The Baltimore That Didn't Happen

Figure 8 of a June working paper from the National Bureau of Economic Research shows a city that doesn't exist. It's built from weighted pieces of 125 American cities, tuned until its homicide rate moves the way Baltimore's did from 2018 through 2021. In January 2022 the two lines part. The real Baltimore keeps falling. The made-up one falls too, just not as far.

Everything worth arguing about in Baltimore's homicide decline lives in the space between those two lines. That includes Malcolm Gladwell's claim, in an interview with the Baltimore Sun published October 17, that Mayor Brandon Scott is “front and center” for the credit. It also includes our own first reaction, which was that homicide is falling nearly everywhere and a mayor taking a bow for it is a mayor standing in front of the weather.

We went through the research. Our reaction holds for the country. It holds less well for Baltimore than we expected. And the best evidence for Baltimore's extra drop was written by people the city pays, about a program run by an office the city is currently fighting in court to keep its records away from its own inspector general. All three of those things are true at once.

The weather

Start with the country, because the scale matters. According to FBI estimates released in August and reported by Stateline, the U.S. murder rate fell 18.1% in 2025, to 4.1 per 100,000. That ties 1955 and 1956 for the lowest rate the FBI has ever estimated. Violent crime fell 9.3%, the largest one-year drop since national estimates began in 1936. No population group the FBI tracks recorded an overall increase in violent crime: not big cities, not suburbs, not rural counties. Preliminary figures for the first half of 2026 put murder down another 23%.

The national decline didn't stop at the border. Statistics Canada reported 672 homicides for 2025, down 16%, the biggest one-year drop since 1986. England and Wales recorded 522 homicide victims in the year ending March 2025, 8.6 per million people. Mexico's rate fell 19.8% in 2025, per InSight Crime's annual round-up, its seventh straight year of decline. None of those places has Brandon Scott.

“Worldwide,” though, is a word we had to give back. Eurostat counted 3,953 intentional homicides across the EU in 2024, up 1%. Australia is flat at a low level. Ecuador's homicide rate rose 31.2% to a record. And the American spike that the decline is unwinding was mostly ours: homicide in England and Wales fell during the 2020 lockdowns, and Canada's rose slowly to a 2022 peak instead of jumping in a single year. The fairer word is the English-speaking world plus Mexico, which is a less satisfying slogan and a more accurate one.

There's an older literature behind the instinct anyway. Jan van Dijk, Andromachi Tseloni and Graham Farrell edited a 2012 volume called The International Crime Drop, arguing that the long fall in Western crime since the nineties ran largely independent of any one government's policy. We've written about our own version of that idea in the criminogenic-entropy post and in why violence fell off a cliff. When crime falls on three continents at once, the local explanations have to fight for whatever is left over.

Forty-eight criminologists and a glass of milk

The Council on Criminal Justice tried to sort the explanations in September. It put 48 researchers in a room at Arizona State for two days, handed them briefs on 48 candidate causes, and polled them. The results came out this month. Only one factor, the easing of pandemic disruption, was rated a major contributor by a majority (57%). Community violence intervention was rated a major or limited contributor by 91%, focused deterrence by 83%. Twenty-six of the 48 factors drew support from fewer than a quarter of the room, among them National Guard deployments, immigration enforcement and increased incarceration.

Adam Gelb, the council's president, had already given the summary in August. The drop is historic, he said, but what's driving it is “about as clear as a glass of milk.” He also warned against handing credit to any one leader, party, policy or program before the evidence catches up.

Two notes on the report. First, it's a poll of informed judgment, and the council says so plainly: these aren't causal findings. Second, the question we care about most, whether a decline shared with Canada and Britain leaves much room for local programs at all, doesn't get its own factor. Cross-national comparison shows up only in the list of things to study next. In our view that's the biggest hole in the exercise. (One of the participants, Andrew Wheeler, whose writing we admire, said in his comments that he was uncomfortable crediting a factor without clearer empirical evidence. We'd have liked more of that.)

Baltimore didn't have a spike

This is where our instinct ran into trouble. The most common way to wave off a city's big decline is mean reversion: places that rose most in 2020 and 2021 had the most to give back. Philadelphia went from 356 homicides in 2019 to more than 560 in 2021. Its drop since then is largely a return trip.

Baltimore never left. In the NBER paper (Kapustin, Chalfin, Biddle, Wade, Khade, Layana, Struhl and Braga, “Reducing Gun Violence at Scale,” Working Paper 35292), Baltimore's annualized homicide rate averaged 57.9 per 100,000 in 2018 and 2019 and 60.0 in 2020 and 2021. That's a bump of about two. The city had been sitting near the 95th percentile of large American cities for most of six decades, and in 2021 its rate was more than eight times the national one. There was no pandemic spike for it to revert from.

