What your incident data is not telling you

Open the annual report. Vehicle incidents, employee injuries, patient handling, near misses, and a bucket for everything that did not fit. Counts against each, probably a comparison to last year.

Every one of those categories answers the same question: what kind of incident was it. That is a question you could have answered on the day it happened, from the first line of the report, without any analysis at all. Sorting a year of incidents by kind produces a tidier version of what you already knew in January.

What it cannot answer is what was going on around any of them. A vehicle incident and a patient-handling injury that both happened in the eleventh hour of a shift, both with a crew member working alone because their partner was finishing a handover, are the same situation twice. Sorted by kind they land in different columns and never meet.

This is not a criticism of your form. Categorizing by type is genuinely useful for insurance, for regulatory reporting, and for knowing where volume sits. It is just not an analysis, and the gap between having categories and having an analysis is an easy one to stall in.

Think about the conditions that surround an incident as sitting in five domains: the system, the process, the equipment, the crew, and the scene.

Now go through the fields on your own form and place each one against those five. Where does the environment go? Where does who was involved and what they did? Then ask the harder questions. Is there a field for whether a procedure existed, whether it was followed, or whether it could be followed? For whether the equipment was the right equipment, available, and working? For staffing, scheduling, training currency, or supervision? Count how many of the five you can actually point at, and note which ones you cannot.

For any domain you could not point at, the consequence is absolute rather than partial. A dimension with no field does not get under-reported; it gets zero, every time, forever. Nothing your crews write can ever be filed under it, so the annual summary shows the whole problem living in the domains the form happens to ask about — which then looks like evidence that those are where the problem is.

It is worth sitting with how circular that is. The form determines the distribution, the distribution drives the interventions, and the interventions target whichever domains the form could see. Three years of that produces a safety program shaped by a stationery decision.

The second problem is subtler and it undercounts the thing you most need counted.

Suppose your form offers a factor dropdown, and it has both No spotter and Spotter n/a. Both mean a spotter was not there. Different crews pick different ones; the same crew picks differently depending on the month. Neither is wrong, and the report shows two modest counts where one substantial one exists.

Free text makes it worse and better at once. Worse because the variants multiply — no spotter, nobody spotting, spotter unavailable, backed without a spotter, partner still inside. Better because the information is at least present, which it would not be if the dropdown were the only route.

A condition described many times has had more opportunities to be described several ways than one described twice, so the conditions you see most of are the ones worth checking for variants. Treat a factor count taken straight from a dropdown as a starting point rather than a total, and read what the narrative says about the same thing before you believe a number is small.

There is a second gap alongside the missing domains, and a severity column being present does not close it. The question worth asking of your own export is not whether it has a severity field, but where the values in it came from. A severity somebody recorded at the time and a severity worked out afterwards — from the injury type, say, or from the recorded cause — are different kinds of fact, and a column of values carries no record of which one it holds. If nobody can tell you, that is itself the answer.

Inferred is not the same as recorded, and the distinction matters more than it sounds. An inferred value tells you what its category implies. A recorded value tells you what this event actually did. Only the second can settle whether the incidents carrying a condition were worse than the ones without it.

The practical consequence: with nothing recorded, findings can count but not weigh. You can say a condition appears in a hundred and forty-four incidents. You cannot say those hundred and forty-four were worse than the rest, because the record never said how bad any of them were. Findings stay at the level of coverage — real, pervasive, and silent about magnitude. A hazard profile can still be recovered from what the narratives describe, and it is worth having, but it is a reading of how hazardous each setup was rather than a record of what actually happened.

This is worth establishing before you commission any analysis, because it sets the ceiling on what that analysis can claim, and it is a question your own people can answer in an afternoon. And if severity was never recorded operationally, it may still exist somewhere else — a claims system, an OSHA log, a field nobody thought to include in the export. Where it does, connecting it is what lets a report weigh a finding by what actually happened rather than only count it.

The narrative box is where crews put everything the form did not ask for. That is not a failure of discipline; it is what a text field is for, and it is the reason the record is worth analyzing at all.

Which means the material worth reading may not be in the structured fields at all. It can sit in two or three sentences a tired person wrote at the end of a shift — and a supervision gap, an equipment complaint, a workaround, or the line telling you a practice has been normal for two years has to go somewhere — if the form offers no field for it, free text is the only place left.

It also means the record has to be read at a scale nobody reads by hand. Eighteen hundred narratives is not an afternoon, so the structured fields become the de facto summary of the year — and they can only summarize the domains they have fields for.

The gap between those two things is the honest answer to what your data is not telling you. It is not that the information is missing. It is that it is present, unread, and in a format that no annual report was ever built to summarize.

You can measure this today, without any data and without buying anything, because the question is about the form rather than the record.

Take your incident report and go through the five domains one at a time. For each, ask whether the form asks about it at all, and if so whether it asks as a required field or leaves it to whoever writes the narrative. Required and structured is full coverage. Present but optional or free-text-only is partial. Absent is a blind spot, and a blind spot means nothing can ever land there.

Then count the domains where the answer was absent. That number is the ceiling on what any analysis of your record can find, no matter how good the analysis is, and it is the single most useful thing you can learn about your safety program in an afternoon.

Doing this first is worth some emphasis. It is answerable immediately, it costs nothing, and it tells you whether your record is worth analyzing before anybody spends money analyzing it. If three domains have no field, the finding is the form — and fixing the form is a cheaper and more certain intervention than any analysis of what the broken one collected.