Two federal recall registers disagree about how fasteners fail

Why this is worth doing at all

Plenty of articles say a fastener can loosen or break. Very few say which one actually shows up in the record, and none we could find compares two registers using one method. That is the gap this fills.

Both sources are public and queryable, so the whole thing is reproducible. The CPSC recall API and the openFDA device-recall endpoint are linked at the bottom.

The counting trap, first

The obvious method is to count recalls whose text mentions screw, bolt, nut, rivet or fastener. On the CPSC set that gives 88 records, 6.0 %.

Reading them breaks the number. Most of those words sit in the product description:

Requiring the fastener word and a failure verb within 60 characters of each other leaves 24 records, 1.6 % — and all 24 were read individually. One dataset, one question, a 3.7× difference depending on how you count.

A smaller error is worth admitting too: without word boundaries, nut matches Nutcracker. A Christmas-goods recall was briefly counted as a fastener failure.

The two results

The top two swap. The firms doing the recalling are not bargain-bin operations either: Trek and SRAM bicycle parts, Lenovo power banks, Honda and Polaris off-road vehicles, BRP snow bikes, Petzl harnesses, IKEA toys, Rinnai boilers. Loosening leading the consumer register is not explained by somebody buying bad screws.

Fifteen ways of counting, one direction

The extraction is ours, so the obvious objection is that the pattern was chosen to fit. Five variants of the failure-verb pattern were crossed with three proximity windows (30, 60, 100 characters) and all fifteen combinations were re-run.

In all fifteen, the consumer register puts loosening first (44.8 % to 75.0 %) and the device register puts breakage first (38.8 % to 65.1 %). The direction never flips.

⚠️ The magnitudes move a great deal — the device-side hit count ranges from 218 to 598. What survives is the ordering, not any particular percentage.

What the regulator’s own field says

openFDA carries a coded root_cause_description, which removes our judgement from at least the attribution step. Against a control group of the non-fastener recalls in the same population (1,496):

Design is the largest single box and 2.6× the control rate. The component being out of specification — the supplier’s side of the line — is 14.0 %, about one in seven, so this is not a page arguing that fastener suppliers are never implicated.

⚠️ Process control runs lower, not higher (10.5 % against 15.4 %). That cuts against the shop-floor assumption that fastener trouble is mostly assembly and torque. It does not prove the assumption wrong; this register does not support it.

What cannot be concluded

What a specifier can take from it

Both registers agree on one thing: loosening and breakage are the top two, together around seven in ten. Which one leads depends on the industry you are in, so the useful question is not which ranks first in general but which is more expensive in your application.

Those two have different countermeasures, and a design that stops one does not usually stop the other. See why screws loosen, what the prevailing-torque figure is, and where the load actually sits.

One outside data point, offered as context rather than evidence: a clinical series on spinal implants reports screw loosening at 0.4 % and breakage at 1.3 %. That is a different population — events after implantation, not recall notices — so it is consistent with the device-register direction here, not a confirmation of it.

This is not one of the six steps. It shows up across them, or after assembly. Where the decisions that lead here were made is in specifying a screw, which sets out the order and why doing it out of order is rework.

Common questions

Does this say screws are more likely to break than to loosen?

Only in one of the two registers. In the medical device data breakage leads at 49.3 % against 22.0 % loosening; in the consumer product data loosening leads at 45.8 % against 33.3 %. The ordering reverses between registers, and fifteen variants of the counting method all reproduced that reversal.

How reliable is a sample of 24?

It is small and the page treats it that way. It is not a sample though — it is every recall in the CPSC set for 2023-01 to 2026-08 where the notice names a fastener as the cause, from 1,458 recalls in total. Widening the criterion to any mention gives 88, but reading those shows most are product descriptions rather than failure causes, so the wider number is worse, not better.

Can this be reproduced?

Yes. Both endpoints are public and linked below. The extraction requires a fastener word and a failure verb within 60 characters of each other in the description, title or hazard fields, followed by reading each hit. The robustness check varied both the verb pattern and the distance window.

Why is there no EU data?

It was attempted. The official EU Safety Gate page is a JavaScript application that returns only a page title to a plain fetch, and the machine-readable interfaces available are third-party scraping services rather than the primary source, so they were not used.

References

• US CPSC recall API: the consumer-product register used here
• openFDA device recall endpoint: the medical-device register, and the root-cause field
• Why screws loosen: two mechanisms under one word

Both registers were pulled in full for their stated windows: CPSC 2023-01-01 to 2026-09-01 (1,458 recalls) and openFDA (1,973 records, retrieved in two calls, complete population rather than a sample). Fisher exact tests were computed here. ⚠️ The fastener subset is selected by our own pattern; the robustness check varied that pattern five ways across three proximity windows and the direction held in all fifteen. ⚠️ The root-cause field is the filer’s attribution, not an independent determination.