A single FDA recall can generate dozens of rows across the agency's public data systems. The Fresh & Ready Foods LLC ready-to-eat sandwich recall from June 2025 produced 86 enforcement report records, or 2.8% of a two-year openFDA pull covering 2024 through 2025, yet those rows represent one recall event [4]. Analysts who count rows instead of events inflate recall statistics by roughly threefold for the broader dataset: 3,026 records mapped to 1,036 distinct event IDs during that window [4]. The duplication arises because the Enforcement Report stores product-level classifications, while press-release feeds, API endpoints, and archived downloads each slice the same underlying event differently. The core workflow is to choose a stable event identifier, consolidate product rows under it, then treat classification, dates, and updates as attributes that can vary depending on when and where the data was captured. No single FDA source is a complete event ledger, so analysts must decide which limitation they can tolerate and document it.
FDA recall data exists at two levels: product rows and recall events
FDA's Enforcement Report includes all recalls monitored by the agency, covering Class I, II, III, and not-yet-classified recalls as described in 21 CFR 7.50 [1]. The report offers product and event views because a recall event may include more than one recalled product [1]. The recall number labeled with a Z-prefix for devices identifies each affected product, not the underlying quality problem [3]. This is the structural reason double counting occurs: a microbiological contamination finding that triggers a recall of 86 sandwich varieties appears as 86 product records, each potentially carrying its own classification and code information.
The Investigations Operations Manual describes the Recall Enterprise System (RES), the electronic data system used by FDA recall personnel to submit, update, classify, and terminate recalls. The RES event record details all products included in a recall throughout the process and is the source of the reference number publicized in the Enforcement Report when a recall is classified [2]. A separate Recall Number, assigned by the responsible center for each recalled product it classifies, follows a letter-prefix format such as F-100-2024, identifying the 100th product classified by the Human Food Program in FY-2024 [2]. Analysts should therefore distinguish event-level identifiers from product-level recall numbers. Counting only recall numbers overstates the number of distinct recall incidents. Counting only RES events understates the number of products affected.
For devices, the openFDA API annotates original records and converts data into JSON, but the agency warns that this data should not be used to track the lifecycle of a recall [7]. The API does not update recall status after hazard classification [7]. This limitation matters because an analyst querying openFDA monthly will see the same recall reappear with unchanged status fields, creating apparent duplicates across snapshots. The Enforcement Report's History feature does track updates to classification, reason, code information, and product description for recalls posted or updated beginning July 24, 2018 [6]. Subscription feeds can notify users of new and updated recalls by keyword or commodity, but each notification is not a new recall event [6].
Aggregate records by event ID before computing any recall metric
A documented R workflow pulls enforcement data from api.fda.gov/food/enforcement.json and then uses count of distinct event_id values to measure how many recall events exist [4]. For the two-year window queried, 3,026 records reduced to 1,036 events, a ratio of 2.92 products per event with a median skew from ten largest events supplying 19.3% of all records [4]. The same grouping logic inside group_by() lets an analyst compute products per event by reason or class. This is the canonical deduplication step: always derive event counts from n_distinct(event_id), never from row counts.
Severity labels further complicate aggregation. Classification is assigned per product, so one recall can carry two or more classes [4]. In the example window, 1,009 events carried a single class, 26 events carried two classes, and one event carried three classes [4]. Summing per-class event counts yields 1,064, which exceeds the true 1,036 events [4]. An analyst who counts Class I events and Class II events separately and adds them will double count the 27 multi-class events. The correct procedure is to classify each event by its most severe product classification, or to report class-specific event counts with an explicit note that multi-class events appear in more than one total. The Fresh & Ready Foods recall contained 85 Class I records and one Class II record, illustrating that one event can straddle severity thresholds [4].
A concrete deduplication workflow for openFDA enforcement data
A suggestion, not a tested validation: begin by pulling all records for a defined report_date window from api.fda.gov/food/enforcement.json in paginated requests. Then group by event_id and compute n_distinct(event_id) for total events, and n_distinct(classification) per event to flag multi-class recalls. The R example shows this approach yields 3,026 records and 1,036 events for 2024 through 2025, with products_per_event derived as records divided by events within each group [4]. Store the event_id, recalling firm, and most severe classification in a distinct event table before computing any headline metric.
