Work · Analysis

Demonstration · Public data · Not client work

Who files the most severe-injury reports? It depends on how you count employers

Question

Which employers file the most severe-injury reports with OSHA — and can you trust the obvious way of counting them?

Data

OSHA's Severe Injury Reports: 105,996 reports of amputations, in-patient hospitalizations and losses of an eye, January 2015 to November 2025. They cover federal-OSHA jurisdictions only; states with their own OSHA-approved State Plans are not included. 2,492 reports carry the file's undocumented FederalState = 0 flag, mostly in State Plan states; they appear as a separate line wherever they matter. Reports dipped from 11,065 in 2019 to 8,908 in 2020, which reflects the pandemic, not a safety trend. Everything below counts reports filed.

Design

I counted reports by employer the obvious way, then repeated the count after each cleaning step, measuring what each step changed.

Result

Counted by the raw employer field, the file lists 71,454 employers; after cleaning it is 56,083 by name, or 65,923 if identical names in different states count separately. How many employers filed reports depends on the definition — identical names in different states are usually different small firms, but the same company for national employers.

71,454Raw employernames−10Duplicatesremoved−14,312Namesnormalized−1,034Fuzzymatching−15Nationalaliases56,083Employers(final)
Most of the distortion is in the names, not the duplicatesDistinct employers in OSHA's Severe Injury Reports, counted by name: the raw count, the change at each audit step, and the final count. The data table also gives the count by name and state after each step.Source: OSHA Severe Injury Reports, January 2015 – November 2025.
View data table
stepvaluekindemployers_by_name_afteremployers_by_name_and_state_after
Raw employer names71454start7145480196
Duplicates removed-10step7144480185
Names normalized-14312step5713267199
Fuzzy matching-1034step5609866132
National aliases-15step5608365923
Employers (final)56083end5608365923

Download the data (CSV)

Duplicates are a minor issue here. No row is an exact copy of another. A stated rule found same-incident pairs filed twice by the same employer; depending on how strict the rule is, 16 to 73 pairs qualify (42 under the rule I used), because the count depends on the matching rule. After checking each one by hand, 39 of 105,996 reports were merged. Two merges were judgment calls (XPO Logistics, 5/3/2019; VA John D. Dingell Medical Center, 5/2/2019: same machine or door and day, a different finger recorded), and three pairs were kept apart as two workers. A merged report keeps the lowest report number and the higher injury counts of the two.

The larger distortion is in employer names. Normalizing case, punctuation, legal suffixes, "dba" and store numbers removes 14,312 names; fuzzy matching removes 1,034 more; a short, hand-approved list of national aliases (USPS for U.S. Postal Service, store formats such as Walmart Supercenter) removes 15. Names that refer to the same operating employer merge (store formats and locations, acronyms); separate legal operating companies stay separate (FedEx Freight vs FedEx; subsidiaries vs parents).

Rank Raw employer field Reports Resolved employer Reports % of reports Name variants States
1 U.S. Postal Service 548 U.S. Postal Service 1,320 1.25 35 50
2 United States Postal Service 290 Walmart 866 0.82 296 29
3 UPS 245 UPS 637 0.60 36 28
4 U.S. POSTAL SERVICE 148 Lowe's Home Centers, LLC 241 0.23 74 25
5 United Parcel Service 134 Publix Super Markets, Inc. 214 0.20 33 3
6 Walmart 126 Tyson Foods, Inc. 201 0.19 17 13
7 USPS 112 Kroger 167 0.16 48 10
8 United Parcel Service, Inc. 109 Target Corporation 131 0.12 22 26
9 Walmart, Inc. 109 Walt Disney Parks and Resorts U.S., Inc. 131 0.12 24 1
10 US Postal Service 87 American Airlines 129 0.12 11 11
11 Walmart Supercenter 83 FedEx Freight 127 0.12 27 25
12 Publix Super Markets, Inc. 82 Menard, Inc. 115 0.11 14 8
13 Tyson Foods, Inc. 70 Home Depot 112 0.11 32 20
14 Lowe's Home Centers, LLC 69 International Paper Company 112 0.11 8 16
15 DHL Supply Chain 65 AT&T 110 0.10 16 15
— FederalState = 0 reports 2,492 FederalState = 0 reports 2,491 — — —

The raw top 15 is really six organizations: the Postal Service alone fills five places under five spellings.

