Key facts
- 220,730 of 427,908 US vehicle complaints filed 2020-2026 (51.6%) record a usable speed at failure. The rest leave the field blank.
- 30.03% of those failures happened at 55 mph or more; 22.72% happened at 0 mph, with the vehicle not moving.
- Only 30.67% fall in the 16-45 mph band. The median reported speed at failure is 35 mph.
- 52.75% are therefore either at highway speed or at rest - the two conditions a short dealer test drive is least likely to reproduce.
- The link between component and speed is sharp: adaptive cruise control 70.0% at 55 mph or more, against back-over prevention 2.6%.
- Counter to the obvious reading, tow and injury rates peak at 16-45 mph, not at highway speed.
What is this field, and why has nobody counted it?
NHTSA's Office of Defects Investigation publishes every consumer complaint as a flat file. Field 32 of that layout is VEH_SPEED, the vehicle's speed when the problem occurred. It is not exposed by NHTSA's public complaints API, it is blank on just under half of all complaints, and it is absent from every commercial vehicle-history and market-data product we are aware of. So the distribution of speeds at which American cars actually go wrong has been sitting in a public file, unaggregated.
We read it across the two current period files, 2020 through October 2026.
At what speed do cars actually fail?
| Speed at failure | Complaints | Share | Cumulative |
|---|---|---|---|
| 0 mph (stationary) | 50,154 | 22.72% | 22.72% |
| 1-15 mph | 27,454 | 12.44% | 35.16% |
| 16-30 mph | 28,775 | 13.04% | 48.20% |
| 31-45 mph | 38,918 | 17.63% | 65.83% |
| 46-54 mph | 9,148 | 4.14% | 69.97% |
| 55-70 mph | 58,270 | 26.40% | 96.37% |
| 71+ mph | 8,011 | 3.63% | 100.00% |
The distribution is bimodal, and that is the finding. The single largest band is 55-70 mph at 26.40%, and the second largest is stationary at 22.72%. The middle of the range, where a test drive spends its time, is comparatively thin.
A dealer test drive is a short loop at suburban speeds. Call that 16-45 mph and it covers 30.67% of reported failures. Widen it generously to 1-54 mph and it still reaches only 47.25%. The 16-45 mph definition is ours, not NHTSA's, which is why the full distribution is printed above and shipped in the dataset: a reader who thinks a test drive covers a different band can recompute the share directly.
Which parts fail at speed, and which fail standing still?
This is where the data becomes useful rather than merely interesting. Components differ enormously in where on the speed range they are reported.
| Component (N>=400) | N | At 55+ mph | Stationary | Median | Towed |
|---|---|---|---|---|---|
| Adaptive cruise control | 2,204 | 70.0% | 7.8% | 65 mph | 2.5% |
| Sun/moon roof assembly | 539 | 61.4% | 11.9% | 60 mph | 2.4% |
| Tires | 1,040 | 56.0% | 11.6% | 55 mph | 5.8% |
| Structure | 1,043 | 48.5% | 29.5% | 50 mph | 4.5% |
| Vehicle speed control | 3,940 | 45.1% | 9.8% | 45 mph | 9.6% |
| Steering | 19,122 | 36.7% | 9.9% | 40 mph | 5.1% |
| Engine | 39,088 | 36.0% | 16.5% | 40 mph | 7.2% |
| Automatic emergency braking | 3,133 | 34.4% | 4.1% | 40 mph | 4.1% |
| Power train | 24,956 | 29.8% | 15.2% | 35 mph | 5.5% |
| Service brakes | 18,718 | 24.6% | 11.3% | 30 mph | 6.4% |
| Electrical system | 26,874 | 21.8% | 35.3% | 25 mph | 2.9% |
| Fuel system, gasoline | 1,409 | 19.9% | 46.3% | 10 mph | 21.6% |
| Air bags | 9,839 | 17.5% | 33.2% | 25 mph | 35.8% |
| Seats | 1,989 | 13.8% | 48.5% | 5 mph | 3.2% |
| Seat belts | 1,331 | 11.9% | 53.3% | 0 mph | 7.1% |
| Back over prevention | 430 | 2.6% | 80.2% | 0 mph | 1.6% |
The pattern is mechanically sensible, which is part of why we trust the field. Adaptive cruise control does not engage below highway speed, so 70.0% of its complaints are reported at 55 mph or above with a median of 65. Back-over prevention is a reversing camera, used at walking pace, and 80.2% of its complaints are stationary. Seat belts and seats are inspected and adjusted at rest.
