FARS vs. CRSS

Quick Answer

Side-by-Side Comparison

The table below summarizes the most important structural differences between the two programs.^1,2,3,4,5

FARS in Detail

NHTSA launched FARS in 1975 to give federal safety researchers and the public a complete, annual count of motor vehicle traffic fatalities in the United States.^1,4 A crash qualifies for FARS only if it occurs on a public trafficway and results in the death of at least one person — a vehicle occupant, pedestrian, or cyclist — within 30 days.¹

State FARS analysts collect data from multiple sources: police crash reports, hospital records, death certificates, and medical examiner or coroner files.¹ This multi-source approach gives FARS more accurate BAC data and injury outcomes than systems that rely only on the first police report. The result is a census — every qualifying fatal crash appears in the dataset, so counts are exact, not estimated.^1,8

FARS publishes a national CSV package containing crash-level, vehicle-level, person-level, and factor-level tables. The key crash table is accident.csv (or accident.CSV in older packages), and the key person table is person.csv.^6,8 State-level fatal crash counts and fatality rates — including the tables on this site — are derived from those tables by grouping records using the STATE FIPS code and summing the FATALS field.^6,8

CRSS in Detail

CRSS replaced the National Automotive Sampling System General Estimates System (NASS GES) in 2016.^2,3 Where FARS is a census of fatal crashes, CRSS is a stratified probability sample designed to be nationally representative of all police-reported crashes — from minor fender-benders to serious injury and fatal crashes.²

NHTSA selects approximately 60 Primary Sampling Units (PSUs) — counties or groups of counties — and then randomly samples police crash reports filed within those PSUs.² A weighting scheme scales each sampled crash to an estimated national total. The published CRSS files contain a case weight variable that researchers must apply to produce nationally representative estimates.^2,7 Because CRSS is a sample, all published figures are estimates with associated coefficients of variation, and small subgroup estimates can have wide confidence intervals.²

CRSS data collection relies entirely on police crash reports. Unlike FARS, there is no secondary check against medical examiner or hospital records.²This means CRSS alcohol and drug fields reflect officer-reported involvement rather than toxicology results, and injury severity is coded from the officer's initial assessment rather than medical records.²

When to Use Each Dataset

The datasets answer different research questions. Mixing them — for example, using a CRSS estimate as a denominator for a FARS fatality count — can produce misleading rates because the numerator and denominator come from programs with different scopes and methodologies.^1,2

Common Reporting Mistakes

Media outlets, advocacy groups, and researchers regularly confuse the two systems. The five patterns below appear most often.^1,2

How Daily Driver Advocate Uses Them

All fatality statistics on this site — state rankings, national totals, drunk-driving fatality counts, and per-capita rates — are derived exclusively from FARS.^1,4,6FARS is the appropriate source because the site's core questions are about fatal crashes, and only FARS provides exact, census-grade counts.

Census state population estimates serve as the denominators for per-capita rates. Population denominators are not drawn from CRSS.

CRSS data is not currently incorporated into site visualizations. When future pages address non-fatal crash patterns — distraction prevalence across all crashes, for example — they will use CRSS and will clearly distinguish CRSS-derived estimates from FARS-derived exact counts.^2,5,7

Important Limitations of Both Systems

• FARS captures only fatal crashes. Injury-only and property-damage-only crashes — the overwhelming majority of police-reported crashes — are outside its scope.
• CRSS produces estimates, not counts. All CRSS figures carry sampling error. Small subgroup estimates (e.g., crashes involving a specific vehicle type in a single season) may have coefficients of variation high enough to make the estimate unreliable.
• Both systems rely on police reports as a primary input, which means crashes that are not reported to police are outside both systems.
• CRSS state-level estimates are not reliable for state comparisons. The sampling design is calibrated for national and broad regional estimates, not state-by-state rankings.
• CRSS has no BAC toxicology data. Alcohol involvement in CRSS is officer-reported, not laboratory confirmed. For impairment analysis backed by toxicology, FARS is the appropriate source.
• NASS GES data (1988–2015) and CRSS data (2016–present) are not directly comparable without methodological adjustment, due to changes in sampling design between the two programs.

Suggested Citations

When citing either dataset, name NHTSA as the issuing agency and link to the official source. When citing this explainer, name Daily Driver Advocate and note it is an independent research summary of the official NHTSA program documentation.