Do most cars that get towed in NHTSA's complaint file crash first?
Key facts
- 5.56% of US vehicle complaints report a tow — 39,331 of 707,854 filed 2016 to 2026.
- 50.4% of those tows involved no crash (19,829 complaints, 2.80% of all complaints). The other 19,502 followed a crash.
- Towed complaints are overwhelmingly more severe on every independent flag: crash 49.58% against 1.58% for complaints with no tow, police 40.61% against 0.72%, medical attention 25.06% against 0.57%, fire 7.41% against 1.45%.
- Ranking parts by raw tow rate is misleading, and the error is systematic. Crash-coupled components rank high only because a tow follows the crash: air bags 5th → 19th, seat belts 12th → 21st, forward collision avoidance 13th → 20th. Breakdown components rise: engine 7th → 4th, power train 9th → 5th, electrical system 15th → 8th, visibility 22nd → 13th.
- The component ordering itself is stable over time (tie-corrected Spearman 0.860 between complaints received 2016-2020 and 2021-2026). The make, model and state orderings are not (0.578, 0.328, 0.651) and are therefore not published here as rankings.
- Engine and engine cooling is the one component that strands drivers without any crash: 22.21% of its complaints report a non-crash tow, against 0.30% for air bags.
Which parts actually leave you on the hard shoulder?
The table below is the whole finding. The first rate counts every tow; the second counts only tows where the complaint reports no crash. The last column is why the two disagree.
| Component group | Complaints | Raw tow rate | Non-crash tow rate | Rank (raw → non-crash) | Share of its tows with a crash |
|---|---|---|---|---|---|
| Engine and engine cooling | 10,344 | 22.40% | 22.21% | 1 → 1 | 0.9% |
| Vehicle speed control | 10,629 | 13.32% | 2.00% | 2 → 9 | 85.0% |
| Fuel system, gasoline | 5,809 | 13.12% | 12.93% | 3 → 2 | 1.4% |
| Service brakes, hydraulic | 5,925 | 11.86% | 9.52% | 4 → 3 | 19.8% |
| Air bags | 82,449 | 11.64% | 0.30% | 5 → 19 | 97.4% |
| Suspension | 15,827 | 7.01% | 4.32% | 6 → 6 | 38.4% |
| Engine | 104,216 | 6.21% | 5.70% | 7 → 4 | 8.2% |
| Service brakes | 50,101 | 6.18% | 1.10% | 8 → 14 | 82.2% |
| Power train | 67,979 | 5.58% | 4.67% | 9 → 5 | 16.3% |
| Steering | 50,742 | 4.76% | 2.74% | 10 → 7 | 42.3% |
| Electrical system | 88,161 | 3.27% | 2.56% | 15 → 8 | 21.8% |
| Seat belts | 6,548 | 4.12% | 0.24% | 12 → 21 | 94.1% |
| Forward collision avoidance | 9,846 | 4.04% | 0.29% | 13 → 20 | 92.7% |
| Visibility | 4,559 | 1.58% | 1.38% | 22 → 13 | 12.5% |
Full 25-row table, including every group not shown here, is in the frozen CSV linked under Sources.
Why does removing crashes change the ranking so much?
Because a tow after a crash measures the crash, not the part. Air bags are the clearest case: they are supposed to deploy in a collision, so a complaint about an airbag is very often a complaint filed after a collision, and the car is towed because it was wrecked. 97.4% of airbag tows had a crash. Reading the raw rate as "airbag faults strand you" inverts the causation entirely.
Engine and engine cooling is the mirror image. Only 0.9% of its tows involved a crash — the car stopped running and had to be carried. That is a breakdown rate, and it is the number a used-car buyer actually wants.
The two orderings correlate at a tie-corrected Spearman of 0.745, so they are not unrelated — but the disagreements are concentrated exactly where it matters, among the safety-equipment groups a buyer is most likely to misread.
Does this hold up if a complaint is counted under every part it mentions?
It holds, and it gets stronger. One complaint can carry several component labels. The table above assigns each complaint to its first component, so no complaint is counted twice. Counting a complaint under every group it mentions instead widens the universe to 27 groups and moves the raw-versus-non-crash correlation from 0.745 to 0.489 — that is, the reversal becomes larger, not smaller. Air bags sit 4th → 21st on that basis. Both bases are published per component in the frozen CSV.
Can you rank brands, models or states on this?
No, and we tried all three. Split by complaints received in 2016-2020 against 2021-2026, the orderings do not hold:
| Dimension | Keys compared | Tie-corrected Spearman | Published? |
|---|---|---|---|
| Component group | 21 | 0.860 | Yes |
| State | 44 | 0.651 | No |
| Make | 27 | 0.578 | No |
| Model (MY2015-2018 cohort) | 253 | 0.328 | No |
The make ordering moves violently between the two periods — Volvo 2nd → 20th, GMC 22nd → 5th, Mazda 12th → 25th. The state spread is too narrow to rank even if it were stable: across the 44 states with at least 2,000 complaints the non-crash tow rate runs from 3.56% (Louisiana) to 1.90% (Nebraska), a spread of only 1.88x.
There is one brand signal large enough to state as a caveat rather than a ranking: Kia complaints report a non-crash tow at 5.87%, against Toyota 1.40%, Honda 1.46% and Tesla 0.78%. That is consistent with the documented Theta II engine failures, but it rests on the same make dimension that only reproduces at 0.578, so it is a lift with its own instability and not a league table.
Is the leading component really about the component?