Compare that event table against the Recalls, Market Withdrawals & Safety Alerts page for the same period, matching on firm name and product description rather than row count. Expect misses: the page warns that not all recalls have press releases or are posted [5]. Record which press-release entries lack a matching enforcement event ID and exclude them from any merged statistic unless you can reconcile the identifier. The agency's subscription service can notify users of new and updated recalls by keyword, but each notification is not a new event [6]. Use it to flag changes for review, not as a counting unit.
Publication date and press-release feeds create snapshot duplication
The report_date field in openFDA is the day the FDA publishes the enforcement report, not the date the recall began [4]. All 86 records of the Fresh & Ready Foods recall share one report_date, so filtering by publication year assigns the entire event to one calendar bucket even if the recall initiation date occurred earlier [4]. Analysts who re-run a date-bounded query after the classification appears will see a not-yet-classified record replaced by a classified one, potentially counting two rows if they merge snapshots without removing the superseded version.
Press releases add another layer of apparent multiplication. The Recalls, Market Withdrawals & Safety Alerts page gathers information from press releases and other public notices about certain recalls of FDA-regulated products, and notes that not all recalls have press releases or appear on that page [5]. This feed remains available for three years before archiving [5]. A recall can have more than one press release, and the Enforcement Report lists all of them [1]. The 2026 Nov 18 entry for Fresh & Ready Foods expansion appears as a separate press-release row from the original June 2025 recall [5]. An analyst who treats each press-release publication as a distinct recall will overcount events that receive initial and updated announcements. The firm-issued recalls dataset files, downloadable by year, contain only press-release data and should not be used as a census of recall events [8].
Search fields can inflate or deflate counts depending on selection
The Advanced Search criteria include Product Description, Code Information, Product Type, Recall Class, Recalling Firm, Status, Recall Number, Reason for Recall, Classification Date, and Event ID [1]. Selecting terms from these fields filters product-level rows. A search for a recalling firm returns every product code under that firm, so the result count is a product count, not a recall count. A search by Event ID returns a single recall event regardless of how many products it contains [1]. Choosing Event ID is therefore the most direct way to retrieve a deduplicated incident.
The Not Yet Classified filter introduces a temporal distortion. FDA began posting certain recalls to the Enforcement Report before hazard assessment was complete, so entries initially display as pending classification and are later updated with Class I, II, or III [6]. A pull that includes not-yet-classified records may capture a recall before its final class, and a later pull may show the updated classification. An analyst comparing month-over-month totals without matching on event ID will not be comparing the same population. The History feature records these classification updates, but only for recalls posted or updated after July 24, 2018 [6]. For older records, analysts have no systematic way to reconstruct pre-classification states from public data.
Device, food, and press-release pipelines use different identifiers
Device recall data in openFDA runs from 2002 to the present and updates weekly [7]. A recall is defined as a firm's removal or correction of a marketed product that FDA considers in violation of the laws it administers, excluding market withdrawal and stock recovery [7]. Device recalls use the Z-number scheme described in the 2013 research analysis, where recall numbers increase sequentially [3]. The responsible-center letter codes vary by commodity: F for Foods, D for Drugs, Z for Medical Devices, V for Veterinary Medicine, B for Biologics, N for voluntary device safety alerts, T for Tobacco, and C for Cosmetics [2]. Merging food and device recall files without normalizing these identifier formats can create false duplicates when two systems assign overlapping numeric suffixes to different events.
The Enforcement Report offers an Event View distinct from Product View, allowing an analyst to retrieve additional information about a recall event via a View Details link [1]. The print view displays all recall information for each product, but the report does not provide a public unique key that spans the press-release feed and the Enforcement Report. An analyst working across both systems must rely on recalling firm, product description, and date matching, which is less reliable than event_id when records differ. The 2026 Nov 19 Grand Strand Sandwich Company listing appears in the press-release feed with product names for four Lunch Box sandwich varieties [5]. Without the corresponding enforcement report event ID, matching this list to product rows requires textual reconciliation that can misalign when a firm uses different product names across systems.
Sources & references
- Enforcement Report Information and Definitions | FDA · www.fda.gov
- FDA Investigations Operations Manual 2026 — Chapter 7 Recall Activities · www.fda.gov
- Software-Related Recalls: An Analysis of Records | Biomedical Instrumentation & Technology · array.aami.org
- Grouping FDA recall records into recall events in R – DataScience+ · datascienceplus.com
- Recalls, Market Withdrawals, & Safety Alerts | FDA · www.fda.gov
- Enforcement Reports | FDA · www.fda.gov
- Device Recall Overview · open.fda.gov
- Recalls Data Sets | FDA · www.fda.gov