At What Mileage Do Used Cars Break? The Median NHTSA Complaint Says 70,000 Miles - and Fixing the Model Year Reverses Which Parts Fail First

At what mileage do used cars actually break?

Key facts

• 70,000 miles is the median odometer reading at failure across 394,120 NHTSA vehicle complaints filed 2015-01-01 to 2026-10-07 that report one. The interquartile range is 33,000 to 110,000.
• Only 50.3% of the 783,760 vehicle complaints in that window report a mileage at all, and coverage falls from 63.2% of model-year 2010 vehicles to 10.5% of model-year 2024 ones.
• Median mileage at failure falls from 140,000 miles for model-year 2000 vehicles to 12,000 for model-year 2020, almost entirely because newer cars have driven less. Any component ranking that ignores this measures fleet age.
• Fixing the model year to the 2015-2018 cohort moves 20 of 23 component groups down and 3 up. The two rankings correlate at only 0.647.
• Back-over prevention is the reversal. Raw, it looks 6th earliest of 23 at 64,000 miles. Model-year-controlled, it is 22nd of 23 at 81,690 miles, nearly the latest-failing group of all.
• The model-level headline a four-year cohort produces (Jeep Wrangler, 11,000 miles) is an artefact: 1,200 of its 1,823 readings are model-year 2018 alone, median 6,200. On model-year 2015 and 2016 separately it sits at 37,753 and 40,000.
• Nothing here is a failure rate. The file has no registration denominator, so it can say when complaints arrive and never how likely a failure is.

Which car parts fail earliest, once you control for model year?

Tyres and wheels come earliest and engine cooling comes last. The table below gives both bases for the 23 component groups with enough readings to rank, sorted by the model-year-controlled figure. The right-hand columns are what the same data says if you do not control for model year, and the gap between the two rank columns is the point of the table.

Rank 7 and rank 20 are held by NHTSA's "unknown or other" and "structure" buckets and are omitted from the table as uninformative labels; both are in the frozen CSV. Suspension is the largest honest mover among named systems, from 17th raw to 5th controlled.

Why does the model year change the answer so much?

Because mileage at failure is mostly a measure of how old the car is, not of how good it is. Median mileage at failure falls almost perfectly monotonically with model year:

A model-year 2000 car reported in this window is at least 15 years old by the time it appears, so it can only be reported at high mileage. A model-year 2020 car has had at most six years to accumulate any. So a component fitted mainly to recent vehicles, like a reversing camera, can only ever be reported at low mileage, and will look like an early failure for that reason alone. That is exactly what happens to back-over prevention, and the mechanism is visible directly in its own model-year curve: a median of 108,000 miles on model-year 2013 vehicles, falling to 9,000 on model-year 2024 ones.

One honest wrinkle in that curve: it is not perfectly monotonic at the recent end. Model-year 2021 vehicles show a median of 30,000 miles against model-year 2020's 12,000, which runs the wrong way. We do not have a confirmed explanation and are not offering one.

Does a four-year cohort fix it?

Not fully, and this is the check that killed our own first headline. Inside the 2015-2018 cohort the four model years still have medians of 49,955, 45,000, 35,000 and 24,000 miles, a two-fold spread. So a nameplate whose complaints happen to concentrate in the newest year of the cohort will look early-failing on composition alone. Measured across the 101 models with at least 200 readings in the cohort, the Spearman correlation between a model's mean model year and its median mileage at failure is -0.463.

The Jeep Wrangler is the clean demonstration. On the cohort it is the earliest-failing model in the file, at a median 11,000 miles, which reads like a finding. It is not:

Two thirds of its cohort readings are model-year 2018, the launch year of the JL generation, whose complaints arrived early and at low mileage. On model-year 2015 and 2016 on their own the Wrangler is ordinary. So the model-level table in this analysis is computed on a single model year and then cross-validated against a second one, and the Wrangler does not survive into either end.

Which models have problems reported earliest and latest?

These are the nameplates that land in the same quartile in both model-year 2016 and model-year 2017, computed separately. Requiring the pattern to repeat in two independent model years is what makes it more than noise.

