Mining Incidents
The quirks ledger, published

A field guide to MSHA data

MSHA's public datasets are the authoritative record of US mine safety enforcement, and several of their fields do not mean what their names suggest. We hit these traps while building this site, verified each one against MSHA's own published files, and wrote them down. This page is that list, plus how our pipeline works, so you can check our numbers and avoid the same mistakes in your own analysis. We are an independent research tool, not MSHA.

18
MSHA datasets, weekly
10
documented traps
0
AI-written figures

How we source the data

  1. 1
    Ingest verbatim, weekly. 18 of MSHA's bulk Open Government datasets: citations, assessments, contests, inspections, accidents back to 1983, employment and production, health sampling, ownership history, addresses of record. Alongside them: DOL wage-and-hour enforcement, SEC mine-safety disclosures, EPA methane reports, Census county context.
  2. 2
    Hash every file before parsing. We can prove which of MSHA's own publications any figure came from.
  3. 3
    Refuse ambiguous files. Parsers enforce the expected column layout; a partial or drifted load fails loudly rather than landing quietly wrong.
  4. 4
    Gate every load. Row-count floors, key uniqueness, cross-file join coverage, value ranges. A load that fails a gate is repaired before it serves a page.
  5. 5
    No AI writes our figures. Every number traces to a direct query over MSHA's own fields, and where two defensible calculations exist we say which one we used.

Found an error anyway? The corrections policy on the about page applies. For citation formats, see how to cite.

What the data does not say

Each entry states the natural misreading, the verified fact, and what we do about it here. The looks like and actually markers are the whole grammar of this page.

01

Vacated and withdrawn citations disappear from the violations file

22,634citations exist only in the contest file

looks like The violations file is the complete count of citations issued.

actually When a citation is vacated or the action withdrawn, MSHA deletes it from the Violations dataset. Its only public trace is the Contested Violations file — so violations-file counts are net of vacaturs, and reconciliations against SEC mine-safety disclosures can legitimately differ by exactly these rows.

On this site: contest sections count vacated and withdrawn citations from the contest record, and per-company attribution notes where that record runs slightly under.

02

A contest's final classification exists only for settlements and prevails

59,320decided contests carry no re-classification

looks like The "decided" columns say whether the S&S designation survived every contest.

actually MSHA re-classifies only settlements and prevails. A dismissed or defaulted contest leaves the citation standing exactly as issued; a vacated or withdrawn citation no longer exists. Neither carries decided columns — reading them alone silently drops sixty thousand decided contests.

On this site: S&S survival counts a designation as surviving when a settlement or decision kept it, or the contest died and the citation stands as issued.

03

MSHA's two contest files disagree about proposed penalties

8,464joined rows disagree on the amount

looks like Two federal files about the same contest state the same proposed penalty.

actually Most of the gap is the dockets file holding no value at all; the rest is consistent with the two files capturing the amount at different moments. Their decided amounts agree better than 99.9%.

On this site: penalty-at-contest figures come from the Contested Violations file, the more complete of the two.

04

The current year's proposed penalties always look low

~24%of the live year is unassessed

looks like Enforcement dollars collapsed this year.

actually MSHA proposes a penalty months after a citation issues, so roughly a quarter of the in-progress year has no amount yet. The dip is assessment lag, not policy.

On this site: an in-progress year's penalty total is never presented as comparable to a closed year.

05

"Proposed" is not "paid", and the delinquency date is not delinquency

100%of assessments carry a delinquency date

looks like A populated delinquency date means an unpaid penalty.

actually The date is populated on every assessment, including fully paid ones — MSHA's own documentation warns against reading it alone. The honest unpaid signal is the currently assessed amount minus the paid amount.

On this site: penalty cards show proposed, assessed, and paid separately; "outstanding" is computed from the difference, never from that date.

06

A zero exposure limit on a health sample is not compliance

39%of personal samples have no limit at all

looks like A concentration-over-limit ratio of zero is a clean result.

actually Those rows are screening measurements of substances with no applicable limit. Averaging the ratio over all samples fakes compliance for four rows in ten.

On this site: every over-limit figure uses only limit-tested samples, and the denominator is stated beside it.

07

Coal and metal/nonmetal exposure records live in different files

2halves of the industry, two file families

looks like The dust, quartz, and noise files cover mining.

actually Those cover coal almost exclusively. Metal and nonmetal exposure — including its silica record — lives in the Personal Health Samples and Area Samples files. Neither family is a subset of the other; a silica analysis from the quartz file alone covers half the industry.

On this site: mine pages show the coal axes and the metal/nonmetal axes side by side, from their respective files.

08

The coal/metal flag on a citation is the inspecting office, not the mine

looks like The violations file's coal/metal indicator classifies the mine.

actually It records which MSHA program area issued the citation, and districts cross-inspect — classifying mines by this flag misfiles thousands of citations.

On this site: a mine's commodity always comes from the mines dataset, never from the citation row.

09

Accident counts are not comparable across the year 2000

1983coverage starts; the file format breaks at 2000

looks like Accidents dropped sharply in 2000.

actually MSHA's archived pre-2000 files include reportable no-injury accidents — roof falls and ignitions that hurt no one — which the modern bulk file largely omits, so raw counts step down artificially at the boundary. Fatality counts are consistent across the whole period.

On this site: trend surfaces either use fatalities or stay within one era, and the coverage note on the about page states the boundary.

10

A "pattern of violations" is almost never on MSHA's list

9POV notices in fifteen years

looks like If an operator had a pattern, MSHA's Pattern of Violations list would say so.

actually The formal designation has been issued nine times in fifteen years; one is in effect today. A pattern argument is nearly always built from the citation record itself — the same violation at the same mine over time.

On this site: POV status appears when it exists, and entity pages carry the citation history the argument is actually made from.

Why publish this

Each of these took real time to figure out, and they are the kind of thing people keep getting wrong. If you use MSHA data for reporting, research, or casework, check your numbers against this list before you publish them. It is faster than finding out the hard way. If we have something wrong here, tell us. The corrections policy on the about page covers this page too.