For a used-car buyer the two ends of that table are two different shopping tasks. The highway-weighted group - cruise control, sunroof, tires, structure, steering, engine - is what a brief drive cannot reach. The stationary group - cameras, belts, seats, electrical, fuel system - is what a careful walkaround can catch before you ever turn a key.
Are the highway failures the dangerous ones?
No, and this is the claim we most expect to be misread, so here is the measurement.
| Speed at failure | Towed | Police report | Medical attention |
|---|---|---|---|
| 0 mph (stationary) | 3.37% | 2.10% | 1.63% |
| 1-15 mph | 8.60% | 5.48% | 2.94% |
| 16-30 mph | 9.46% | 4.90% | 3.12% |
| 31-45 mph | 9.42% | 4.49% | 3.02% |
| 46-54 mph | 9.29% | 4.65% | 2.85% |
| 55-70 mph | 7.26% | 2.87% | 1.70% |
| 71+ mph | 7.25% | 2.65% | 1.69% |
Tow, police-report and medical-attention rates all peak in the 16-45 mph range and fall at highway speed. We are not going to over-explain a result we did not predict; plausibly, low-speed and urban events are likelier to end with a tow truck and a report filed, while a highway fault is often a warning light or a feature dropping out that the driver carries home. Whatever the mechanism, the honest statement is narrow: highway-speed faults are harder to detect before purchase, not more severe when they occur.
Does the finding survive the obvious objection?
The obvious objection is selection: VEH_SPEED is blank on 48.4% of complaints, so perhaps only people with dramatic highway stories bother to fill it in, and the 30.03% is inflated.
We ran the check most likely to reverse the result. For every component we took its band shares among complaints that do report a speed, then reweighted those by that component's share of the full 427,908-complaint universe, blanks included. If speed-reporters skewed toward moving failures, the reweighted figure would fall.
| As computed | Reweighted to full universe | |
|---|---|---|
| 55 mph or more | 30.03% | 29.09% |
| Stationary | 22.72% | 24.71% |
| 16-45 mph | 30.67% | 30.10% |
| Outside 16-45 mph | 52.75% | 53.80% |
It moves up by about a point, not down. The reweighting covers 99.9% of the universe.
The seven-year series is the second check, and it is the stronger one, because field coverage collapsed over the period while the headline did not:
| Year | Speed-reported N | Coverage | 55+ mph | Stationary | Outside 16-45 |
|---|---|---|---|---|---|
| 2020 | 40,308 | 69.0% | 25.6% | 28.3% | 53.9% |
| 2021 | 28,709 | 58.8% | 27.4% | 26.2% | 53.6% |
| 2022 | 28,941 | 54.6% | 32.6% | 20.6% | 53.2% |
| 2023 | 32,083 | 51.3% | 32.2% | 20.6% | 52.9% |
| 2024 | 30,648 | 45.6% | 31.3% | 20.6% | 51.9% |
| 2025 | 33,466 | 45.4% | 31.1% | 19.2% | 50.2% |
| 2026 | 26,575 | 41.4% | 31.5% | 22.1% | 53.6% |
Coverage fell from 69.0% to 41.4% and the outside-band share stayed in a 3.7-point range. Note the internal movement that the total hides: the highway share rose from 25.6% to 31.5% while the stationary share fell from 28.3% to around 20%. Those two trends are real in the data and we are not claiming a cause for either.
How do we know the field means what it says?
VEH_SPEED is a number in a file. Before computing anything we checked it against the complaint narratives, which are a different field filled in by the same person.
| Speed band | "highway/freeway/interstate" | "parked/parking lot/driveway" | "would not start" |
|---|---|---|---|
| Blank | 5.86% | 5.61% | 0.99% |
| 0 mph | 3.60% | 13.10% | 2.69% |
| 1-15 mph | 8.73% | 16.63% | 1.06% |
| 16-30 mph | 10.46% | 6.14% | 1.33% |
| 31-45 mph | 13.25% | 5.16% | 1.22% |
| 46-54 mph | 25.20% | 4.23% | 1.14% |
| 55-70 mph | 44.78% | 3.43% | 1.03% |
| 71+ mph | 55.86% | 3.27% | 1.64% |
Highway language rises monotonically across all seven reported bands, from 3.60% to 55.86%. Parking language runs the other way, and would-not-start language peaks where the car is not moving. A field that was noise could not produce that gradient.
