How this was built

Where the data comes from

Two FDA sources, both public:

• The refusals themselves come from the FDA Data Dashboard API endpoint https://api-datadashboard.fda.gov/v1/import_refusals. It requires a free authorization key, obtained through FDA's Office of Inspections and Investigations Unified Logon. The key is not stored in this repository.
• The charge text comes from FDA's own crosswalk, ACT_SECTION_CHARGES.CSV, which maps each numeric charge ID to a short code, the statutory sections and the full charge statement. A copy as downloaded is in data/charge-codes-source.csv.

Every word of charge language on this site is quoted verbatim from that file. Nothing was paraphrased or shortened. The charges are what FDA writes on the Notice of FDA Action.

The query

The endpoint caps a response at 5,000 rows, so the fetcher walks the record month by month from January 2002 to the present, paging within each month until a short page comes back, and de-duplicating on ShipmentID. That produced 525,457 refusals dated January 1, 2002 through August 2026. The script stops only after eighteen consecutive empty months and exits non-zero if it fetches nothing at all, so a silent failure cannot be published as an empty database.

What a refusal is, and is not

A refusal is FDA's final decision that a detained shipment appears to violate the Food, Drug, and Cosmetic Act. The shipment must be exported or destroyed under Customs supervision within 90 days. It is an administrative finding about the appearance of a violation. It is not a court judgment, not a finding of fact about the firm and not proof that anyone was harmed. Firms may contest refusals, and a refusal can follow a paperwork failure as easily as a filthy one.

One shipment can carry several charges, so charge counts sum to more than the number of shipments. The unit of a charge count is charge-on-a-shipment, not shipment.

Counting rule for shipments

Each row returned by FDA is one line on a refused entry, identified by a ShipmentID like 1FX-5639990-0/20002/1. A single physical container can therefore appear as several rows when it holds several refused product lines. This site counts rows, which is how FDA's own dashboard counts them. Read "525,457 shipments refused" as "525,457 refused entry lines."

Charge codes without published text

279 distinct charge codes appear in the refusal record. 22 of them are not present in FDA's published crosswalk, and those are shown here as bare code numbers labelled "no published text." They are a small share of all charge hits. If FDA updates the crosswalk, re-running the build will fill them in.

Why only recent refusals are searchable line by line

The full 525,457-record history is used for every count, chart and ranking on the site. Shipping all of it as a browsable file would mean a download in the tens of megabytes, so the line-by-line browser covers refusals since September 2025 — 39,386 of them. Country, firm, charge and year figures are computed across the entire quarter-century, not the recent slice.

What this cannot tell you

• What share of imports get refused. There is no denominator here. A country with many refusals may simply ship enormous volume. Do not read the country ranking as a quality ranking.
• What happened next. Refusals can be appealed, and reconditioning can be approved. Outcomes are not in this data.
• Anything about products that were never sampled. FDA examines a fraction of entries, targeted by risk. Refusals reflect where FDA looked as much as where problems are.
• Firm identity over time. Firms are matched on the name and country as filed. Spelling variants of the same company will appear as separate firms.

Rebuilding it

Two scripts in this repository: fetch_fda.py pages the API month by month and writes one JSON object per line; build_fda.py joins the charge crosswalk, aggregates and writes the files in data/. Set FDA_USER and FDA_KEY in the environment. FDA updates the underlying report monthly.

Built with AI assistance. The fetching, aggregation and this page were produced with Claude. Figures are computed from FDA's API rather than transcribed, but check FDA's dashboard before publishing any number from here.

← Back to the database