Partly not, and this is the trap that has caught this desk before. The non-crash tows under engine and engine cooling are 37.3% Ford (857 of 2,297 — Escape 252, EcoSport 251, Focus 109, Fusion 81) against a Ford baseline of 18.57% of all complaints: a 2.01x concentration. So that group's rate dates one manufacturer's defect wave as much as it measures a component. It is reported here rather than dropped, because the rank ordering survives temporally (0.860) while the magnitude of that one group should be read with the concentration in mind.
Does an older car get towed more often?
Not in a straight line, which is itself useful. Tow rate by model year runs 9.05% (2001), 4.35% (2008), 6.26% (2015), 4.82% (2025) — a U-shape then a hump, not the steady climb an ageing-fleet effect would produce. This matters because our own earlier work on the odometer field found a strong fleet-age gradient there. A mileage reading accumulates with age; whether a given breakdown required a tow does not. The trap belongs to quantities that accumulate.
What a VIN check can and cannot tell you here
This analysis is model-level, not per-VIN, and it is worth being exact about the difference.
It can inform what to inspect. If engine and fuel-system complaints are the ones that actually strand drivers, those are the systems to have inspected on a specific used car, and the ones where a service record matters most.
A Zilocar VIN check returns what has been reported on one specific vehicle: accident and damage records including airbag deployment, odometer history and rollback indicators, salvage-auction and theft records, ownership and sales-listing history, and any recall present on that car.
It cannot tell you whether that car was ever towed. NHTSA's tow flag lives on a complaint, not on a vehicle record, and complaints are recorded against a make, model and year. Nor can a VIN check show a recall's remedy status, an open NHTSA investigation, or per-unit firmware — a franchised dealer and NHTSA hold those.
What this analysis cannot claim
Four things, stated plainly. It is not a failure rate: the file carries no registration denominator, so every figure here is a share of complaints, never a share of cars. It is not a ranking of brands, models or states, all three of which failed cross-validation above. "Towed" is complainant-reported and unverified. And a component count is an upper bound on any one part, because the generic buckets absorb a great deal — note that ENGINE (104,216 complaints, 6.21% tow) and ENGINE AND ENGINE COOLING (10,344, 22.40%) are two separate buckets that both mean engine, and their rates differ by 3.6x.
Method
Source. NHTSA ODI consumer-complaint period files COMPLAINTS_RECEIVED_2015-2019.zip, COMPLAINTS_RECEIVED_2020-2024.zip and COMPLAINTS_RECEIVED_2025-2026.zip from https://static.nhtsa.gov/odi/ffdd/cmpl/, public domain, downloaded 2026-10-13 07:21 UTC. The zip members carry timestamps of 2026-09-25 09:46, 2026-10-11 09:23 and 2026-10-11 09:24, and all three archives passed a CRC check. The combined FLAT_CMPL.zip was deliberately not used: it has silently served stale extracts under a fresh timestamp on three separate occasions.
Freshness assert, run before anything was computed. The newest receipt date in the three files is 2026-10-08, five days before retrieval.
Universe. 1,111,402 component rows → 1,098,186 with PROD_TYPE = V → 784,010 distinct vehicle complaints after deduplicating on ODINO, because one complaint can carry several component labels while the tow and crash flags are properties of the complaint. Restricting to receipt years 2016-2026 leaves N = 707,854.
Why 2015 is dropped. · VEHICLES_TOWED_YN reads Y on only 1.07% of 2015 complaints against 4.81% to 6.79% in every year from 2016 to 2026 — a five-fold discontinuity at the very start of the file, which is a field not yet populated rather than a safer year.
The field-integrity check that had to come first. · VEHICLES_TOWED_YN (position 49) has zero blanks in any year, which is the signature of NHTSA's 2021 retroactive rewrite of previously blank Y/N fields to N. That rewrite destroyed two neighbouring fields outright — in these same files ORIG_OWNER_YN returns just 9, 6 and 16 · Y records in 2024, 2025 and 2026 against roughly 70,000 complaints a year, and PURCH_DT coverage falls from 2.84% in 2015 to 0.00% by 2024. The tow field did not die the same way: its Y share is stable at 4.81-6.79% across all eleven years with no collapse. The test that separates a surviving field from a dead one is whether the Y share holds over time, not whether the column is fully populated. The consequence for every rate here is that "not towed" and "did not answer" are the same stored value, so each figure is a floor, not a point estimate.
Definitions. Raw tow rate is the share of a group's complaints with VEHICLES_TOWED_YN = Y. Non-crash tow rate is the share with VEHICLES_TOWED_YN = Y and CRASH = N. Each complaint is assigned to the component group of its first · COMPDESC label, taking the text before the first colon, so no complaint is counted twice; the all-labels variant is reported alongside. Groups with fewer than 2,000 complaints are excluded, leaving 25.
Rank statistics are tie-corrected. Tied values take the average rank. An uncorrected rank function returns a figure that depends on input order, which silently flattered a weak correlation in this desk's work on 2026-10-11.
Robustness, five checks, three of which bite. (1) The component-assignment rule changes the headline correlation from 0.745 to 0.489, in the direction that strengthens the reversal; both are published. (2) The make, model and state orderings fail a temporal split (0.578, 0.328, 0.651) and are not published as rankings; the component ordering passes at 0.860. (3) The leading component's non-crash tows are 37.3% Ford against an 18.57% baseline, a 2.01x concentration, so its magnitude partly dates one defect wave. (4) Fleet age does not produce a monotonic gradient here, unlike the odometer field. (5) The severity flags corroborate the tow flag independently — towed complaints run 49.58% crash, 40.61% police and 25.06% medical against 1.58%, 0.72% and 0.57% for the rest — so the field is recording real events rather than noise.
Limits. No registration denominator, so nothing here is a failure rate. Complainant-reported and unverified. Every rate is a floor, per the field-integrity check. Component buckets are upper bounds.
