Methodology
Foodborne Watch aggregates published U.S. records about foodborne illness. It adds no epidemiological judgement of its own: every field is either copied from a source, or derived from a source field by a rule described on this page.
What this is not
- Not a count of foodborne illness. CDC estimates most foodborne illness is never reported. These are records of what agencies published, and that is a much smaller number.
- Not real time. Outbreak reporting lags investigation by months to years; recalls appear after FDA classification. Recent periods always look emptier than they were.
- Not a risk map for a place. A marker is a state centroid, not a location where anyone got sick. See “Geography” below.
- Not medical, legal or regulatory advice. For a recall affecting you, always read the source notice.
Source coverage and update timing
| Source | Type | Updates | Caveat |
|---|---|---|---|
| CDC National Outbreak Reporting System (NORS) | outbreak | Annual, with revisions | NORS is voluntarily reported by state and local health departments and is historical: outbreaks appear only after investigation closes, counts are final-as-reported, and many outbreaks are never reported at all. |
| CDC / state food safety advisories | advisory | As published | Advisories are point-in-time public communications. Case counts in an advisory are preliminary and are routinely revised upward or downward as investigation continues. |
| Cyclosporiasis weekly surveillance (fixture) | surveillance | Weekly during season (May–Aug), fixture until a real feed is configured | PLACEHOLDER DATA. No stable machine-readable weekly cyclosporiasis feed is configured yet; these rows come from a checked-in fixture and must not be cited. Weekly domestically acquired case counts are provisional by definition. |
| openFDA Food Enforcement Reports | recall | Weekly | Enforcement reports describe a firm's recall of a product; they are not outbreak reports. Distribution states are what the firm reported and do not indicate where illness occurred. Records are added after FDA classification, so recent recalls may be missing. |
Definitions: incident types
- Outbreak investigation
- A public health agency investigated two or more illnesses linked to a common food. Outbreak records are reported after the fact and are the slowest to appear.
- Surveillance signal
- A periodic count of laboratory-reported illnesses. A surveillance signal is not an outbreak determination — it describes reporting activity, not a linked event.
- Recall / enforcement
- A firm removed a product from distribution, usually after its own testing or an FDA finding. A recall is about a product, not about illnesses; most recalls have no associated illness.
- Advisory
- A point-in-time public communication — a do-not-eat notice, an exposure notification, or an investigation update. Numbers in advisories are preliminary and change.
Definitions: status
Status is mapped from the source's own wording, which is always shown beside it. A record is labelled confirmed only where the source itself says confirmed. Where a source states nothing, the status is “not stated” — never an assumption.
- Confirmed by source
- (
confirmed) - Probable (source)
- (
probable) - Suspected (source)
- (
suspected) - Under investigation
- (
under_investigation) - Closed / historical
- (
closed) - Status not stated by source
- (
unknown)
Geography
The finest geography stored anywhere in this system is the state. Points on the map are state centroids. Patient addresses, restaurant and retail locations, farm locations, and inferred exposure points are not collected, not stored and not displayed. A multistate record is drawn once per affected state; a nationwide record is not drawn at all and appears only in lists. Recall records carry the states a firm reported for distribution, which is where a product went, not where anyone was harmed.
How records are processed
- Each source has an adapter that fetches raw records and maps them to one normalized shape, validated by a schema. A record that fails validation is rejected and counted, never stored half-parsed.
- The untouched source payload is stored separately from the normalized record, so any field can be re-derived later without re-fetching.
- Records are deduplicated on (source, source record id). If a source revises a record, the previous version is preserved as a revision and the incident's revision number increments — visible at the bottom of every incident page.
- Pathogen is derived from the source's etiology or recall-reason text. Records outside the tracked pathogens are skipped, not filed under “other”.
Known gaps
- Weekly Cyclospora surveillance is served from a checked-in fixture, clearly labelled as placeholder data, because no stable machine-readable feed is configured. It must not be cited.
- USDA/FSIS meat and poultry recalls are not yet ingested; FDA-regulated foods only.
- Recall distribution states are parsed from a free-text field and are coarse; the firm's own state is the most reliable geography on a recall.
- Case counts in advisories are not scraped from prose — they show as “not stated”.