Here is the count, as the police department has reported it:

Baltimore homicides
2019  348
2020  335
2021  337
2022  333   GVRS launches in the Western District (Jan)
2023  261   expands to Southwestern (Jan)
2024  201   Central (Jan), Eastern (Apr); restated to 194 in Dec 2025
2025  133   Southern (Jul)

2019 -> 2025:  (133 - 348) / 348 = -61.8%
2026 YTD thru Oct 1:  83 vs. 98  (-15.3%)

The authors then do the thing a skeptic would ask for. They regress each comparison city's 2022 to 2025 homicide change on its pre-2022 level. The slope is about 0.11, which at Baltimore's starting rate predicts a decline of six to eight per 100,000. Baltimore fell roughly eighteen. Only one of the 125 comparison cities had a bigger raw decline. Their synthetic Baltimore, the city in Figure 8, puts the real one about 25% below where it would otherwise have been over 2022 through 2025 (placebo p < 0.001), and they get nearly the same answer from a second pool of 83 large cities.

Inside the city, the first district to get the strategy is where the effect shows up first. In the 18 months after January 2022, shootings and homicides in the Western District fell by roughly a third relative to comparable Baltimore neighborhoods, carjackings by about 40%, with no sign the violence moved next door. Total arrests didn't rise. Arrests for serious violent crimes did.

So the national tide explains a lot of Baltimore's drop, and it doesn't explain all of it. Something specific to Baltimore happened. We came in expecting to write that it didn't.

Read the disclosure first

The first page of the NBER paper, before the abstract, is worth reading slowly. Five of the eight authors, through the University of Pennsylvania, support the city's implementation and expansion of the Group Violence Reduction Strategy. Three are currently paid under that contract. Anthony Braga, the senior author, is a principal investigator on the contracts and says he has drawn no compensation from them since 2022. The funders run from Bloomberg Philanthropies to the Ravens. The acknowledgments thank the mayor by name.

None of that makes the math wrong. The authors disclose it fully, which is more than most advocates do. But it's the city's own consultants grading the city's program, in a paper that hasn't been peer-reviewed. A few things in the paper itself deserve the same scrutiny the headline gets:

• The Western District effect alone would move the citywide number by only about 5%, the authors calculate. The rest of the citywide gap is attributed to the expansion and a “broader change in law enforcement culture,” which is harder to test.
• With a standard synthetic control instead of their preferred estimator, the citywide homicide effect shrinks to −4.6 per 100,000 and isn't statistically significant (p = 0.12). The authors explain why they discount that version, and the explanation is reasonable. It's still the estimate a skeptic would want to see.
• The cross-city comparison, in their words, can't isolate GVRS from anything else that changed in Baltimore. Ivan Bates took office as state's attorney in January 2023. The governor sent state resources. A Gonzales poll this winter found Baltimore voters split between crediting Scott (47%) and Bates (37%).

Our reading: the evidence that Baltimore beat the trend is reasonably strong. The evidence that one program did it is weaker. And the evidence that one man did it is a story.

What Gladwell got right, and what he didn't

Credit where it's checkable. Gladwell told the Sun the same strategy had been offered to Baltimore thirty years ago and fifteen years ago, and nobody followed through. That's accurate. The NBER paper describes Operation Safe Neighborhoods, launched in Park Heights in 1998 and dismantled amid infighting, and a 2014 to 2015 pilot in the Western District that ended after leadership turnover and Freddie Gray's death. It also says Scott engaged Penn's lab in June 2020, as the presumptive mayor-elect, and that sustained leadership from the mayor and police commissioner kept the routines in place. Fidelity is the authors' whole explanation for why this attempt worked when the first two didn't. If you want a single thing to give Scott, that's the honest one.

His description of the mechanism is also right, and it's worth noticing that it isn't what the Sun's introduction says. The introduction has him crediting Baltimore's strategy of sending “violence interrupters” into the streets. In his own answers, Gladwell describes focused deterrence: police and partners identifying a small number of people at risk, talking to them, and making clear the law will come down if they retaliate. Those are different programs. Baltimore's violence-interrupter program is Safe Streets, and the newest Johns Hopkins evaluation estimated 42% fewer youth homicides and 21% fewer youth nonfatal shootings near its sites, neither statistically significant.

Then there's the money. Gladwell told the Sun that the U.S. and Western Europe spend about the same share of GDP on public safety, but that “we're spending the bulk of the money on prisons.” The European half is close: Eurostat puts EU police spending at 0.9% of GDP in 2024 against 0.2% for prisons. The American half isn't. State and local governments spent $135 billion on police and $87 billion on corrections in 2021, according to the Urban Institute. Police win by about half again. His broader point, that America is thinly policed, has real support. The U.S. had roughly 233 sworn officers per 100,000 people in 2021 against about 340 across the EU, and the Lewis and Usmani paper behind the argument makes its case in officers per homicide. That's a headcount argument. Somewhere on the way to the book it became a spending one.

