FARS: What a Fatality Census Cannot Support

Most people who cite this dataset get the first thing right. It is a census and not a sample, so its counts are counts rather than estimates with error bars around them. Then they stop, and the stopping is the problem, because a census is an ambition about coverage rather than a guarantee of it. What a census actually contains is decided by its inclusion rule, and this one has three clauses.

The rule, in the agency's words

The analytical user's manual states the qualifying condition in a single sentence, and every limit discussed below is already in it:

To qualify as a FARS case, the crash had to involve a motor vehicle traveling on a trafficway customarily open to the public and must have resulted in the death of a motorist or a non-motorist within 30 days of the crash.

NHTSA, Fatality Analysis Reporting System Analytical User's Manual, 1975-2024 (DOT HS 813794)

Read as three filters rather than as one sentence, that is a motor vehicle, a public trafficway, and a death inside thirty days. Each one removes a population that somebody, eventually, assumes is in the file.

Filter one: a motor vehicle

The least troublesome of the three, and still worth stating, because the question people usually want answered is about road danger rather than about motor vehicles specifically. A fatal collision involving no motor vehicle at all is outside the system by definition. The census counts what it says it counts.

Filter two: a trafficway customarily open to the public

This is the filter that surprises people, and its consequences are large. A fatal crash in a private car park, on a driveway, on private industrial land, or on any surface not customarily open to public travel does not qualify, however ordinary the vehicle and however real the death.

The strongest evidence that this excludes a substantial population is that the agency built a separate system to cover it. The same publications library that holds the FARS manual also holds an analytical user's manual for Non-Traffic Surveillance, which exists to describe non-traffic crash injuries and fatalities. Two systems are maintained because one boundary line runs between them.

The practical consequence: any question phrased as fatalities involving vehicles is answered incompletely by this dataset alone, and the shortfall is not distributed randomly. It sits disproportionately in the categories that happen on private ground, which includes a good deal of low-speed and reversing incidents.

Filter three: death within thirty days

The sharpest edge, because it is a bright line drawn across something continuous.

A person who dies of crash injuries on the twenty-ninth day is a fatality in this file. A person who dies of the same injuries on the thirty-first day is not. Nothing about the crash differs. What differs is which side of an administrative threshold the death fell on, and the threshold exists so that a file can be closed and published on a schedule rather than staying open indefinitely.

Two things follow. Any comparison against a source using a different survival window is comparing different quantities, and several international road death series use exactly that kind of differing window, which is why cross-national comparison of road fatalities is harder than it looks. And the count is systematically lower than a count of deaths caused by crashes, by however many people die on day thirty-one and after. That difference is unknown rather than small, and the file cannot tell you its size, because the cases that would answer the question are the ones it declines to record.

Three files, three units, and the denominator error that follows

Beyond inclusion, the structure of the data invites a specific mistake. The manual's contents describe an ACCIDENT file, a VEHICLE file, and a PERSON file, which is to say three different units of analysis in one dataset.

One crash may involve several vehicles and several people. A row count is therefore meaningless until you have said which file it came from, and a rate built by dividing a person-level numerator by a crash-level denominator will be wrong in a direction that depends on how many occupants were involved. This is the most common error made with the data and it is entirely avoidable by naming the unit before computing anything.

What the coding manual being annual actually means

Coverage runs from 1975 to the present across the fifty states, the District of Columbia, and Puerto Rico, and raw files for the whole period are downloadable. That length is genuinely valuable and it is also a trap.

The reason the analytical manual spans 1975 to 2024 in a single document is that coding practice has evolved over that period, and the manual exists to document the evolution. NHTSA additionally publishes a coding and validation manual for each individual year of collection. A data element in a 1985 file and the same-named element in a 2024 file are not automatically the same measurement.

So a fifty-year series is available and a fifty-year trend is a claim requiring work. The work is reading what changed for the elements you are using, in the years you are crossing, and the documentation to do it is published. Skipping that step produces trend lines that are partly measuring the history of a coding scheme.

What it supports well

Within its rule the file is excellent, and the limits above are not reasons to avoid it. It supports counts rather than estimates. It supports deep breakdowns, because the vehicle and person files carry a large number of coded elements. It supports state and long-period analysis better than almost any alternative.

What it does not support is the sentence most people want to write, which is a claim about how many people are killed by vehicles. It answers a narrower question with precision, and the narrowness is documented rather than hidden. A reader who quotes the number without the rule has borrowed the precision and left the definition behind.

The companion guides in this section cover the federal commercial vehicle sources that draw on this file, and the Texas state collection whose inclusion rule is drawn in a different place entirely. For the court records material, the rest of this publication is the way across.