One Point Three Miles of Sidewalk

On page 816 of a 2011 issue of Criminologythere is a line of arithmetic with six numbers in it. Worked through, it comes to 53.11. That is the paper's estimate of how many violent crimes didn't happen in Philadelphia in the summer of 2009 because pairs of police officers, fresh from the academy, walked the same few blocks all day.

We spend most of this blog on technology. License plate readers, drones, acoustic gunshot sensors, AI alert apps. This post is about the oldest tactic in the book: an officer on foot, on a corner, being seen. Is there good evidence that it cuts crime? There is, and it comes with conditions attached that matter as much as the headline number.

The Newark verdict

For about thirty years the answer was supposed to be no. The Police Foundation's Newark Foot Patrol Experiment, published in 1981, compared beats that kept, lost, or gained foot patrol and measured what changed. Residents noticed. They felt safer, and they rated crime problems as less severe. Crime itself didn't move in any way the study could detect.

That became the settled view. Jerry Ratcliffe, the Temple University criminologist who later led the Philadelphia study, puts it plainly on his project page: since the 1980s, many police and researchers held that foot patrol improves how people see the police and reduces fear, “but they don't prevent actual crime.”

Foot patrol became the thing you did for the community meeting. Cars did the crime work.

How Philadelphia drew the beats

In 2009 the Philadelphia Police Department had two waves of recruits coming out of the academy, in March and late June. Commissioner Charles Ramsey wanted them on small foot beats in violent areas, mainly to hold down summer violence, and the department invited Ratcliffe's team to turn the deployment into a randomized trial. The design is worth walking through, because the details turned out to be the finding.

The researchers started with the city's nearly twenty-two thousand street intersections. Around each one they drew a polygon containing all the space closer to that intersection than to any other (a Voronoi network, if you want the term). Then they counted violent crime from 2006 to 2008 in each polygon and ranked them.

The top 220 corners, roughly the top 1 percent, held 15% of the city's 2008 robberies, 13% of its aggravated assaults, and more than 10% of its homicides. Widen to the top 5 percent of corners and the shares climb to 39% of robberies, 42% of aggravated assaults, and 33% of homicides. If you have read our piece on hotspot mapping, that concentration will look familiar. It is the reason place-based policing exists at all.

District commanders drew 129 candidate foot beats, each containing at least one of the top 220 corners. After splitting and merging some, the team had 124, dropped the four quietest, and ranked the remaining 120 by violence. The first and second went into one pair, the third and fourth into the next, and so on. A random number generator picked one beat in each pair. Sixty got officers. Sixty didn't.

Each target beat got two pairs of officers covering a 10 a.m. to 6 p.m. shift and a 6 p.m. to 2 a.m. shift, Tuesday through Saturday, swapping shifts every other week. On paper that came to 57,600 hours of foot patrol over the first twelve weeks. The beats were small. They averaged about fifteen intersections and 1.3 miles of street.

The arithmetic, in the open

The paper reports a 23% relative reduction in violent crime in the target beats compared with the controls. That figure comes from an inverted odds ratio of .77. The more interesting number is the one that accounts for displacement, because the first thing any patrol officer will tell you is that crime just moves around the corner.

So the team drew a buffer around each target beat, generally no more than two long Philadelphia block-lengths deep, and counted crime there too. Here is the equation from the paper, with the real counts filled in:

TNE = [target_before × (control_during / control_before) − target_during]
    + [buffer_before × (control_during / control_before) − buffer_during]

TNE = [359 × (327/296) − 306] + [320 × (327/296) − 391]
    = 90.60 + (−37.49)
    = 53.11

Read it left to right. Control areas went from 296 violent crimes to 327, so the city as a whole was running a bit hotter that summer. Scale the target beats' 359 by that trend and you would expect about 397. They had 306. That's ninety-one fewer.

The buffers went the other way. Expected about 353, got 391. Some crime did move next door, around thirty-seven offenses of it. Net it out and you get fifty-three. Displacement was real and it was smaller than the benefit.

Two notes on the method, because they are the kind of thing we look for before trusting a result like this.

First, the simplest possible test fails. Compare violent crime counts in target and control beats during the experiment and there is no significant difference (a mean of 5.10 against 5.45). The authors report that in a footnote rather than burying it. The effect only shows up once you control for where each beat started, which is exactly what you should do. High-crime places drift back toward average on their own, the regression-to-the-mean problem we wrote about in Three Burglaries Is Not a Trend. A study that picks the worst corners and then celebrates a drop, without a control group, is measuring gravity.

Second, the effect wasn't spread evenly. The abstract says it plainly: only target areas in the top 40 percent of pre-experiment violence had significantly less violent crime. Below that, foot patrol didn't register. The authors' phrase is that it works “as long as a threshold level of violence exists initially.”

(We'll skip most of the modeling. The short version is that they used ordinary least squares with the pre-period count as a covariate, checked it against Poisson and negative binomial models, and got the same answer.)

