How we count

Every count on this site comes from one public file: NHTSA's Fatality Analysis Reporting System. The rates per resident and per mile add two more public sources, from the Census Bureau and the Federal Highway Administration. This page lists every rule applied on the way from those files to a page.

The source: NHTSA's Fatality Analysis Reporting System

The Fatality Analysis Reporting System (FARS) is the US Department of Transportation's census of every crash on a public road in which someone died within 30 days. The National Highway Traffic Safety Administration (NHTSA) compiles it from police reports, death certificates and state records, one file per year since 1975. The files are public domain and can be downloaded from NHTSA.

This site uses the 2024 Annual Report File, released April 2026, the latest year NHTSA has published; the 2024 figures are from the Annual Report File and may be revised when NHTSA releases the final file. The 2025 file, with the final 2024 counts, is expected in spring 2027. State and national lines since 1975 use every yearly file; roads, counties and shares use 2015 to 2024, because road names are patchier in older files.

What a fatal crash is, and what a death is

A fatal crash is one crash, however many vehicles or people were in it, in which at least one person died within 30 days. Deaths are the people who died in it. A crash with three deaths counts once in a crash table and three times in a death table.

Most figures on this site are counts, or shares of fatal crashes or of deaths: FARS records crashes, not how much people drive, so a state with many crashes may simply have many drivers. The only rates are the three in the next section, and each page that shows one names its source.

The rates per resident and per mile, and where their other half comes from

Deaths per 100,000 residents: FARS deaths divided by the Census Bureau's population estimate for the same year (vintage 2024 for 2020 on, vintage 2019 for 2015 to 2019). A ten-year rate divides the ten years' deaths by the ten years' populations added up. Where a year has no estimate, there is no rate: Connecticut's counties have none after 2019, because the Census Bureau now publishes Connecticut by planning region.
Deaths per 100 million vehicle-miles: FARS deaths divided by the miles driven in the state that year, from FHWA Highway Statistics, table VM-2.
Fatal crashes per 100 miles of interstate: the ten-year average of fatal crashes a year on an interstate in one state, divided by its length from the FHWA Interstate route log, table 1, for stretches of 20 miles or more. It is per mile of road, not per vehicle: it says where crashes cluster, not how risky a trip is. Two route-log rows were plainly swapped (I-75 and I-81 in Tennessee) and are corrected by hand.

A rate needs 10 or more deaths under it. Rates per resident and per mile can rank a state differently: a rural state with long drives has more miles per resident.

How roads are named and ranked

Only numbered interstate, US and state routes are ranked and get pages, read from the road name NHTSA filed. County roads, city streets and unnamed roads stay in every state and county total but not in road tables. Business routes, loops and spurs named after a parent route ("US-83 Business") are left out of the parent's count.
NHTSA files the road name as free text. We read the route at the start of it: "I-35", "I-35/PURPLE HEART TRL" and "I 35" are the same road. The spellings each road page was built from are listed at its foot.
Crashes NHTSA filed on a ramp, frontage road or service road of a numbered route are left out of the road's count ("Left out" on the state page, with the number).
Where an interstate splits into lettered branches (I-35E and I-35W around Dallas and Fort Worth and around Minneapolis and St Paul; I-69E, I-69C and I-69W in South Texas), a crash filed under the plain number on a branch is counted on that branch. The county decides where it holds one branch only; elsewhere the crash's coordinates put it on the nearest branch.
Where two routes share a road (US 59 and I-69 through Houston, for one), NHTSA files each crash under one route, so the shared stretch counts on one route only and the other route's count is lower than its length would suggest.
Roads and counties are ranked by how often a fatal crash happens there, with the people killed beside each count; states, years and days of the year are counted in people killed. Click a column head to rank a table the other way.
A county gets its own page with 30 or more fatal crashes from 2015 to 2024. Smaller counties stay in their state's county table and in every total.
A road gets its own page with 15 or more fatal crashes in the state from 2022 to 2024. Roads are counted within a state; an interstate that crosses several states has a page per state.

The floors under every table, headline and share

Lists: a road or county appears in a ranking table from 5 fatal crashes in the window.
Headlines: a road or county only tops a ranking, holds a sentence ("the deadliest road in Vermont") or answers a question from 10 fatal crashes. Below that, the heading changes to say what the table shows without naming a worst.
Shares: a share (alcohol, speeding, dark, pedestrians) is only given over 20 crashes or deaths. Below that the page says "Not enough data".
Above or below: a county or road is marked above or below its state's share only when the gap is more than a point and more than chance would give on that many crashes (twice the standard error of the state's share at the county's size). Smaller gaps read "about the same".
Ties: equal values share a rank, and the next rank is skipped (1, 2, 2, 4), on every page that ranks them.
Changes on the year before: under 20 deaths the page gives the two counts ("down from 10 in 2023") instead of a percent, which would swing wildly.