Neither list is a reliability ranking, and the reason is specific rather than a general disclaimer. Each of these medians is dominated by one problem. 64.6% of the Kia Soul's model-year 2016 complaints with a mileage attached are engine complaints (234 of 362). For the model-year 2016 Ford Escape, 30.4% are power train and another 21.8% are back-over prevention (190 and 136 of 625). So the number dates a known defect for that nameplate; it does not say the car is better or worse than one absent from both tables. A model with no concentrated defect wave tends to sit in the middle of the distribution and appear in neither list.

What a VIN check can and cannot tell you here

A Zilocar VIN check can show:

• the odometer readings reported on that specific vehicle over time, and rollback indicators where the sequence does not make sense,
• accident and damage records, including airbag deployment,
• junk and salvage auction records, theft records (NICB), ownership history, sales-listing history,
• specs, NHTSA and IIHS ratings, valuation, and whether a recall is present on the car.

It cannot show:

• what will fail next, at any mileage. Nothing in this analysis supports a per-vehicle prediction, and the absence of a registration denominator means the complaint file cannot even produce a per-model failure rate;
• whether a recall was actually completed, the remedy or flashed status, which lives with NHTSA and the dealer;
• open NHTSA investigations, which attach to a manufacturer and component rather than to a VIN;
• the legal title brand. Zilocar surfaces the salvage auction record and title history; the brand classification itself sits with NMVTIS and state DMVs.

The practical use of this analysis is narrower than a prediction and still worth having: the odometer reading on the car in front of you is one of the few hard numbers you get, and a VIN check tells you whether that reading is credible. What the national complaint record adds is which systems owners of that generation were reporting, and at roughly what point.

Method

Source files. NHTSA's Office of Defects Investigation consumer-complaint period files: COMPLAINTS_RECEIVED_2015-2019.zip, COMPLAINTS_RECEIVED_2020-2024.zip and COMPLAINTS_RECEIVED_2025-2026.zip. Downloaded 2026-10-10 at 07:24 UTC. The servers reported Last-Modified of 2026-09-25 09:46:34 GMT, 2026-10-09 09:34:18 GMT and 2026-10-09 09:35:17 GMT. Together they hold 1,111,007 component rows, and the newest filing date inside them is 2026-10-07, three days before retrieval.

Why not the combined file. The single FLAT_CMPL.zip download is not used here. It has silently regressed to a stale extract more than once while still serving a fresh Last-Modified header, once stopping at filings from 2018 and once at 1996, in both cases with an intact archive that parsed cleanly. The period files are the stable source, and this analysis asserts that the newest filing date is within 30 days of the run before computing anything. That assert passed at 3 days.

Universe and filter. Rows with PROD_TYPE = V only, which drops 13,206 equipment, tyre-as-product and child-seat rows. Deduplicated on ODINO, because one complaint can carry several component labels while the odometer reading is a property of the complaint, giving 783,760 distinct vehicle complaints filed 2015-01-01 to 2026-10-07.

The mileage field. · MILES, position 18 of the ODI complaint layout, is the odometer reading the complainant reported at the time of the failure. It is free text. A reading of 0 or blank is treated as not supplied rather than as a failure at zero miles, and readings above 500,000 are dropped as not credible for a passenger vehicle. 394,120 complaints (50.3%) carry a usable reading. All central figures are medians, not means.

Thresholds. Component groups are included when they have at least 2,000 all-year readings and at least 400 inside the 2015-2018 cohort, which leaves 23. The model-level table requires at least 80 readings within a single model year, which leaves 279 model-year-and-model combinations across 2015-2018.

Robustness checks, reported either way. Three of the four bite.

