Leading vs Lagging Safety Indicators: What Each One Can and Cannot Tell You

The useful question is not which type is better. It is which decision each number is allowed to support.

Lagging indicators: what they are and why they stay

Lagging indicators measure what already happened. OSHA describes them as measures of the occurrence and frequency of events that occurred in the past, such as the number or rate of injuries, illnesses, and fatalities. CCOHS lists injury frequency and severity, lost time injuries, incidents including property damage and environmental releases, near misses, and workers compensation costs.

They are not going anywhere, for good reasons. They are the only indicators tied to actual harm, they are defined consistently enough to compare between organisations, and much of the set is a legal reporting obligation rather than a management choice.

CCOHS is direct about their limits. Lagging indicators cannot predict future effectiveness at preventing incidents. They report only retroactively. There is a time gap between the event and the point at which the outcome is measured. And used alone, they push an organisation into reacting rather than preventing.

Why a low rate at one site proves so little

The Bureau of Labor Statistics reported 2.5 million nonfatal workplace injuries and illnesses in private industry in 2024, a total recordable case rate of 2.3 per 100 full-time equivalent workers, down from 2.4 in 2023. Apply that base rate to a site of 200 people and a normal year produces a small handful of recordable cases.

That arithmetic has an uncomfortable consequence. A quarter with zero recordables at that site is the expected outcome even if nothing about the risk has changed, and one bad event moves the annual rate more than any programme did. Reporting culture moves it further: make reporting harder and the rate falls with no change in exposure at all. A number that responds that strongly to something other than risk cannot be the primary measure of whether controls are working.

Leading indicators: measuring the work, not the outcome

OSHA describes leading indicators as proactive and preventive measures that can shed light on the effectiveness of safety and health activities and reveal potential problems in a safety and health programme. CCOHS adds the practical framing: proactive, preventative, and predictive measures taken to identify and eliminate hazards in the workplace.

The Campbell Institute at the National Safety Council, which has run a research series on leading indicators since 2013, groups them into three families. Systems-based indicators measure whether the management system is doing what it says: audits completed, actions closed, procedures reviewed on schedule. Operations-based indicators measure the condition of the work itself: equipment maintained on time, permits issued correctly, exposure monitoring performed. Behaviour-based indicators measure what people do: hazards reported, participation in safety activity, observed use of controls.

The families matter because a set drawn entirely from one of them has a predictable blind spot. Systems-only sets measure paperwork. Behaviour-only sets drift towards counting people. Operations-only sets miss whether anything was learned.

A worked set for one warehouse

Six indicators for a distribution site with forklift traffic, manual handling, and a racking system. Each one comes with the thing it is actually measuring and the way it fails.

Read down the failure column and a pattern appears. Every one of these indicators fails in the same direction: the number improves while the risk does not. That is the property to design against.

What leading indicators cannot do

Two honest limits, both routinely ignored in vendor material and board packs.

Proactive is not the same as predictive. An indicator earns the word predictive when a link to outcomes has been demonstrated in your own operation over enough history to see it. Adopting a metric because it worked at another organisation, in another industry, with a different workforce, is borrowing a correlation you have not tested.

The Campbell Institute case study on Cummins shows what testing it looks like. Cummins picked a few starting indicators, one of them training hours, then took twelve months of data and calculated a correlation coefficient for each one against the incidence rate over the same period. Training hours came back at minus 0.86, a strong negative correlation. Rather than stop at a good number they asked what sat underneath it, and found the effect concentrated in specific training on risk assessment and job safety analysis rather than in training hours as such. Their director of corporate health and safety describes the programme as evergreen and says there is "no such thing as a perfect leading indicator", with indicators reviewed every six months and updated annually.

Two things in that are worth copying. The correlation was calculated on their own data, not borrowed. And a strong result triggered a further question instead of closing the topic, which is what turned a number into a decision about where training should go.

Any indicator that is easy to score becomes a target. Once a number appears in a monthly report with someone's name against it, the cheapest way to move it is to change how it is recorded. This is not cynicism about safety teams; it is what happens to every metric under management pressure. The defence is pairing each activity count with a quality measure that moves in the opposite direction when the count is gamed.

Root cause work has the same discipline problem. Our comparison of 5 Whys, fishbone, and fault tree analysis covers how a method produces a confident answer from a weak evidence base, which is the analytical version of the same failure.

Building a set that survives a year

Five rules, in the order they matter. Start from a hazard in the risk register rather than from a list of popular metrics. Check who can move the number, and prefer measures produced as a by-product of the work rather than ones someone reports about themselves. Pair every activity count with a quality check. Fix the review period before the first reading, so a single good quarter cannot be presented as proof. And at each annual review, retire any indicator that has not changed a decision.

Where controls are the subject of the measurement, the ranking of the control matters as much as its completion. Our guide to the hierarchy of controls covers why a completed briefing and an installed barrier should not carry the same weight in a metric.

Where SoterAI fits

Most indicator sets fail on collection rather than on selection. The numbers have to be assembled by hand from separate systems, so they arrive late and stop being assembled at all. When inspections, hazard reports, incidents, and corrective actions are captured as connected records, most of this set is a by-product of the work rather than a reporting exercise, which also removes the incentive to shade it.

That accumulated history is what Risk Intelligence is built from: not a generic model, but the record of what happened in your operation, which controls were verified, and where the same hazard keeps returning. Our short definition of Risk Intelligence covers the term, and the risk intelligence layer piece covers how the layer is assembled. The records pages show which fields each record type holds, which is what determines whether an indicator can be calculated at all.

See which fields each record type holds before you design the indicator set. Browse the records library.