That move will be familiar to anyone who remembers the ten-thousand-hour rule. Anders Ericsson's 1993 violin study found that the best students had practiced more, on average. In a 2016 excerpt from his book Peak, Ericsson pointed out that half of the top group hadn't reached ten thousand hours at all. A 2014 meta-analysis by Brooke Macnamara and colleagues found practice explained about 21% of the variation in musical performance and under 1% in the professions. A real finding, sanded down until it fits on a cover. The Baltimore result is real too. “Mayor Scott deserves the credit” is the sanded version.

The office getting the credit is the office in court

The Group Violence Reduction Strategy is run out of the Mayor's Office of Neighborhood Safety and Engagement, known as MONSE. In March, Inspector General Isabel Mercedes Cumming reported on SideStep, a MONSE youth-diversion pilot in the same Western District. The city had paid about $694,799 to community organizations. Cumming found invoices that appeared falsified, and a 2023 email in which a MONSE employee sent a table of more than 700 names, many of them juveniles, to a relative's personal Gmail. Only 24 of 51 participant files she reviewed had case notes. She referred both matters to law enforcement. MONSE said it was deeply concerned and that the report lacked the details it needed to act.

The records fight grew out of that review. The city's Law Department heavily redacted the MONSE records Cumming had subpoenaed and argued her access should be limited to what the public could get under the state's public-records law. She sued. On August 14, Judge Pamela White denied the city's motion to dismiss, writing that the inspector general's office is not controlled by and does not answer to the mayor. In September she granted Cumming summary judgment and ordered access restored. The mayor's office responded by accusing Cumming of breaching privilege “for personal motives,” without naming the records. The city has appealed. As of October 15, Cumming said about 67 subpoena requests were outstanding and her office had not received a single document. The court's deadline is November 5.

We have some history with this. SpotCrime is a Baltimore company. In June 2025 we noted that when the city moved its open crime feed to NIBRS, the “Shooting” category disappeared. Shootings now sit inside aggravated assault with a weapon field that doesn't say whether a gun was fired. The city was announcing drops in shootings the same season the public lost the simplest way to check. (We wrote about how this happens in general in the Phoenix records-migration post.)

We think the homicide decline is real. Homicide is the hardest crime to make disappear, and Maryland's medical examiner is a state office, not a city one. But the case for crediting a particular office rests on that office's records: who was contacted, which services were delivered, what was paid for. A city that wants the credit should be the first to open those files. Baltimore is spending money on lawyers to keep them closed.

So why is it down?

Nobody knows to a decimal place, and anyone who says otherwise is selling a book. The best reading we can give is three layers, from the widest down.

The long decline underneath.Homicide in America has been falling since the mid-nineties, and the 2020 to 2021 surge was an interruption, not a reversal. Some of what's pushing it down is structural and shared with Canada and Britain. Fewer young people are entering crime at all. The CCJ report puts estimated U.S. arrests at 7.6 million in 2025, about 25% below 2019, and cites cohort research showing each generation arriving at adulthood with less criminal involvement than the last. Teenagers spend less unsupervised time together, and they drink less. No mayor did that.

The American spike unwinding.Whatever made 2020 uniquely violent here, whether pandemic disruption, police pulling back after George Floyd's murder, or local governments losing staff, mostly reversed. Local government employment was back to its pre-pandemic level by September 2023, helped by $350 billion in American Rescue Plan money. That explains much of 2022 and 2023 and less of 2024 and 2025, when rates fell well below 2019.

Targeted work in a few high-homicide cities. Homicide is concentrated, so a handful of cities can move the national number. Brandon del Pozo of Brown put it well in the CCJ report: community violence work probably mattered in a handful of places that drove the national figures, not as a broad national force. Baltimore is plausibly one of those places, and the GVRS evidence, conflicts and all, is the best evidence any city has.

So were we right? About the country, yes. A decline that shows up in Canada and England isn't a policy victory for anyone in particular, and the credit-taking around it (from mayors, prosecutors, governors and presidents) deserves the skepticism we came in with. About Baltimore, partly. The city did better than its peers, by a margin that survives the obvious objections, and the sustained adoption of a strategy that failed twice before is a real thing Scott's administration did. Whether it was the strategy, the prosecutor, the commissioner, the state or the mayor is a question the city could help answer, if it would hand over the files. (We've written before about how small numbers lie. A city of 133 homicides is still big enough to measure.)

The synthetic Baltimore got safer too, by the way. It's made of 125 other cities, with their own mayors and their own programs, and its homicide rate still fell. Nobody gets to take credit for that line. Everywhere else, it's most of the story.