Key facts
- The file splits 85,690 hotline complaints (23.12%) against 285,013 filed on NHTSA's website (76.88%), out of 370,703 distinct US vehicle complaints filed 2021-01-01 to 2026-10-22.
- 17.60% of hotline complaints record a towed vehicle, against 2.39% of web complaints: a 7.35x gap.
- Removing recall correspondence moves it up, to 20.74% against 2.45%, a 8.47x gap.
- The gap is about immobilisation, not danger. Crash rates are 4.23% against 3.67%, a ratio of only 1.15.
- 13 of the 24 component labels with 2,000+ complaints draw 90% or more of them from a single channel.
FUEL SYSTEM, GASOLINEis 98.73% hotline;FUEL/PROPULSION SYSTEMis 0.10%. - 45.19% of hotline complaints reference a recall or an unavailable part, against 6.90% of web complaints.
- Hotline share does not rank the models that strand owners: the Tesla Model 3 has one of the lowest hotline shares and an above-average 6.38% tow rate.
What are NHTSA's two intake channels?
NHTSA's Office of Defects Investigation publishes every consumer complaint as a flat file, and field 21 of that layout, CMPL_TYPE, records how the complaint arrived. Fifteen codes are documented. Two of them carry essentially the entire modern file: IVOQ, defined as "NHTSA WEB SITE", and EVOQ, defined as "HOTLINE VOQ", a report taken down by an operator on the agency's Vehicle Safety Hotline. Together they are 370,703 of the 370,886 complaints filed since 2021. The remaining 183 are consumer letters, congressional referrals and defect petitions.
That field is not exposed by NHTSA's public complaints API, so anyone reading complaints through the API or through a commercial product sees the complaints with the channel stripped off. It exists only in the flat file, and as far as we can establish nobody aggregates it.
How differently do the two channels record severity?
Starkly, and in one dimension only. Every complaint carries structured yes/no flags for whether the vehicle was towed, crashed or caught fire, whether police were called, whether medical attention was needed and how many people were injured or killed. Set side by side, six of the seven barely move and one moves by a factor of seven.
| Severity flag | Hotline (n=85,690) | Website (n=285,013) | Ratio |
|---|---|---|---|
| Vehicle towed | 17.60% | 2.39% | 7.35 |
| Medical attention needed | 2.32% | 1.46% | 1.58 |
| One or more injured | 3.28% | 2.15% | 1.53 |
| Fire | 2.14% | 1.41% | 1.52 |
| Police report filed | 3.60% | 2.39% | 1.51 |
| Crash | 4.23% | 3.67% | 1.15 |
| One or more deaths | 0.08% | 0.07% | 1.21 |
The mechanism this points at is mundane and, once stated, hard to unsee: a stranded owner reaches for a phone, and an owner whose car still drives gets round to a web form. The hotline is not where more dangerous faults go. It is where undriveable cars go. We have measured the non-crash tow flag on its own before, and this is the clearest structural reason that flag behaves the way it does.
Does the finding survive the check most likely to break it?
It does, and it moves in the unhelpful direction for the obvious objection. The objection is that the hotline is really NHTSA's recall help desk: owners ring up because a recall part is unavailable, those calls land in the complaint file, and they would not involve towed cars at all. The first half of that is correct and substantial.
| Basis | Hotline tow rate | Website tow rate | Ratio |
|---|---|---|---|
| As filed | 17.60% (n=85,690) | 2.39% (n=285,013) | 7.35 |
| Recall correspondence removed | 20.74% (n=46,967) | 2.45% (n=265,341) | 8.47 |
| Recall correspondence only | 13.79% (n=38,723) | 1.63% (n=19,672) | 8.48 |
45.19% of hotline complaints reference a recall, a campaign number, a recall notice or a part that was not available, against 6.90% of web complaints. So the hotline really is, in large part, the recall mailbag, which is a finding in its own right for anyone reading these records. But recall traffic has a lower tow rate than the rest, so removing it raises both channels' rates and widens the gap. The severity difference is not recall paperwork. This is the same contamination mechanism we found dominating the gasoline fuel-system complaint bucket, measured here on a different axis.
Three further checks, each reported whichever way it came out:
- By filing year. The hotline tow rate runs 15.62% to 19.88% across 2021-2026 and the web rate 2.00% to 2.80%. No year reverses it, and no year is close.
- By component. Of the 23 labels carrying 100 or more complaints in both channels, the hotline tow rate is higher in 22. The single exception is
AIR BAGS, at 20.26% by phone against 20.58% online, and that is the one bucket already known to be about half recall correspondence because of the Takata campaigns. - By vehicle age. On the recall-screened basis the hotline tow rate sits between 17.7% and 22.0% at every vehicle age from 0 to 20 years, and the web rate between 1.6% and 3.5%. The gap is flat in age even though the share of complaints arriving by phone rises with age, from 10.27% on cars in their first year to about 21% on twenty-year-old cars.
