In October 2025 the National Highway Traffic Safety Administration published a fact sheet reporting that 40,901 people were killed on American roads in 2023. In April 2026 the same agency published a research note reporting 41,025 people killed in 2023. Neither number is a correction of the other and neither is an error.
The reason is a mechanism the agency documents in the back of its own publication, and understanding it is the entry price for using this data at all. The Annual Report File is the version of the Fatality Analysis Reporting System released for the most recent available year. A year later it is replaced by the Final File, and the interval exists so that variable data requiring outside sources can be submitted, which can change the counts. The October fact sheet carried the 2023 Annual Report File. The April research note carried the 2023 Final File.
The same publication supplies the previous instance of the same adjustment, which is the best evidence of how large it can be. The final 2022 count was 42,721, updated from 42,514 in the 2022 Annual Report File. That is 207 additional deaths appearing in a national total two years after the year ended.
What has to be true for a death to appear
FARS holds records on every fatal motor vehicle traffic crash in the fifty states, the District of Columbia and Puerto Rico. Two tests decide inclusion. The crash must involve a motor vehicle travelling on a trafficway customarily open to the public, and it must result in the death of a vehicle occupant or a nonoccupant within thirty days of the crash.
Both tests exclude real deaths. A crash in a parking lot or a driveway not regularly used by the public for transport is not a motor vehicle traffic crash, so a child killed in a private driveway is absent from the number politicians cite. Someone who dies of crash injuries on day thirty-one is absent as well. The thirty-day rule is an international convention rather than a clinical judgment, and it is what makes the series comparable across decades, which is a real benefit paid for with a real omission.
Geography adds one more wrinkle worth knowing before quoting a state table. Puerto Rico is included in the FARS database and excluded from the United States total, which reported 307 deaths there in 2023 and 288 in 2024.
The counts of fatalities and fatal crashes are actual counts. Injuries and total crashes are estimates from a sample survey, the Crash Report Sampling System. NHTSA marks the boundary in its statistical language: changes in fatal crash counts are not tested for statistical significance, because they are a census. So when a research note says that a 0.8 percent decrease in injuries was not statistically significant while reporting a 4.3 percent fall in deaths without qualification, that is not inconsistency. Two different kinds of number are sitting in the same table.
Why the recovery from 2021 is still incomplete
The series that matters starts before the disruption. Using the final files compiled in the 2023 fact sheet, deaths ran 36,835 in 2018 and 36,355 in 2019, then 39,007 in 2020, 43,230 in 2021 and 42,721 in 2022. The 2023 and 2024 figures from the later research note are 41,025 and 39,254.
So the peak was 2021, the decline has now run three consecutive years, and 2024 was the first year below 40,000 since 2020. It was also still about 8 percent above 2019. The recovery is real and it is not complete, and any account that stops at either half of that sentence is selling something.
Fatal crashes moved with the deaths but not identically: 37,769 in 2023 against 36,297 in 2024. Injuries are flatter, at an estimated 2.44 million then 2.42 million, with the injury rate per 100 million vehicle-miles moving from 75 to 74.
One year, two denominators, two verdicts
The most useful thing in this data is not a number but a disagreement between two normalisations, and 2020 is where it shows.
Deaths rose from 36,355 in 2019 to 39,007 in 2020, an increase of 7.3 percent. Travel fell over the same year from 3,261,772 million vehicle-miles to 2,903,622 million, a decrease of 11.0 percent. Divide deaths by travel and the fatality rate per 100 million vehicle-miles jumps from 1.11 to 1.34, a rise of about 21 percent. Divide the same deaths by population instead and the rate per 100,000 people goes from 11.07 to 11.77, a rise of 6.3 percent.
Both figures are published in the same table, computed from the same deaths. The mileage rate says 2020 was a catastrophe, because the roads got radically more dangerous per mile driven. The population rate says 2020 was a bad year in a bad decade. The mileage rate is the right instrument for asking whether driving became more dangerous. The population rate is the right instrument for asking whether more people died.