Why totals here differ from state and news figures

A state highway department counts from its own crash reports, often within weeks, and a news report quotes whichever count was out that day. NHTSA's file comes a year or more later, after every death within 30 days is known and each case is checked against police reports, death certificates and medical records. NHTSA also publishes early estimates for a year before the file; news from that year often quotes those. So a state total here can differ from the state's own by a few deaths, and a year page from the news of that year by a few hundred. Road totals differ more, because this site leaves out ramps and frontage roads and counts only crashes filed under the road's number.

Four labels for what is missing, and what a zero means

A zero on this site is always a count of nothing: a year in which no fatal crash was filed on a road is a 0, and a real one. Missing figures are never shown as 0; they carry one of these labels.

No data: the source has no figure: no Census population for the place that year (Connecticut's counties after 2019), no FHWA mileage for a route, no year before the first year.
Not reported: NHTSA coded the field as unknown or not reported. We never guess.
Not enough data: the figure exists but fewer than 20 crashes are under it, so the share would mean little.
Left out: the row is in the file but removed by one of our rules (ramps, frontage roads, unnumbered routes), with the reason and the count printed where it would have been.

In the CSV files an empty cell means no figure, for any of these reasons; a 0 is a real zero.

What the factors mean

Drunk driving deaths (alcohol-impaired): NHTSA's own measure, and the one the drunk driving page ranks by: people killed in crashes where a driver had a blood alcohol concentration of .08 or more. Many drivers are never tested; NHTSA estimates their level ten times over from the crash circumstances (multiple imputation) and puts the ten estimates in the FARS file. We average them as NHTSA does, which reproduces its published total (11,904 deaths, 30%, in 2024).
A drinking driver: NHTSA's drinking-driver flag for any driver in the crash: police-reported drinking, or any positive alcohol test (.01 or more), as a share of fatal crashes. No estimate is involved, so it is a floor, and it is the measure on state, county and road pages.
Speeding: any driver coded as exceeding the limit, too fast for conditions, racing, or speeding with specifics unknown.
Unbelted: occupants of passenger vehicles (cars, SUVs, vans, pickups) killed and coded as using no restraint, as a share of those whose belt use is known. Motorcyclists, and vehicles without belts such as ATVs, are not in the base.
Hit-and-run: any vehicle in the crash coded as having left the scene.
A distracted driver: any driver in the crash with a distraction or inattention recorded in NHTSA's driver-distraction table. Drivers coded "not reported" or "unknown" (half of all drivers since 2019) count as not distracted, so the share is a floor, and it depends on how fully each state's police record distraction. NHTSA calls distraction under-reported.
In the dark, road unlit: light condition "Dark, not lighted". Dusk, dawn, lit roads and "dark, lighting unknown" are separate and stay in the base.
Rural: NHTSA's rural or urban code for the crash location.
First harmful event: what the crash first hit: another vehicle in motion, a pedestrian, a tree, a rollover.
Vehicle types: NHTSA's body type codes grouped into cars, pickups, SUVs, vans, motorcycles (including mopeds and scooters), truck-tractors, other trucks and buses. Off-road vehicles, golf carts and farm equipment are counted in totals but never ranked.

When, where and who

Hour, weekday and month are as NHTSA records them. The weekday-by-time grid pools 2015 to 2024 so small states have enough crashes per cell. Who died comes from the person file: pedestrians, cyclists (bicyclists and other pedalcyclists), motorcyclists (occupants of a motorcycle, moped or scooter), occupants of other vehicles, and others (people on personal conveyances, in parked vehicles or in buildings). The five rows add up to every death.

What Fatal Crash Data leaves out on purpose

No names, no crash narratives, no addresses or coordinates of individual crashes, no photos. FARS is a public record of crashes, and this site shows it as counts, shares and the three rates above. No risk scores: FARS has no record of how many vehicles use a particular road, so a "risk of driving this road" number cannot be computed honestly.

When each page last changed

The foot of every page gives the date its figures last changed. A rebuild that changes nothing on a page leaves its date alone.

Corrections to a figure

If a figure looks wrong, tell us with the page and the number, and we check it against the file. The build is one set of scripts that run from the raw file to every page, so a fix applies everywhere.