I Counted 59,455 Device Recalls. Factories Beat Design, Except for Hips.

My thesis was that FDA recalls name a design cause more often than a manufacturing cause. In the public record, that is false for the database as a whole. Manufacturing causes outnumber design causes 31.9% to 21.9%. This is a count of FDA root-cause labels, which is post-market evidence. It says nothing about patient harm.

Question

After my hip implant post, I promised to count how often recalls blame design, how often they blame manufacturing, and how often they name no cause. I expected design to win. I thought a bench-tested design was the weakest link more often than assembly. I also expected the split to differ by device type.

The question has three parts:

1. Across all FDA device recall records, which cause group is larger?
2. Does the answer change by device family?
3. Which assumption moves the answer most?

Data and where it came from

I used the openFDA device recall file. Its last update was 2026-10-08, and it holds 59,468 records [2]. The openFDA page says the data are unvalidated. It also says FDA may correct records later, and it says not to use the data to track a recall's life [8]. Coverage starts in 2002 [8].

Each record carries a field called root_cause_description. The firm and FDA fill it in. I counted its values [1]. The values sum to 59,455, which is 13 fewer than the record total. I do not know why. Perhaps those 13 have no cause value.

One independent count agrees on the biggest single labels. A 2026 blog analysis of 58,374 records reports Device Design at 13.0% and Process Control at 11.0% [9]. My own figures from the live file are 7,676 of 59,455 (12.9%) and 6,557 of 59,455 (11.0%) [1]. The record count is higher because the file grew.

I could not split by recall class. The recall file has no class field that I could count. The enforcement file has class but no root cause [7]. I counted its classes anyway: Class I 3,659, Class II 35,424, Class III 1,029 [7]. Class I is 9.1% of those 40,113 entries. The link between the two files needs a join by event number. I did not run it. This post therefore does not test the Class I and II part of my thesis. I say so plainly.

For device families, I searched the product description text. This is a keyword match, not an FDA device class. I read parentheses with spaces as "OR". I did not verify that reading. The term "ct" may also match other text. Treat family numbers as rough.

Method

I put each cause label into one of four groups. This is my judgment. Another reader would draw the lines differently.

• Design: Device Design, Component design/selection, Software design, Software Design Change.
• Manufacturing: Process control, Nonconforming Material/Component, Process design, Process change control, Employee error, Equipment maintenance, Material/Component Contamination, Component change control, Mixed-up of materials/components, Vendor change control, Environmental control, Release of material before test results, Finished device change control, Manufacturing material removal.
• No usable cause: Other, Under Investigation by firm, Unknown/Undetermined by firm, Pending.
• Left out of the comparison: all labeling, packaging, use error, storage, regulatory-status labels, and three software labels about manufacturing or deployment (483 records).

I summed counts by hand. I did not use the Lab. Anyone can repeat this from the cited counts.

For the gap between two shares I use this standard approximation. Let $p_d$ be the design share, $p_m$ the manufacturing share, and $n$ the number of records.

$$SE(p_d - p_m) \approx \sqrt{\frac{p_d + p_m - (p_d - p_m)^2}{n}}$$

This formula measures the noise in a gap between two shares of one set. The records are a census, not a sample. So the interval shows only how fragile a small family is. It does not describe a population.

Result

Whole database

Design labels sum to 13,033 records: 7,676 + 1,609 + 3,519 + 229 [1]. That is 21.9% of 59,455.

Manufacturing labels sum to 18,943 records [1]. That is 31.9%.

The gap is 10.0 points. The approximate 95% interval is plus or minus 0.6 points. Manufacturing leads.

A fair reader will say the "no usable cause" group hides the answer. That group is 16,889 records: 8,928 + 6,967 + 650 + 344 [1]. It is 28.4% of the file. Among the 42,566 records with a usable cause, design is 30.6% and manufacturing is 44.5%. Manufacturing still leads.