What the officers actually did

Walking the beat sounds passive. The activity records say otherwise. In the target areas, compared with before, pedestrian stops rose 64%, arrests rose 13%, drug-related detections rose 15%, and vehicle stops rose 7%.

Field researchers who went along saw a wide range of styles. Some officers visited child care centers and talked to people on the block. Others stopped cars at stop signs and questioned pedestrians. A follow-up study found that the areas officers really patrolled ran about 0.13 square miles larger than the ones they were assigned. Some strayed because they thought offenders had moved one street over. Some strayed because they were bored.

In our view this is the part people leave out when they talk about “just putting cops on the street.” The researchers' own explanation for the drop is deterrence through the certainty of being disrupted, stopped, or arrested. A visible officer is part of that. So are a lot more stops, and stops carry costs of their own that this experiment wasn't built to measure. And, awkwardly for the Newark story, a 2015 survey from the same research program found that residents of Philadelphia foot patrol hot spots didn't feel any safer.

Then the officers left

The experiment ended and most of the beats were withdrawn. Evan Sorg, Cory Haberman, Ratcliffe, and Elizabeth Groff went back to see what happened next, and published the answer in Criminology in 2013. The beats went back to their pre-experiment crime levels. Ratcliffe's summary is blunt: there was no evidence the benefits lasted beyond the time officers were assigned.

There was a second, quieter finding. Beats that ran longer than three months showed diminishing effects even while officers were still walking them. Whatever deterrence the officers built up, it wore off with familiarity.

The same group ran a second trial, the Philadelphia Policing Tactics Experiment, published in 2015. It tested three approaches across 81 violent hot spots. Offender-focused policing, which targets the specific people driving violence in an area, produced a 42% drop in violent crime and a 50% drop in violent felonies against its controls, according to the National Institute of Justice's CrimeSolutions review. Foot patrol and problem-oriented policing didn't significantly reduce violent crime in that trial.

That second result deserves weight. It's a smaller trial (twenty treated places per tactic), so it had less power to detect a modest effect. It still means the 2009 result is one well-run experiment with a replication that didn't come out the same way. That's the honest state of the evidence.

The dosage question

The 2011 paper offers a guess at why Philadelphia worked where Newark hadn't, and it is the most useful idea in the whole body of work. Newark used existing foot beats, some of them commercial corridors up to sixteen blocks long. An officer on a beat like that doesn't come back to the same corner very often.

The Philadelphia beats averaged 1.3 miles of street. The authors cite Flint, Michigan, where the department expanded its foot patrol areas over the research team's objections, in one case to twenty times the original size, and the crime reduction fell off substantially (Trojanowicz, 1986). The authors call their own work “an important first step rather than the final word.” We think that is the right register.

The pattern that emerges, and it is a pattern rather than a law, looks like this. Foot patrol seems to work when the beat is small, the place is among the most violent in the city, and the deployment is fresh. Stretch the beat, pick a quieter corner, or leave the same pair there for six months, and the effect thins out. Take the officers away and it disappears.

Against the national decline

It's tempting to connect this to the drop in crime now under way. The Real-Time Crime Index, drawing on 565 agencies covering 113.2 million people, has murder down 16.3% and violent crime down 5.4% for January through July 2026 compared with the same months of 2025. Robbery is down 13.3%.

We wouldn't make that connection. Nothing in the Philadelphia work says anything about why crime is falling across the country, and a lot of cities are posting similar declines under very different patrol strategies. What the experiments offer is narrower and, we think, more useful. They tell you how a specific tactic shows up at a specific place, and how long it lasts. (For the broader question, our posts on the historic decline and criminogenic entropy are a better starting point.)

Reading a foot beat in the data

If you build on block-level crime data, you'll eventually see a place where a patrol deployment started or stopped. Here is what the Philadelphia work suggests you check before calling anything a trend.

1. Split by how the incident started. Victim-reported violence and officer-initiated activity move in opposite directions under foot patrol. A combined count can stay flat while both halves change.
2. Look two blocks out.Philadelphia's buffers were about two block-lengths deep, and they absorbed thirty-seven extra offenses. A drop inside a hot spot and a rise next door can be the same event.
3. Don't score a drop without a baseline. The worst corners improve on their own. Compare against similar places, not against the same place last quarter.
4. Mind the calendar. The 2009 beats ran in summer, when violence runs high anyway, and had no coverage from 2 a.m. Sunday to 10 a.m. Tuesday. Our seasonality post covers why a summer window needs its own comparison.
5. Expect decay. A deployment that has been in place for six months is not the same treatment it was in month one. When it ends, assume the old level comes back unless the data says otherwise.

None of this needs new technology. It needs knowing where the officers were, which is often the hardest field to get.

In Newark, a beat could run sixteen blocks. In Philadelphia it averaged a little over a mile of sidewalk, and the researchers' best explanation for the difference is how often an officer came back past the same corner. Fifty-three violent crimes, by their count, came down to that.