Why does this change how a complaint count should be read?
Because the two channels do not share a component vocabulary, so for several components a published count is a count of one channel's paperwork. This is the part with consequences beyond this article.
| COMPDESC label, as filed | Complaints | From the hotline |
|---|---|---|
SERVICE BRAKES, HYDRAULIC | 4,867 | 98.79% |
FUEL SYSTEM, GASOLINE | 4,267 | 98.73% |
ENGINE AND ENGINE COOLING | 8,301 | 98.05% |
BACK OVER PREVENTION | 2,662 | 97.15% |
STRUCTURE | 3,156 | 97.12% |
ENGINE | 62,635 | 16.88% |
SERVICE BRAKES | 25,634 | 6.06% |
STRUCTURE:BODY | 11,359 | 3.18% |
UNKNOWN OR OTHER | 33,060 | 1.09% |
FUEL/PROPULSION SYSTEM | 13,280 | 0.10% |
Read the pairs. NHTSA carries two fuel labels, two brake labels, two engine labels and two structure labels, and in each pair one belongs to the hotline and the other to the website. Both labels of each pair are in active use in every year from 2021 to 2026, so this is not a vocabulary that was renamed at some point: it is two vocabularies running in parallel. FUEL SYSTEM, GASOLINE sits between 96.6% and 99.5% hotline in each of the six years while FUEL/PROPULSION SYSTEM sits between 0.0% and 0.6%.
The consequence is direct. A ranking of models by SERVICE BRAKES, HYDRAULIC complaints is a ranking of hotline callers. A ranking by FUEL/PROPULSION SYSTEM is a ranking of web filers. The two are not comparable, they cannot be summed without saying so, and because the channels differ 7x on tow rate, they are not even measuring events of the same seriousness. 13 of the 24 largest labels are in this position.
Does a model's hotline share tell you anything about the model?
Less than it appears to, and we tested it rather than assuming. Controlling for vehicle age by ranking within a single model year and then cross-validating on an adjacent one, and with recall traffic removed, the ranking is stable and strongly brand-shaped: seven Ford nameplates land in the top quartile across two adjacent model years, and the Honda Civic, Honda Odyssey and Tesla Model 3 land in the bottom.
| Nameplate | Hotline share, MY2017 / MY2018 / MY2019 | Tow rate, all complaints |
|---|---|---|
| Ford Fusion | 39.9% / 36.1% / 36.4% | 7.38% |
| Ford Edge | 34.9% / 37.3% / 21.2% | 4.22% |
| Ford Escape | 32.8% / 31.9% / 32.3% | 7.13% |
| Ford F-150 | 27.4% / 26.8% / 24.0% | 3.47% |
| Ford Explorer | 23.1% / 24.8% / - | 4.96% |
| Honda Civic | 6.6% / 8.4% / 10.4% | 4.02% |
| Honda Odyssey | - / 7.1% / 6.7% | 2.06% |
| Tesla Model 3 | - / 5.7% / 7.0% | 6.38% |
| all models pooled | 5.20% |
The two columns do not line up, and that is the result. If hotline share were a severity signal, the Fords would carry high tow rates and the Teslas low ones. Instead the Tesla Model 3, with the lowest hotline share in the table, has a tow rate above the all-model average, and the Ford F-150 has one of the lowest tow rates on the list. What the hotline-share ranking measures is how a given model's owners get in touch with NHTSA, which plausibly reflects owner demographics, dealer instructions and the age and price of the car. It does not identify the cars that leave people at the roadside.
We are stating that as a negative finding on purpose. The tempting version of this article ranks brands by hotline share and calls it a breakdown league table. The data does not carry that claim.
What this means if you are buying a used car
NHTSA's complaint records are genuinely useful and free, and they are about the model, not about the car in front of you. Three practical rules follow from the numbers above.
Read the flags, not the count. A model with 400 complaints of which 3% record a tow is in better shape than one with 150 complaints of which a third do. The flags are on every complaint record.
Check whether a component figure is one channel's. If you are comparing a figure for SERVICE BRAKES against one for SERVICE BRAKES, HYDRAULIC, you are comparing two different populations of filers.
Never read a complaint count as a failure rate. The file has no denominator. We have made the same point about mileage at failure and about how long a defect takes to reach the record, and it binds here too.