By device family

Family (text match)	Records	Design	Manufacturing	Gap, design minus manufacturing
Hip (femoral or acetabular)	630 [3]	238 (37.8%)	126 (20.0%)	+17.8 points, plus or minus 5.8
Imaging (x-ray, MRI, CT, ultrasound)	8,241 [6]	2,177 (26.4%)	2,469 (30.0%)	-3.5, plus or minus 1.6
Pump (infusion or pump)	2,576 [4]	666 (25.9%)	978 (38.0%)	-12.1, plus or minus 3.0
Catheter	4,152 [5]	703 (16.9%)	1,629 (39.2%)	-22.3, plus or minus 2.2
All records	59,455 [1]	13,033 (21.9%)	18,943 (31.9%)	-10.0, plus or minus 0.6

Hip recalls are the one family where design leads clearly. Of 630 records, 237 carry the plain label "Device Design" [3]. That fits the story in my earlier post. A hip joint cleared by comparison, with bench wear data and little patient follow-up, fails by design more often than by assembly. But this table does not show that the cleared-by-comparison route caused it. It shows only what firms and FDA wrote down.

Catheters show the opposite. Manufacturing causes are 2.3 times the design causes (1,629 against 703) [5]. A catheter is a simple object made in huge numbers. That is a plausible reason for process faults. It is my opinion, not a finding.

Imaging is nearly tied, and software explains why. Software design alone has 806 imaging records [6], which is 9.8% of the family. If software design were excluded from "design", imaging would show design at 16.6% and manufacturing at 30.0%.

What I did not measure

I did not measure harm. A recall record counts a product listing, not a patient. One bad lot of a cheap catheter can create many records. A single design flaw in an implant can create few records but serious harm. Counts of records are not counts of injuries. Class I is the group built for serious harm, and I could not split by it.

Sensitivity: which assumption moves the result most

I tested four assumptions.

1. The "Nonconforming Material/Component" label. This label has 5,965 records [1]. It means a part or material did not meet its specification. A supplier may have failed. Or the specification may have been poor, which is a design matter. If I move it out of manufacturing, manufacturing falls to 12,978 records (21.8%). Design stays at 13,033 (21.9%). The two groups tie. This single choice moves the result more than any other I tried.

2. The "no usable cause" group. For design to beat manufacturing, at least 5,910 of the 16,889 undetermined records must be design causes with no manufacturing share: 18,943 minus 13,033. That is 35.0% of them. If I assume all of them are design, design reaches 50.3%. That is an extreme bound, not a forecast. Only the nonconforming-material move comes close to a tie without assuming facts about unknown records.

3. Adding design-like labels. I added Labeling design (856) and Package design/selection (848) [1]. Design rises to 14,737 records (24.8%). Manufacturing still leads by 7.1 points.

4. Counting single labels only. Device Design alone is 12.9%. Process control alone is 11.0% [1][9]. In that narrow view, design leads by 1.9 points. The blog analysis I checked makes the same observation. It notes that design leads as a single label, but a process group reaches 17.0% when four labels are added [9]. The word "leading cause" depends on grouping.

So the honest summary has two parts. If you compare one label with one label, design wins narrowly. If you compare design as a group with manufacturing as a group, manufacturing wins. The result flips with the treatment of one supplier-related label.

What changed my mind, and what stays

I held the view that recalls mostly teach design lessons. I now hold it more weakly for devices in general. The data do not support it as a general rule. I still hold it for hip implants, where the design count is nearly twice the manufacturing count. That is one family with 630 records, and text matching might include non-implant items.

I also keep one weaker claim: the cause labels are soft. About 28% have no usable cause [1]. A label such as "Other" cannot teach a design lesson. I like a recall notice that names its cause. By that standard, the record is a poor teacher for more than a quarter of its entries.

Evidence strength

My current view is that the evidence for "manufacturing leads overall" is moderate. The count is large, but the grouping is mine and one label can tie the result. The evidence for "hips are design-led" is weak to moderate: one family, a text match and no class split. The evidence for my original thesis is weak.

The gap I would close first is the join from event number to recall class. Class I recalls are where serious harm sits. If design leads there, my instinct survives in the part of the record that matters most. If it does not, I will say so.