The same table settles a specific trap. NHTSA rewrote this file in 2021 so that previously blank numeric fields read as zero, which would make a stationary bucket meaningless. It did not happen to this field in these period files: 206,956 complaints are still genuinely blank, and the blank row's profile (5.86% highway, 5.61% parked) is distinct from the 0 mph row's (3.60%, 13.10%). Blank and zero are different states here, and we treat them as such.
What a VIN check can and cannot tell you here
It cannot tell you at what speed a particular car's faults will appear. No vehicle history report carries a speed-at-failure field, and it could not: the speed in this data comes from other owners' complaint narratives about the same model, not from the car in front of you.
What the written record does carry, for that specific VIN, is accident and damage history including airbag deployment, odometer readings and rollback indicators, theft records, junk and salvage auction records, ownership and sales-listing history, specifications, safety ratings and valuation, plus whether recalls exist on the vehicle and how many. A Zilocar report shows the auction record and title history; it does not show the legal title brand, which belongs to NMVTIS and the state DMV, and it does not show recall remedy status, open NHTSA investigations or whether a software fix was ever applied - those live with NHTSA and the franchise dealer.
The practical reading of this analysis is that the drive and the paperwork check different things, and the paperwork is doing more of the work than buyers assume. Roughly half the failure conditions in this file are ones a test drive does not reach.
Method
Source. NHTSA Office of Defects Investigation consumer complaint flat files, COMPLAINTS_RECEIVED_2020-2024.txt and COMPLAINTS_RECEIVED_2025-2026.txt, from https://static.nhtsa.gov/odi/ffdd/cmpl/. Downloaded 2026-10-25. Server Last-Modified 2026-10-24 09:28:27 GMT and 2026-10-24 09:29:25 GMT; zip member timestamps 2026-10-24 09:28:14 and 09:29:16; ZipFile.testzip() returned None on both archives. 626,410 tab-delimited rows, 51 fields on every row with no short rows.
Field positions were pinned against the authoritative layout at https://static.nhtsa.gov/odi/ffdd/cmpl/CMPL.txt (last updated 2026-04-30) rather than from the older 41-field description: these files carry 51 fields, so VEH_SPEED is field 32, PROD_TYPE field 46 and COMPDESC field 12.
Universe. A complaint repeats one row per component, so rows were deduplicated on ODINO (195,374 duplicate rows dropped) and restricted to PROD_TYPE='V' (dropping 1,242 child-restraint, 1,021 tire and 859 equipment complaints and 6 with a blank type). That gives 427,908 distinct vehicle complaints.
Analysis set. Of those, 220,730 carry a VEH_SPEED value that parses as an integer in 0-200. 206,956 are blank and are not counted as zero. 222 parse outside 0-200 (values such as 354, 450, 670, almost certainly odometer or kilometre entries) and are excluded; they are 0.05% of the universe and cannot move any figure here. The three counts plus the usable set reconcile exactly to 427,908.
Aggregation. Bands are on the reported integer mph, with 0 held as its own band throughout. Shares are of the 220,730 analysis set unless stated. Medians are the middle reported value. Component figures group on COMPDESC exactly as filed, with no consolidation of NHTSA's label variants, and components with N<400 are suppressed from the table but retained in the totals. VEHICLES_TOWED_YN, POLICE_RPT_YN and MEDICAL_ATTN were used only as within-band rates.
Reversal check. Each component's within-component band shares were reweighted by that component's share of the full 427,908 universe, blanks included, to simulate the distribution had every complainant reported a speed. The outside-band share moves from 52.75% to 53.80%, covering 99.9% of the universe. The seven-year series was computed separately and holds in a 50.2-53.9% range while coverage falls from 69.0% to 41.4%.
Field validation. · VEH_SPEED was cross-checked against the free-text CDESCR narrative for monotonicity (highway language 3.60% to 55.86% across bands) and against the blank row to confirm blank is not stored as zero.
Known limits. Five, and they bound every number above.
- Complaints are self-selected. This is a distribution of reported failures, not of failures, and not a fleet failure rate. It cannot support any claim about how often a model fails.
- Every figure is conditional on the 51.6% that report a speed. The reweighting and the stable seven-year series are evidence that the conditioning is not driving the result; they are not proof.
- The 16-45 mph "test drive band" is our definition. NHTSA publishes no such band. The full distribution is given so any other band can be computed.
- Reported speed is the complainant's recollection, often months after the incident, and clusters on round numbers (65, 70, 35, 55, 45).
- No model-level or make-level claim is made or supported here.
MODELTXTis free text with known label splits, and we did not consolidate it, so the dataset carries component and band cuts only.
