We graded 13,216 American cities with FBI data. Here's what we found.

SafeNest ships a database of 13,216 American cities, each carrying a letter grade built from FBI crime data. When we finished it, the first thing we did was look at the distribution — and the first thing the distribution said was something we had to sit with for a while.

57.1% of American cities get an A. Only 8.4% get a D or an F.

That looks like grade inflation. It isn't. It's what national-relative grading does to a country whose municipalities are mostly small — and understanding why is the difference between reading our grades correctly and reading them badly. Here's what the full set actually contains.

Where the anchor is

The grade is not a curve. Nobody is graded against the other 13,216 cities in the file. Each city is measured against a fixed national baseline: about 381 violent and 1,954 property offenses per 100,000 residents, weighted 70% violent and 30% property.

That fixes the anchor, and the anchor is the whole story:

Once the anchor sits at the national average, the distribution follows mechanically. The national average is population-weighted: it is dragged upward by large cities, where most Americans live and where most crime is reported. But most municipalities are not large. The median city in our file has 6,723 residents and a violent crime rate of 139.9 per 100,000 — well under half the national figure. A country of small towns, measured against an average set by its big cities, produces a lot of A's.

So the 57.1% is real, and it is not a compliment we paid ourselves. It is a fact about American municipal geography, showing up in a letter.

The part that trips people up is the C. A C sounds like the middle of the pack — the average city, the median result. It is the grade an exactly-average city earns, but it is nowhere near the middle of the distribution:

• A — 7,547 cities (57.1%)
• B — 2,572 cities (19.5%)
• C — 1,988 cities (15.0%)
• D — 809 cities (6.1%)
• F — 300 cities (2.3%)

A C city has 76.6% of cities grading above it and 8.4% below. If you are looking at a C and reading it as "average, therefore fine," you are reading it wrong. It means average for the United States — and the United States average is set in places with a great deal more reported crime than the typical American town.

The 15× separation

A distribution tells you how the letters are spread. It doesn't tell you whether they mean anything. So we tested the grades against a year of data they had never seen: assign each city a grade from one year's FBI data, then go look at what actually happened in that city the following year.

The bands hold their order, every year:

Stated the way the underlying study states it: city crime grades assigned from one year's FBI data correctly rank next year's crime — cities graded A have a median violent crime rate 15× lower the following year than cities graded F. In the pooled data that comes to 15.4×: an F city's median is 1,037.6 violent crimes per 100,000 against an A city's 67.4. The median rises monotonically from A to F, and it did so in all 21 reporting years we tested. This is the strongest thing the grade has going for it, and it is the reason we ship a letter at all.

Now the part that usually gets left out of a chart like this. The bands separate on average. They do not partition. Look at the tails: the 90th percentile of the A band (189.9 per 100,000) sits well above the 10th percentile of the B band (61.7). Every adjacent pair overlaps that way. A specific A city and a specific B city can trade places, and often do.

What that table supports is a statement about groups: the A group and the F group are genuinely, hugely different populations of cities. What it does not support is a statement about any pair: that this A city is safer than that B city. The grade sorts the country. It does not adjudicate between neighbours on the list.

The large cities with the lowest reported crime

Among the 92 cities of 250,000+ residents that have their own municipal police department, these have the lowest reported violent crime rates in the file:

One note on how that list was built, because it changes the answer. The FBI's data is organised by agency, not by city, and county-wide agencies sit in the file next to municipal ones. Rank the raw data and you get county police and sheriff's departments at the top — a county's average is diluted by everything that isn't its urban core. 48 such agencies appear at this population size. Nassau County, New York would have placed seventh on this list, and Nassau County is not a city. They are excluded above. It is the kind of error that makes a chart look great and be wrong.

An A in a town of 370 means something different

This is the finding we most want a reader to leave with, because it is the one most likely to mislead someone in practice.

There is no population floor. The smallest city we grade has 5 residents. Cities under 5,000 people are 5,654 of the file — 42.8% of every grade we publish. And at that size, the arithmetic of a crime rate stops behaving:

Read the last two columns together. Of the 364 cities under 500 people, 69.5% reported zero violent crimes, and 69.0% of the bucket is graded A. In a town of a few hundred people, reporting no violent crime in a year is the expected outcome. It is not evidence of anything in particular. And a single incident — one assault, one robbery, in one year — would put that same town at roughly 269.9 per 100,000, which is most of the way to a C.

So: an A in a town of 370 people and an A in a city of several hundred thousand are not the same statement. The second is a rate computed from hundreds of incidents, and it is stable. The first is a rate computed from zero or one, and it would change completely on the strength of a single police report. Both print the same letter. If you are looking at a small town, look past the letter to the incident count underneath it — and treat the letter itself as close to no information.

Three more things in the data

The grade is fragile at the boundary — often on one incident

We asked the file a blunt question: how many cities would print a different letter if their violent crime count changed by one? The answer is 1,868 cities — 14.1% of the database.

Almost all of them are small, which follows from everything above. But 8 have populations over 50,000, and 3 of those sit exactly on the A threshold: Canton Township, Michigan; Rowlett, Texas; Rockwall, Texas. One additional reported assault in any of them, and the letter on the page becomes a B. Nothing about those places would have changed.

This is a real property of the grade and not a rounding artefact we can engineer away — every threshold system has a boundary, and a boundary always has cities sitting on it. It is a good reason to treat the difference between adjacent letters as noise, and the difference between distant letters as signal.

States differ enormously — and it is not a ranking of states

Across the 45 states with at least 50 graded cities, the share of cities earning an A runs from New Hampshire (87.7%), Kentucky (84.1%), Maine (83.2%) at the top to South Carolina (17.2%), California (18.3%), Arkansas (26.7%) at the bottom. That is a spread of 70.5 percentage points, and it is tempting to read it as a league table of state safety.

Don't. The A-share of a state is partly a fact about that state's crime and partly a fact about how its municipalities are drawn — a state incorporated into hundreds of small townships will bank a lot of A's for the reasons described above. That said, the effect is weaker than we expected: the rank correlation between a state's A-share and the median population of its cities is only -0.145. Municipal size explains some of the spread, not most of it. Both things are in the data, and we can't cleanly separate them here, so we're reporting the number and the caveat together rather than the number alone.

Why there is no "most dangerous cities" list here

1,109 cities grade D or F — 8.4% of the file. We are not going to rank them for you, and the reason is in the rest of this article rather than in our manners.

A SafeNest grade is a city-wide average of crime reported to the FBI by that city's police department. It describes the municipality as a whole. It does not describe a street, a block, or an address, and crime is not distributed evenly inside a city — it is concentrated, everywhere, in every city. A list of low-graded cities invites exactly the reading we have spent this article arguing against: that the letter on a city tells you what will happen to a person on a particular corner of it. It does not. Publishing the D and F counts is reporting the data. Ranking them would be dressing up a city-level average as a verdict on the people who live there.