Is the Safety Triangle Valid?

The Opportunity

A client for whom we had implemented a predictive safety analytics dashboard came to us with an interesting research proposition. They had seen a constant decline in reportable incidents in recent years, but no improvement in first aid cases.

This is inconsistent with a nearly century-old theory known as the Heinrich/Byrd Safety Pyramid, commonly referred to as the “Safety Triangle”. It proposes that the ratios between fatalities, accidents, injuries, and minor incidents follow a fixed pattern (the ratios represented in progressive stages of a pyramid).

Our client wanted to understand why this commonly employed theory was not holding true.

Our Approach

We used monthly key safety metrics data across multiple regions. Key elements of the reportable incident rate, such as injuries, fatalities, lost time incidents, and medical treatment cases, served as potential variables to compare with first aid cases. We also had potentially related variables such as safety observations, hours worked, training, etc.

We started with simple correlation analysis and moved on to regression modeling, in order to statistically test hypotheses. This method had capabilities to account for regional differences in the relationship between the studied variable and potential explanatory variables.

The Impact

Formal statistical hypothesis testing could find no relationship between first aid cases and reportable incidents. In other words, the ratio, moving up the pyramid, did not hold.

In fact, in one region, there was strong evidence that first aid cases were increasing while reportables were decreasing. This raised suspicions that perhaps medical treatment cases, which count as reportables, were incorrectly being classified as first aid cases, which are not.

While in some ways the research raised questions beyond the ones it answered, it did provide some conclusions to speculations by management. And the data-driven approach helped promote the notion that decisions are better supported by data analytics than by relying on hunches.