PlantOps BI

TRIR with Small Numbers

A plant with 25 thousand hours a month sees a recordable every few months. Its TRIR is then a count of a handful of events, and a count that small moves a lot by chance. Here is how much, on the demo data, and how to show it.

TRIR by plant and year with exact 95% Poisson limits. 2026 covers January to August.

PlantYearRecordablesHoursTRIR95% range
Portland Plant 2024 4 401,937 1.99 0.54 to 5.10
Portland Plant 2025 2 403,234 0.99 0.12 to 3.58
Portland Plant 2026 (YTD) 1 273,557 0.73 0.02 to 4.07
Boise Plant 2024 6 298,540 4.02 1.48 to 8.75
Boise Plant 2025 1 300,252 0.67 0.02 to 3.71
Boise Plant 2026 (YTD) 4 200,226 4.00 1.09 to 10.23
Reno Plant 2024 0 204,499 0.00 0.00 to 3.61
Reno Plant 2025 3 203,823 2.94 0.61 to 8.60
Reno Plant 2026 (YTD) 2 137,339 2.91 0.35 to 10.52

Did Boise Get Safer?

Boise's TRIR went from 4.02 in 2024 to 0.67 in 2025, a drop of 83% that would headline a safety review. It rests on 6 recordables against 1. If the true rate had not changed at all, a split at least that lopsided would happen with probability 0.12 (an exact test of two Poisson rates, conditional on the 7 events). That is weak evidence at best, and consistent with no change at all. Don't settle it by eye from the ranges in the table, either: two 95% ranges can overlap while the difference between them is still significant, so compare the rates with a test like this one.

This cuts both ways. Reno's 0.00 in 2024 doesn't mean it was the safest plant that year; zero recordables in about 204 thousand hours is consistent with a TRIR up to 3.61.

Monthly TRIR Is Mostly Noise

By month, Boise reads 0 in 22 of 32 months and between 7.83 and 15.59 in the rest, because a single recordable in a month of about 25 thousand hours is a TRIR near 8. A monthly trend line of that is a picture of when incidents happened to fall, not of a rate.

Boise recordables by month. Each square is one recordable.
  1. Jan 24 1 recordable
  2. Feb 24 0 recordables
  3. Mar 24 0 recordables
  4. Apr 24 1 recordable
  5. May 24 0 recordables
  6. Jun 24 0 recordables
  7. Jul 24 0 recordables
  8. Aug 24 0 recordables
  9. Sep 24 1 recordable
  10. Oct 24 1 recordable
  11. Nov 24 1 recordable
  12. Dec 24 1 recordable
  13. Jan 25 0 recordables
  14. Feb 25 0 recordables
  15. Mar 25 0 recordables
  16. Apr 25 0 recordables
  17. May 25 0 recordables
  18. Jun 25 1 recordable
  19. Jul 25 0 recordables
  20. Aug 25 0 recordables
  21. Sep 25 0 recordables
  22. Oct 25 0 recordables
  23. Nov 25 0 recordables
  24. Dec 25 0 recordables
  25. Jan 26 2 recordables
  26. Feb 26 1 recordable
  27. Mar 26 0 recordables
  28. Apr 26 0 recordables
  29. May 26 0 recordables
  30. Jun 26 0 recordables
  31. Jul 26 1 recordable
  32. Aug 26 0 recordables

What to Show Instead

The DAX

DAX has CHISQ.INV, which gives the exact Poisson limits directly: the lower limit on a count k is half the chi-square quantile at 2.5% with 2k degrees of freedom, the upper is half the 97.5% quantile with 2k + 2.

Recordables =
CALCULATE (
    COUNTROWS ( fact_safety_incidents ),
    fact_safety_incidents[incident_type] = "Recordable"
)

Hours Worked = SUM ( fact_labor_monthly[hours_worked] )

TRIR = DIVIDE ( [Recordables] * 200000, [Hours Worked] )

-- Exact 95% Poisson limits on the count, scaled the same way as TRIR.
-- A period with no recordables still has an upper limit, so COALESCE to 0.
TRIR Lower 95% =
VAR k = COALESCE ( [Recordables], 0 )
VAR h = [Hours Worked]
RETURN
    IF (
        h > 0,
        IF ( k = 0, 0, DIVIDE ( CHISQ.INV ( 0.025, 2 * k ) / 2 * 200000, h ) )
    )

TRIR Upper 95% =
VAR k = COALESCE ( [Recordables], 0 )
VAR h = [Hours Worked]
RETURN
    IF ( h > 0, DIVIDE ( CHISQ.INV ( 0.975, 2 * k + 2 ) / 2 * 200000, h ) )

The site computes the same limits by searching the Poisson distribution directly, and its tests check them against published table values. The p-value uses the conditional binomial test: given the total count, the second year's share of events is binomial with the second year's share of hours.

TRIR KPI · XmR chart · About the demo dataset