The most consequential number in workplace safety has no statistical relationship to the outcomes it is supposed to prevent. That is the finding of peer-reviewed research on TRIR, the Total Recordable Incident Rate, the number contractors win bids on and boards read as the state of safety.
Hallowell and colleagues examined the metric's statistical behaviour in The Statistical Invalidity of TRIR as a Measure of Safety Performance (Professional Safety, 2021). There is no statistical relationship between a company's TRIR and its risk of serious injury or fatality, and at the size of most real workforces, year-to-year movement is dominated by random variation. The decimal-point changes that drive celebration and alarm are mostly noise.
The national data says the same thing
US Bureau of Labor Statistics injury and illness data shows recordable injury rates falling by nearly half since 2006. Over the same period the fatal injury rate fell by roughly a fifth. If recordables were a faithful proxy for catastrophic risk, the two curves should have moved together. They did not.
The global picture is harsher. The International Labour Organization estimates around 2.93 million work-related deaths per year, the large majority from occupational disease, harm that appears in no incident log. And IOGP data on oil and gas fatalities shows 81% were contractors, the segment the reporting systems cover least.
Why a real number can measure nothing
Every recordable is a real event. The rate still fails, for three reasons.
First, serious events are too rare to count. A rate built on total recordables is dominated by frequent, minor events, which arise from different causes than the catastrophic ones.
Second, the metric invites its own corruption. A number tied to contracts and bonuses attracts classification pressure (e.g. the case managed into first aid, or the report discouraged). It can improve while nothing about the hazard changes.
Third, it reports after the harm. Even a statistically sound count of injuries arrives too late to prevent them.
A number with those three properties is not a safety instrument. It is an accounting convention.
What to steer by instead
The data to do better usually already sits in the safety function's own records (e.g. observations, permits to work and near-misses). Assembled, it supports different instruments.
Count the warning conditions, not just the injuries. The dangerous situations and failed safeguards that research links to serious events can be counted directly.
Predict where the next incident is likely, with odds per site and per activity tested against history, so the stated odds have actually matched what happened.
And wire the warnings to people. Predictions go to named owners with agreed trigger points that schedule an intervention.
None of this retires the recordable count; regulators require it. The shift is in what the function steers by.
Predicted incidents with tested odds are what the Prophesee Compliance Suite produces from the records a safety function already keeps. Built with and proven inside global enterprises. See your own sites' odds. Start here.