What a VIN check can and cannot tell you here
It cannot tell you any of this. Complaints are filed against a model by other owners, and no vehicle history report carries complaint narratives, NHTSA's intake channel or a per-complaint severity flag. Zilocar's does not, and any report claiming to show which NHTSA complaints apply to your specific car is overstating what exists. Nor can any report tell you whether a recall on the car has actually been remedied, or whether NHTSA has an open investigation: remedy status comes from the manufacturer or a franchised dealer, and investigations are published by NHTSA.
What a VIN check does carry is the written history of that one car. Accident and damage records including airbag deployment, odometer readings and rollback indicators, junk and salvage auction records, theft records from NICB, ownership history, sales-listing history, specifications, safety ratings and valuation, plus whether recalls exist on the vehicle. Where that matters most for the finding above is the auction and damage record: a car that was towed, written off and resold leaves a trace in its own history even though the complaint that described the failure was filed against the model.
The division of labour is clean. Use NHTSA, free, for what a model tends to do. Use the VIN report for what happened to this car. A Zilocar report costs $39.50 for a single vehicle, or $29.50 for the first month and then $59.75 a month for 20 reports.
Method
Source. NHTSA Office of Defects Investigation consumer complaint flat files, COMPLAINTS_RECEIVED_2020-2024.zip (75,874,105 bytes) and COMPLAINTS_RECEIVED_2025-2026.zip (42,560,546 bytes), from https://static.nhtsa.gov/odi/ffdd/cmpl/, downloaded 2026-10-27. ZipFile.testzip() returned None on both. The combined FLAT_CMPL.zip download was not used: it has regressed silently to stale extracts more than once, so the period files are the default source.
Freshness assert, run before anything was computed. The newest filing date in the data is LDATE 2026-10-22, five days before the run. A stale-but-parseable file is the failure mode on this endpoint, so the maximum filing date is checked against the run date first and the analysis aborts if it is more than about 30 days old.
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 CMPL_TYPE is field 21, COMPDESC field 12, PROD_TYPE field 46 and VEHICLES_TOWED_YN field 49.
Universe. A complaint repeats one row per cited component, so rows were deduplicated on ODINO and restricted to PROD_TYPE='V', giving 429,351 distinct vehicle complaints across the two files. Filings before 2021-01-01 were dropped: the dealer-detail fields that first drew attention to CMPL_TYPE are only 1.5% populated in 2020 against 9% to 15% afterwards, which is a field not yet in service rather than a safer year.
Analysis set. Of the 370,886 complaints filed 2021-01-01 to 2026-10-22, 370,703 carry CMPL_TYPE of EVOQ (85,690) or IVOQ (285,013). The 183 on the five residual codes (LETR 177, CON 5, DP 1) are excluded as too few to report. The counts reconcile exactly.
Severity. Read from the structured flags as filed, with Y counted and everything else not counted. Note that NHTSA's 2021 flat-file rewrite converted previously blank Y/N fields to N, so every rate computed from these flags is a floor, not a point estimate. VEHICLES_TOWED_YN passes the two-step test for a field that survived that rewrite: its Y share is stable across the whole period rather than pinned at a corner.
Recall screen. A complaint counts as recall correspondence when its free-text CDESCR matches a loose screen for an NHTSA campaign number, a recall notice or notification, a recall repair, a part that was not available, or an unreasonable wait. This is the loose definition, the only one comparable between studies. It is approximate in both directions and a second screen would plausibly land a few points away, so the contamination shares here should be read as points in a range. The reversals are what the screen is being used for, and those are what survive.
Model-level control. Complaint counts carry no denominator, and a four-year model-year cohort is not tight enough at the nameplate level, so nameplate rankings were computed on a single model year with a minimum of 250 complaints and then cross-validated on an adjacent model year. Only nameplates landing in the same quartile in two adjacent years are reported. MODELTXT is free text with known label splits and was not consolidated, so a nameplate split across strings may be understated.
Known limits. Five, and they bound every number above.
- Complaints are self-selected. This is a distribution of reported problems, not of problems. There is no denominator anywhere in this file, so nothing here is a failure rate and no "worst model" claim is supported.
- The channel is correlated with the complainant, not only the event. Who reaches for a phone rather than a web form varies with age, region and vehicle value. We show the tow gap is flat in vehicle age and stable across six years and 22 of 23 components, which is evidence against a pure demographic explanation, not proof.
- The severity flags are floors, because NHTSA's 2021 rewrite made blank and genuine
Nthe same stored value. - The recall screen is approximate, as stated above, and the contamination percentages carry a few points of classifier spread.
- No causal claim is made about why a given model's owners phone. The nameplate table is published specifically to show that hotline share does not track tow rate.
