Off the Shelves

Every recall in the Enforcement Report lands on the same three-rung ladder. Class I means a reasonable probability of serious harm or death. Class II means temporary or reversible harm; Class III, little chance of any. The ladder is the report's whole editorial judgment, and it is where this page starts: food carries the largest serious-harm load by far - 44% of food recalls are Class I, against 10% for drugs and 9% for devices.

Stream	Class I	Class II	Class III	Total
Drug	1,732	14,341	1,700	17,773
Food	12,806	14,670	1,740	29,216
Device	3,576	34,828	1,025	39,430

The record explains itself, in free text: every recall carries a reason_for_recall, here themed into recurring buckets across the Class I subset. The pattern is banal and that is the point. The most-serious recalls are rarely exotic - they are a label that forgot the allergen, a line that grew Listeria, a device that was never sterile. Food fails at contamination and the label; devices fail at sterility and mechanical defects; drugs fail at cross-contamination and potency.

1. 01
   Microbial contamination (Listeria, Salmonella) Food 10,823
2. 02
   Undeclared allergen Food 3,719
3. 03
   Lack of sterility assurance Device 863
4. 04
   Component or mechanical defect Device 764
5. 05
   Foreign material / particulate Device 479
6. 06
   Software or firmware malfunction Device 151
7. 07
   Cross-contamination with another drug Drug 126
8. 08
   Super- or sub-potent formulation Drug 86

Fig. 2Class I recalls attributed to each themed reason, ranked. The chip names the stream the theme dominates; the printed count carries the value. Counts are real, themed from the openFDA reason_for_recall free text.

The most-serious tier is not moving as one. Across 2017-2025, Class I device recalls climbed steeply while drug and food counts stayed noisy - ending near or below where they began. The three panels share one vertical scale, so the heights compare directly across the gutter: food still posts the largest raw counts most years, but the device line is the one bending upward.

Fig. 3Class I recalls per report year, one panel per stream, shared vertical scale. The delta compares 2017 to 2025. Counts are distinct Class I recalls per report year from the openFDA enforcement files.

● Full The current calendar year is still in progress, so the series ends at the last complete year.

Rising volume could just mean more products and better surveillance. The sharper question is the mix: of what each stream recalls, how much sits on the top rung? Over the past decade that mix moved unevenly - it eased for Drug and Food, but climbed sharply for Device: from 2.0% to 12.8% of its own records, more than six times over. It is the same stream driving the raw Class I count higher (Fig. 3) that is also getting more severe.

Stream	2017	2026	Change
Drug	6.2%	4.3%	-1.9 pp
Food	35.9%	27.3%	-8.6 pp
Device	2.0%	12.8%	+10.8 pp

● Full Share of the stream's own records, so a stream can grow safer in absolute terms while its mix worsens - and the reverse.

The filing rate does not hold steady across the year. Class I volume crests in Mar (220 filings) and falls to its low the very next month, in Apr (50) - a swing of about 340 percent, trough to peak. There is no single tidy season behind it: a second, almost-as-tall crest sits in July, and the low months are scattered rather than clustered in any one part of the calendar.

Monthly table

Month	Class I / mo
Jan	74
Feb	77
Mar	220
Apr	50
May	91
Jun	136
Jul	209
Aug	87
Sep	56
Oct	117
Nov	101
Dec	109

● Full Averaging over full years removes the partial-year lag that would sag the tail months.

Every enforcement record names the recalling firm and where it sits. Mapped that way, a recall is a return address: the record piles up where things are made - California, the mid-Atlantic drug belt, the industrial Midwest - not where they are eaten, swallowed, or implanted. Foreign and blank-state records are excluded. Darker states carry more records; the ranked table below is the source of truth.

Hover a state for its figures

● Full Counts are raw record volume, not per-capita; big-population, high-manufacturing states dominate by construction.

Big manufacturing states file the most recalls - that is Fig. 6, and it is mostly a census of factories. Divide instead: of each state's records, how many sit on the top rung? The answer scatters. Some states run far above the national 21.5 percent line, others well below it. A state can be a heavy filer with a mild mix, or a light filer whose record skews serious. Small states ride on few records, so their shares swing hardest.

● Full Small states ride on few records, so their shares swing hardest - read the left edge of the plot loosely.

The recall system runs on confession. Across all three streams just 0.8% of recalls are FDA-mandated; the rest are firms pulling their own products before the agency compels them. That is the design working - and it is also the design's exposure, because the ledger only records the failures firms notice and admit to. The rose slice marks the rare forced action.

Fig. 8Recall initiation per stream: the neutral track is voluntary, the rose slice is FDA-mandated (drawn at a minimum visible width; the printed share is exact). Split is real, from the openFDA enforcement files.

● Full FDA gained mandatory food-recall authority only in 2011 (FSMA); most mandated actions post-date it.

Termination is the FDA's word for over: the agency has verified the product is off the market and the fix held. It is slow. Of the terminated records, roughly 55 percent stayed open longer than a year, and the single most common outcome is one to two years on the books. Devices close slowest - a firmware fix must be verified unit by unit; a contaminated lot of food just has to be gone.

Open span	Records	Share
under 6 mo	14,881	22%
6-12 mo	15,780	23%
1-2 yr	18,537	27%
2-3 yr	9,521	14%
3-5 yr	7,441	11%
5+ yr	1,767	3%
All terminated	67,927	100%

● Full Ongoing recalls are excluded by construction, which biases these spans short - the still-open tail hasn't finished growing.

Recalls cluster. A minority of firms account for a repeated share of events, usually for the same recurring failure - the compounding pharmacy that cannot hold sterility, the diagnostics maker that keeps correcting the same device. These are the real firms that appear most often in the openFDA enforcement files, counted by distinct recall event. Read the count with care: a long catalog of products earns more recalls than a short one, so volume here is not the same as negligence.

#	Firm	Stream	Worst class	Events
01	Siemens Medical Solutions USA, Inc	Device	Class I	313
02	Siemens Healthcare Diagnostics, Inc.	Device	Class I	238
03	GE Healthcare, LLC	Device	Class I	224
04	Baxter Healthcare Corporation	Drug	Class I	184
05	Boston Scientific Corporation	Device	Class I	151
06	Beckman Coulter Inc.	Device	Class I	150
07	Zimmer Biomet, Inc.	Device	Class I	144
08	Hospira Inc.	Drug	Class I	139

Table 1Firms ranked by distinct recall events in the openFDA enforcement files, with the worst hazard class each reached. Counts track product breadth as much as safety.