How good is the matching? The matcher is precise but conservative. Of 150 merges it made, 145 were right: precision 96.7% (95% Wilson interval 92.4% to 98.6%). It finds an estimated 47.7% of true matches (95% stratified-bootstrap interval 38.1% to 61.5%). In the band just above the threshold, nearly all misses are typos inside common words that the distinctive-part check rejects ("Manufacturin", "Distibution", "Serive"); but the larger estimated share is pairs scoring just under the threshold, a mix of typos and variants such as "+" vs "and" — and that share rests on few labeled matches, so it is uncertain. The missed matches mean the employer count above is overstated by at most about 2.0% (1.2–3.0%). If every judgment-call label flipped, recall would be 50.2%. In the lowest score band none of 50 sampled pairs matched; if the true rate there were at its upper 95% bound, recall would be 28.2%. The rule was not tuned on these labels. Accuracy figures are illustrative: they were estimated against match labels drafted by an AI model. In client work, labels are checked by hand.

A count ranking is not a safety ranking. Report counts track how big an employer is, among other things. As an illustration only, here are the ten largest filers next to their own reported headcounts. Headcounts are national (or company-wide) and for one fiscal year; the reports cover federal-OSHA states only and span 2015–2025. So this is not a rate comparison and not a safety ranking. The association is loose: industry, and how much of a workforce sits in federal-OSHA states, matter too.

Employer Reports, 2015–2025 Headcount What the headcount covers Source
U.S. Postal Service 1,320 639,000 Total (substantially all U.S.) U.S. Postal Service, FY2024 Form 10-K (Sept 30, 2024)
Walmart 866 1,600,000 U.S. associates Walmart Inc., Form 10-K, fiscal 2025 (Jan 31, 2025)
UPS 637 406,000 U.S. employees (excl. temporary seasonal) United Parcel Service, Inc., Form 10-K, 2024 (FY2024)
Lowe's Home Centers, LLC 241 270,000 Total full- + part-time (mainly U.S.; no U.S. split) Lowe's Companies, Inc., Form 10-K, fiscal 2024 (Jan 31, 2025)
Publix Super Markets, Inc. 214 255,000 Total (stated as 'more than') Publix Super Markets, Inc., Form 10-K, 2024 (Dec 28, 2024)
Tyson Foods, Inc. 201 120,000 U.S. team members (all Tyson entities) Tyson Foods, Inc., Form 10-K, fiscal 2024 (Sept 28, 2024)
Kroger 167 409,000 Total (stated as 'over') The Kroger Co., Form 10-K, fiscal 2024 (Feb 1, 2025)
Target Corporation 131 440,000 Total incl. seasonal Target Corporation, Form 10-K, fiscal 2024 (Feb 1, 2025)
Walt Disney Parks and Resorts U.S., Inc. 131 171,000 Disney company-wide U.S. — NO parks figure is published; not comparable The Walt Disney Company, Form 10-K, fiscal 2024 (Sept 28, 2024)
American Airlines 129 133,300 Group total, full-time equivalents American Airlines Group Inc., Form 10-K, 2024 (Dec 31, 2024)

What this does not show

  • Which employers are safer. Counts reflect size, industry and federal-OSHA coverage; State Plan states are missing entirely, and state totals reflect coverage and workforce size, not relative safety.
  • A complete entity map. The matcher misses some true matches and keeps subsidiaries apart from parents by design; other definitions would give other counts.
  • Unreported injuries. The file holds what employers reported.

Reproduce it

The audit code, the decision log (including every hand-review decision) and the results files are at https://github.com/ts2427/ThresholdDataSciences/tree/main/demos/osha_audit; the run instructions are in the demos README. The label files themselves are not published.

If this were your data

Before a number like "reports per employer" drives a decision, I would find out what it is actually counting: the duplicates, the spellings, the definitions and the coverage — and say how much each one moves the answer.