and look at them this year, those blocks will likely have fallen on their own, whether or not anyone deploys more police officers to to them.
Earlier this month, the NYPD doubled down on the zone strategy, launching what it called the “largest summer deployment in the department’s recorded history,” sending up to 3,800 officers a night to 72 zones in 40 of the city’s 78 precincts. Each zone averages just over a tenth of a square mile, small enough to walk across in under 10 minutes. The department’s January-through-April press release credited the deployment with major crime going down 22.7%, shooting incidents down 62.5% and shooting victims down 66% inside the zones. Last summer’s 19-week version, by NYPD’s accounting, produced similar numbers, with shootings down 47% and overall crime down 18%. The story is the same season after season: Heavy deployment seems to yield dramatic drops, and the deployment gets credit.
Commissioner Jessica Tisch called the zones “the one strategy that is game changing” in her January 2025 State of the NYPD address, a “hyperlocal, data-driven policing model” in which an algorithm selects “clusters of violence and disorder.” She has framed the approach as “the 2025 modernization of the classic cops-on-the-dots strategy.”
That is shorthand for the Jack Maple formulation, developed under Bill Bratton in the early 1990s, which became the operational core of CompStat. The system paired precinct-level crime data with regular command reviews, making commanders personally answerable for patterns on their watch. It has underpinned every place-based strategy since, from Operation Impact in 2003 to the current Violence Reduction Zones.
The underlying approach has evidence behind it. Decades of research show that concentrating police resources on small, high-crime hotspots reduces violence. A 2024 meta-analysis of 32 evaluations confirmed the finding and added an important distinction. Problem-solving approaches, in which officers and analysts identify and address the specific conditions sustaining crime on a block (an unsecured vacant lot, a failing streetlight, an open-air drug market), outperform programs that simply flood those blocks with additional officers. The question is not whether police presence in small high-crime areas can reduce crime. It can and does. The question is whether the headline drops the NYPD reports from its zones are evidence that this particular deployment is doing so.
To be clear, citywide violence has fallen. Shootings last year hit a recorded low, and murders are also on track to hit historic lows. But what does analysis show when the focus is not on the citywide drop but on the much sharper declines the NYPD reports inside its zones?
It’s not new for the NYPD to attribute significant crime declines in narrow parts of the city to its own efforts when another likely suspect is in the shadows.
In January 2003, the NYPD launched Operation Impact, designating Impact Zones (high-crime areas judged to be experiencing recent spikes) and assigning roughly two-thirds of police academy graduates to flood them on foot. Over the program’s 12 years, it averaged about 25 zones across roughly six square miles. Commissioner Bill Bratton ended it in 2015. “A generation of cops had grown up in a culture of enforcement, not one of relationship building,” he wrote. “When I arrived, cops were still graduating from the Academy and going straight into Impact Zones.”
The zone strategy launched in May 2023 has the same operational logic as Impact: identify blocks that are spiking and flood them with uniformed officers during peak hours. Both versions use algorithms, mathematical tools that look for where crime is concentrating, to pick the blocks. Both rely heavily on rookies straight out of the Academy.
What has actually grown is the footprint. Last summer’s plan covered 12 square miles, twice Impact’s average of six.
Operation Impact was evaluated by independent researchers. In 2016, John MacDonald, Jeffrey Fagan and Amanda Geller compared zones to similar non-zone blocks in the same precincts. Their estimate, based on a more rigorous methodological analysis, came in at 10% to 14%, far smaller than the figures NYPD leads with. The department typically touted higher declines, ranging between 20% and 44%, by comparing crime in a zone during the program to the same period the year before. That arithmetic takes the entire drop at face value, with no control for what would have happened without the deployment.
The selection itself does the work
Regression to the mean works like this. Crime on any given block in any given year reflects two things. One is the block’s underlying rate — the level of crime its population tends to produce. That rate is real and changes slowly. The other is the year’s particular bad luck: a crew that moved through, a hot summer, one repeat offender out and active. By definition, bad luck has no reason to repeat the next year.
So when the department picks blocks at the very top of last year’s count, it has picked partly the blocks that need the most intensive intervention and partly just those that are experiencing bad luck or statistical noise. The first part stays. The second part doesn’t.
Next year, crime levels on the same block are bound to decline, whether or not anyone did anything.
The effect is amplified under three conditions: when the unit being measured is small, so a single event swings the count; when the window used to select a block is short, so one unlucky stretch is the whole record; and when the thing being measured swings around a lot from year to year on its own.
The zone strategy checks all three boxes. Zones average about 0.17 square miles, or about 24 (Manhattan-sized) blocks. The NYPD told a joint Council hearing on April 29 that the School Safety Zones are designated on the prior year’s trends. And shootings, prominent among the reported outcomes, fluctuate block-to-block more than almost anything else public-safety agencies track.
So how much would crime drop if no one did anything?
The NYPD has not published the algorithm behind how it selects its Violence Reduction Zones, but the department has said they are picked based on recent spikes. Take the blocks scoring highest on a recent crime measure, stack them until the total footprint hits about 12 square miles, and call the result the zones. Whatever the formula, the act of picking the worst blocks in a given moment is itself enough to produce a drop the following year.