By 2024 the mileage rate had come down to 1.19 on travel of 3,294,031 million miles, against a decade low of 1.08 in 2014 and a decade high of 1.38 in 2021. Five years after the disruption the per-mile risk is still above where it was before, on more driving than there was before.
The decline is not evenly distributed
One long-run shift cuts across everything above. The proportion of people killed inside a vehicle has fallen from a high of 80 percent in 1996 to 62 percent in 2024, and the proportion killed outside one, meaning motorcyclists, pedestrians, cyclists and other nonoccupants, has risen from 20 percent to 38 percent.
The most recent year shows why that trend is durable. Total fatalities fell 4.3 percent in 2024, and nonoccupant deaths fell 3.7 percent, so the outside-the-vehicle share held. But the injury figures went the other way: nonoccupant injuries rose 5.7 percent and motorcyclist injuries rose 5.2 percent, both statistically significant increases, while passenger car occupant injuries fell 4.6 percent. Vehicle crashworthiness improvements accrue to the people inside the vehicle, and the countermeasures that protect everyone else are geometric rather than mechanical, which is the argument running through intersection conversion evidence.
Rural and urban roads have also diverged. Urban fatalities have exceeded rural fatalities every year since 2016, and between 2015 and 2024 urban deaths rose 30 percent while rural deaths fell 9.6 percent, ending at 22,905 and 16,006. Rural roads remain disproportionately lethal on the travel they carry: in 2023 rural areas held 20 percent of the population and 31 percent of vehicle-miles, and produced 41 percent of the fatalities. Read that alongside the classification caveat NHTSA attaches, which is that rural and urban categories were updated in 2022 FARS, changing previously reported figures.
Where the data stops being able to help
The behavioural categories are widely quoted as causes and they are not causes. They are coding categories with definitions worth reading.
A traffic crash is speeding-related if any driver was charged with a speeding-related offence, or if an officer indicated that racing, driving too fast for conditions, or exceeding the posted limit was a factor. Nothing in that requires a measurement. On that definition, 11,288 deaths in 2024 were speeding-related, 29 percent of the total. Alcohol-impaired driving deaths were 11,904, also 30 percent, and NHTSA imputes blood alcohol concentration where test results are unknown. Distraction-affected crashes account for 3,208 deaths and drowsy-driver crashes 644, and restraint use was unknown for 11 percent of passenger vehicle occupants killed.
Two definitional seams also catch out comparisons. Vehicle type classification changed in 2020, when NHTSA began decoding vehicle identification numbers against the manufacturer’s own body class, so 2020 and later categories are not comparable to 2019 and earlier for any analysis with a vehicle component. And from 2022 motorized bicycles are no longer collected as motor vehicles, so their riders count as pedalcyclists rather than motorcyclists, and a crash involving only motorized bicycles leaves the database entirely. A decade-long chart of motorcyclist deaths crosses both seams.
None of this explains 2021. The publications record what happened with unusual precision and they decline to adjudicate why, and the honest position is that the causal account remains contested. The data can rule one thing out. Reduced exposure cannot be the mechanism, because travel fell in 2020 while deaths rose, and travel had recovered past its 2019 level by 2024 while deaths stayed above 2019. Beyond that the record offers correlated categories rather than a cause, and NHTSA does not claim otherwise.
The scale of what is riding on that gap is itself published. NHTSA puts the economic cost of all traffic crashes in 2019, the most recent year with cost data, at $340 billion, and the total value of societal harm including quality-of-life valuations at $1.4 trillion, with an average discounted lifetime economic cost of $1.6 million and an average comprehensive cost of $11.3 million per fatality. The federal data holdings behind those figures are mapped in FHWA data sources, the vehicle-specific series in large-truck crash data, and the countermeasure literature in highway safety coverage.
Which returns the argument to the choice of denominator, because that choice is where policy actually gets made. An agency reporting per-mile risk is answering a question about the road. An agency reporting per-capita deaths is answering a question about the population. A program evaluated on the first can succeed while more people die, and a program evaluated on the second can fail while every mile driven gets safer. NHTSA prints both columns and takes no position on which one an argument should use. Everybody quoting the data does take one, usually without saying so.