1. The naive component ranking is an artefact of fleet age. Median mileage at failure falls monotonically from 140,000 miles (model-year 2000) to 12,000 (model-year 2020) because newer cars have driven less. Holding the model year to 2015-2018 moves 20 of 23 component groups down and 3 up, and the Spearman correlation between the raw and controlled rankings is only 0.647. The largest move is back-over prevention, from 6th earliest at 64,000 miles to 22nd of 23 at 81,690. Its own model-year curve shows the mechanism: 108,000 miles on model-year 2013 vehicles down to 9,000 on model-year 2024.
2. A four-year cohort is still not tight enough, and this killed our first model-level headline. Within the cohort the four model years have medians of 49,955, 45,000, 35,000 and 24,000 miles, and across the 101 models with 200 or more cohort readings the Spearman correlation between mean model year and median mileage is -0.463. The cohort's apparent earliest-failing model, the Jeep Wrangler at 11,000 miles, draws 1,200 of its 1,823 readings from model-year 2018 alone (median 6,200); on model-year 2015 and 2016 separately it is at 37,753 and 40,000. The published model table is therefore computed on a single model year and cross-validated on a second, requiring a nameplate to land in the same quartile in both model-year 2016 and model-year 2017. The Wrangler appears in neither list.
3. The field's coverage is not missing at random. The share of complaints supplying an odometer reading falls from 63.2% (model-year 2010) to 10.5% (model-year 2024), and from 60.3% to 39.7% inside the 2015-2018 cohort itself. Differential non-response therefore pushes in the same direction as fleet age, and no cross-model-year mileage comparison in this file can be cleanly separated from it. This is a limit on the analysis, not a result we can correct for.
4. The model-level medians each date one defect wave rather than ranking durability. 64.6% of the Kia Soul's model-year 2016 readings are engine complaints (234 of 362); the model-year 2016 Ford Escape is 30.4% power train and 21.8% back-over prevention (190 and 136 of 625). The tables are published with that stated rather than as a reliability ranking.

Corroboration, and it is not a clean 4 of 4. The ingest was checked against NHTSA's own live complaints API (api.nhtsa.gov/complaints/complaintsByVehicle) on four model-year-and-model pairs. Two matched exactly: 2016 Toyota Tacoma 323 against our 323, and 2017 Honda CR-V 1,773 against our 1,773. Two differed by under 1% in opposite directions: the API returned 823 for the 2016 Kia Soul against our 816, and 1,143 for the 2016 Ford Escape against our 1,145. A refresh lag alone cannot produce a difference in both directions, and the discrepancy is not double counting on our side, because no ODINO in these files appears in more than one period file. It is most likely a difference in how the API normalises make and model names and deduplicates against the flat file. We report it rather than round it away; none of the figures published here is a total complaint count, so nothing load-bearing rests on it.

Precision and denominator limits. · 30.0% of supplied readings are exact multiples of 5,000 and 68.3% are exact multiples of 1,000, so the underlying distribution is coarse and small median differences between adjacent rows should not be read as meaningful. The file carries no registration denominator, so nothing here is or can be converted into a failure rate: every figure describes when complaints arrive among people who complained, not how likely a failure is.

Known limits.

• Complaints are self-selected. Only owners who chose to file appear, and filing is correlated with how serious and how surprising a problem felt, not with how common it is.
• A recall or news coverage can produce a complaint wave at whatever mileage the affected cars happened to be at, which is the mechanism behind check 4.
• The window itself truncates old model years: a model-year 2000 vehicle can only appear in this file if it was complained about in 2015 or later, by which time it was already 15 years old. That biases old model years upward and is part of why the fleet-age curve is as steep as it is.
• COMPDESC is a free-text component label and its generic buckets are large. "Unknown or other" alone carries 12,787 cohort readings, so a component group's figure is an upper bound on how specific it is.
• NHTSA republishes these files continuously, so a later download will not reproduce these counts exactly. The two frozen CSVs below are the reproducible artefact.

Frozen datasets.

• Per-component, both bases, with the rank on each: when-used-cars-break-mileage-at-failure-nhtsa-data-vin-check-2026-10-10.csv
• Per-model within a single model year, with each model's dominant component group and its share: when-used-cars-break-mileage-at-failure-nhtsa-data-vin-check-by-model-2026-10-10.csv

Sources

• NHTSA Office of Defects Investigation, consumer-complaint flat-file downloads, static.nhtsa.gov/odi/ffdd/cmpl/ - the three period files listed in the Method, retrieved 2026-10-10 07:24 UTC. Last verified 2026-10-10.
• NHTSA, ODI complaint record layout (field positions, including MILES at position 18 and PROD_TYPE at position 46). Last verified 2026-10-10.
• NHTSA datasets and APIs index, https://www.nhtsa.gov/nhtsa-datasets-and-apis. Last verified 2026-10-10.
• Zilocar vehicle-history report example, https://zilocar.com/vehicle-history-report-example - for what the report does and does not return. Last verified 2026-10-10.

Analysis and both frozen datasets by the Zilocar vehicle-data desk, 2026-10-10. Figures are computed from the public NHTSA files named above and can be reproduced from the frozen CSVs.