For the purposes of this exercise, the exact method matters less than the act of selection itself. So we built a simple model that mimics the selection process, picking the highest-crime blocks each year and tracking what happens next, to show how much of the decline regression to the mean alone can explain. For every summer from 2010 through 2025, we ranked blocks by their crime incident count in a defined look-back window, i.e. the prior summer, the prior year or the prior three years of crime or shootings numbers. We then took the top blocks until the cumulative area reached 12 square miles, and recorded what happened in those blocks during that summer compared to the previous summer. We tested two outcomes (shootings and street violence like robberies and assaults) across three look-back windows (the prior three years, the prior year and the prior summer alone). Find our detailed methodology here.
Two windows in this period are worth keeping straight. Operation Impact ran from 2003 through 2015. The current Violence Reduction Zones launched in May 2023 and are still running. In the years between Impact and the current zones, the department stepped back from centralized, year-round hot-spot deployment. It adopted a Neighborhood Policing plan instead, a precinct-driven model that left enforcement priorities to local commanders and communities. To deal with increasing summer violence, the department instituted Summer All-Out, which sent a few hundred administrative cops to precincts with a lot of violence. Their deployment was left up to police commanders and has been shown to have no impact on summer crime.
In every summer from 2010 to 2025, whether or not a zone strategy was in place, the blocks that spiked the previous year fell, and they fell far more than the city as a whole.
From 2015 through 2023, when no zone strategy existed, crime in simulated zones fell faster than the citywide rate every summer, with no extra officers deployed. For shootings, picking blocks on the prior year’s count produced an average drop of 76%, even as citywide year-over-year shootings averaged a 6% rise across the window. That is a gap of over 80 percentage points. For outdoor street violence, the same selection produced an average drop of 32%, against an average citywide rise of 2%.
The narrower the selection window, the bigger the apparent drop. Picking on the prior summer alone, the narrowest window we tested, averaged a drop of 88% for shootings and 55% for street violence. During this period, there were no zone deployments, yet areas that had been spiking previously still saw drastic declines.
The NYPD’s reported drops sit right inside the mechanical range, meaning they’re what one would expect to see even if police did nothing.
For the broader major-crime category, the September 2025 press update reported that total crime in its zones went down 18.3% over the first four-plus months of the summer plan, as compared to a year-end citywide major-crime drop of 3%. That 15-point gap also falls well within the simulated range, slightly above what our most conservative look-back produces.
Regarding School Safety Zones, which the NYPD has stated are chosen based on the prior year’s trends, the department reported crime down 53% and shooting incidents and victims each down more than 75% in the program’s early months. Those numbers also sit squarely inside what our simulation produces with no deployment at all, under both prior-year and prior-summer look-backs.
The acid test was summer 2020. If anything should overwhelm a selection effect like the one described here, it is a year when the entire city moves in the wrong direction. Summer 2020 was such a year. New York’s shootings increased 150% compared to the summer of 2019, the largest single-year increase in modern city history. By most measures, police activity collapsed over the same period.
Even in the summer of 2020, when New York’s shootings increased 150%, the blocks selected based on 2019 spikes still fell.
If prior-year trends were used to select them, they fell 47%; if the prior-summer trends were used, they fell 80%. The reason is that the 2020 surge was spatially diffuse. It spread to blocks that had not been on the 2019 hot list. The 2019 hot blocks reverted toward their long-run rate even as the city around them more than doubled.
If a program had targeted those 2019-hot blocks in 2020 and reported a 50% to 80% drop, it would have been hailed as a triumph. That drop, however, would have happened with or without any program. The city’s worst summer in two decades did not overpower this effect.
So what should we make of the zone numbers?
While all of this complicates the NYPD’s simple cause-and-effect assertions, none of this proves that it’s useless to deploy police to places in the throes of crime spikes. “Hot-spot policing” is proven to be effective in many contexts.
In their 2016 independent evaluation, MacDonald, Fagan and Geller found Operation Impact reduced crime by about 10% to 14%, and the gains spilled over to neighboring blocks rather than pushing crime around the corner. The current program could plausibly produce something similar. But the headline numbers NYPD publishes don’t tell us whether it does.
To distinguish a real deployment effect from the drop that selection alone produces, researchers need a credible counterfactual. They need to isolate what the intervention did beyond what would have happened anyway. For example, one could consider comparable blocks where police were not deployed, and observe those also during the same period. This is the standard in the criminology literature, what grant-funders expect in their reviews and what MacDonald and his coauthors did with Operation Impact a decade ago.
The City is looking to spend real money on the zone program, expanding it to what the department calls the largest scale in its recorded history. Whether it works, and how much, deserves a proper evaluation, using sophisticated analytical methods. The data to do such an evaluation largely exist on NYC Open Data or in NYPD’s own systems. Philadelphia commissioned an independent foot-patrol evaluation a decade ago. Other cities do this routinely.
Public safety spending that can’t survive rigorous analysis shouldn’t survive the budget process.
