The same TRIR of 3.20 scores p76 against its industry and p93 against its true peers
A rate carries no verdict on its own. Whoever picks the comparison group picks the answer.
600,000 hours and 10 recordable cases give a TRIR of 3.20
One plastics resin plant, NAICS 3252, about 300 employees, CY2025. The rate is fixed for the rest of this page. Only the comparison group changes.
Against all filers, 3.20 is 8% below the 3.49 national rate
3.49 is the hours-weighted aggregate across 375,635 CY2025 filings. This is the comparison that usually reaches the board deck.
Against NAICS sector 32, 3.20 is 37% above the 2.34 sector rate
Sector 32 aggregates 20,388 establishments. The plant did not change between these two slides. The denominator did.
Against the 478 filers in NAICS 3252, 3.20 ranks p76
The industry median is 0.57 and 45% of filers reported no recordable case. Three quarters of the industry did better than this plant.
Against the 34 same-size peers, 3.20 ranks p93
Hold NAICS 3252 and add the 250 to 499 employee band. The aggregate there is 0.77. The defensible sentence is "near the bottom of our peer group", not "below the national average".
45% of this industry reported zero cases, so the average describes almost nobody
The aggregate is 1.37 and the median is 0.57. A mean quoted against this shape flatters everyone above it.
45% of NAICS 3252 establishments reported no case at all
That block of zeros is the left edge of the figure. The median rate for the industry is 0.57.
The 1.37 aggregate sits more than twice the 0.57 median
The gap is the long right tail: p90 is 5.56 and p99 is 14.21. A handful of high-rate sites carry the average away from the middle of the group.
3.20 lands between p75 at 3.01 and p90 at 5.56
That is inside the worse half of a distribution whose better half is mostly zeros. Low percentile is the good direction, so this site reads as a problem the moment the group is right.
Changing the peer group moves the same rate by more than 20 percentile points
Move the dropdowns yourself. If a menu can change the benchmark, the menu is part of the reported number.
Choose a peer group.
Keep 3.20 fixed and the rank moves up to 20 points
Pick a different industry code or size band in the picker. The rate on the left never changes. Only the percentile does.
A p93 computed on n = 34 is two plants from a different answer
The readout prints n for whatever group you chose. Several NAICS and size combinations hold fewer than 50 establishments. Any rank from a group that small belongs in a deck with its n beside it.
Fix the peer group before you compute the rank
Write down the industry code and size band the site genuinely belongs to, then compute once. Picking the group after seeing four candidate ranks turns a safety number into a negotiation.
Size band alone swings the verdict from p72 to p93
Industry held fixed at NAICS 3252. Only the employee band changes, and the rank moves 21 points.
Show the numbers
The same 3.20 ranks p72 to p93 across four size bands
Four bands in NAICS 3252 have enough filers to publish. The harshest verdict lands in the band this plant actually sits in.
63% of the 20 to 49 band reported zero, against 45% industry-wide
Fewer exposure hours produce more true zeros, which pins the lower half of that band to the floor. Ranking a 300-person plant against it says nothing about the plant.
Rank each establishment in its own band, then report the spread
A corporate roll-up mixes size bands and industries and lands on a rate that matches no peer group in the file. Site-level ranks with their n attached are the reportable unit.
Every percentile on this page is interpolated from 7 breakpoints, not a computed rank
The limits of the method, written down before anyone quotes a number from it.
Limits, stated up front
- These are self-reported employer filings. Under-reporting is not visible in this data and is not corrected for anywhere on this page.
- The dots in the density charts are reconstructed from published percentile breakpoints, not individual establishment records. They show the shape of a distribution. They are not 478 real filings drawn one by one.
- Percentile rank is interpolated between seven breakpoints with an explicit anchor at the zero mass. It is an approximation of a rank, not a computed order statistic on the full sample.
- Several NAICS and size combinations have small n. The example group used through this page has n = 34. A rank on 34 establishments is unstable and should never be quoted without the n.
- A percentile is a position, not a cause. Nothing here says why a rate is high, and nothing here supports a claim that one site is safer than another.
- The 3.20 example site is an illustrative input, not a real establishment. Every figure it is compared against is real.
Fix the peer group before you fix the rate
- For each site, write down its six-digit NAICS code and its employee size band. Do this before anyone computes a percentile.
- Re-rank every site against that fixed group. Expect some sites that looked fine against the national average to move into the worst quartile.
- Put n next to every percentile in the deck. Any group under about 50 establishments gets a caveat line, not a headline.
- Stop reporting a single enterprise-wide rate against a national average. Report the distribution of site ranks instead.
What is missing
This was the last of the six instruments. The hub lists what none of them has answered yet, and what would have to be built to answer it.