The occurrence reporting curve has been falling for three consecutive safety reviews, and the room splits in two: those who read into it the return on last year’s corrective actions, and those who dare not say that they no longer write anything. Neither camp can be right on the strength of the graph alone, because a volume of reports measures neither safety nor silence: it measures the meeting between events and a willingness to talk about them. Three tests separate the two hypotheses, and none of them calls for data the operator does not already hold in house.
A counter that adds up three independent quantities
The number of reports received in a month is the product of three factors that have nothing to do with one another: the number of opportunities — sectors operated, flight hours, maintenance tasks, cabin rotations —, the rate at which notable events occur within them, and the proportion of those events that someone takes the trouble to write up. The first is a matter of exposure, the second of safety, the third of culture. A fall in the product points to none of the three, and that is precisely what allows everyone to read into it whatever they wish.
The first discipline is therefore to take exposure out of the equation before any interpretation. A rate is built with a denominator consistent with its numerator: reports per thousand sectors when the population is flight crew, per thousand flight hours when a fleet is being followed, per hundred crew members when bases are compared, per hundred tasks when a maintenance shop is concerned. Many systems nevertheless record the numerator monthly and activity quarterly, which makes it quite simply impossible to calculate a rate per base: that is a design flaw in the information system, not an inevitability of safety.
Once normalisation is done, a good share of the observed “falls” disappears of its own accord: the season, a fleet withdrawal, a route lost, a shorter winter programme. What remains is the other share, the one in which the rate itself is falling — and only there does the question begin to arise.
The total number of reports is not a direct measure of safety. It’s a mix of how often things go wrong, how many chances there are for things to go wrong, and how willing people are to report them. To understand if fewer reports mean things are getting safer, you have to separate these three factors.
An airline operates 1,000 flights per month and receives 20 reports. If it reduces its flights to 800, the number of reports might drop to 16, even if safety hasn’t changed. Only by calculating reports per 1,000 flights (2.0 before, 2.0 after) can you see if the change is real or just due to fewer flights.
Three administrative discontinuities also produce false breaks that no normalisation can correct: a change of reporting tool, an overhaul of the taxonomy, the merger of two systems after a takeover of operations. Each one shifts the baseline without anything having moved on the line, and a form burdened with a few extra screens is enough to make throughput collapse. Before interpreting an inflection, one therefore rereads the version history of the reporting system, and dates changes of form in the same way as changes of procedure.
First test: set it against a measurement that depends on no one
Flight data analysis has one property that a report will never have: it records without asking the crew’s opinion. An approach that is unstabilised against the chosen criterion, a late go-around, a speed exceedance in a given configuration, a ground proximity warning, a landing outside the energy envelope adopted by the operator — all of that is extracted from the recordings by analysis algorithms, regardless of what anyone has decided to write.
The crudest comparison is already instructive. If the rate of exceedances detected by the flight data programme stays stable or rises while the reporting rate falls, the improvement hypothesis no longer holds: it assumes that events are becoming rarer, whereas a counter that no one has been able to influence states the opposite. The reverse is also true, and that is what makes the test honest: a concordant fall in both series is the strongest argument an operator can produce in support of a genuine improvement in its safety.
The finer test is better still, because it removes exposure from both sides at once. For the only classes of event that both systems see — those whose triggering is measurable and whose reporting is mandatory or expected — one calculates the share of events detected in the flight data that carry a matching report. This coverage rate depends neither on the number of sectors nor on the event rate, since numerator and denominator come from the same flights: it measures nothing but the propensity to report. It is one of the few just culture indicators calculated on facts rather than on a climate survey answered by those who already report.
Two precautions make the exercise workable. The first is procedural: matching means tracing a recorded event back to a flight, and therefore to a crew, something the flight data monitoring agreement concluded with staff representatives strictly governs. It is carried out by the designated trusted third party, who returns a count and never an identity — the product is an aggregate figure, and the exercise keeps its value only if everyone knows it can produce nothing else. The second concerns scope: flight data see only what the sensors see. A misunderstanding with air traffic control, pressure at departure, a loading error, a cabin event, a workshop non-conformity leave no trace there, and the coverage rate says nothing about those families.
Second test: the mix, not the total
Not all reports cost the same effort to write, nor the same risk. An event subject to mandatory reporting almost always leaves a trace elsewhere: a part removed, an entry in the technical log, a message from air traffic control, a recorded warning, a maintenance task scheduled in its wake. Not writing it up erases nothing and exposes the writer to having to explain one day why the trace exists without the report. At the other extreme, the voluntary report in which the reporter implicates himself — a setting missed, a chart misread, a loose readback, a decision he judges questionable in hindsight — leaves no trace at all if no one writes it. Keeping quiet costs nothing whatsoever.
The result is a very recognisable signature. A genuine improvement brings down all families roughly in proportion, because it acts on events and not on pens. A withdrawal of trust, by contrast, distorts the profile: the share of mandatory and technical reports rises mechanically, not because they are increasing, but because everything else is fading away around them. The share of reported human factors recedes, and the share of reports in which the reporter is the author of the deviation goes first, because it is the most expensive in trust.
This test is conducted on rates per family as much as on shares. An internal campaign on fatigue can inflate one family and give the illusion of a healthy profile while all the others are collapsing; conversely, a family eliminated by a successful corrective action distorts the profile without any trust having been lost. One therefore looks at both readings, and treats with suspicion any variation that coincides with an internal communication rather than with a change in operations.
If fewer reports are being written, but the only ones still being reported are the ones that are mandatory, it’s a sign that people are only reporting when they have to. This means they’re not reporting smaller issues or things they might have caused, which is a sign that trust in the system is breaking down.
An airline has 100 reports per month, with 40% being mandatory (e.g., technical issues) and 60% being voluntary (e.g., human factors). If the total drops to 80 reports, but 60% are now mandatory, it shows that voluntary reports are dropping faster, suggesting people are less willing to admit mistakes.
The trap. Setting a number of reports as a quantified target, in either direction. An upward target makes a single event burst into three reports and fills the system with noise, to the point of drowning the signals the analysis was looking for. A downward target — or a fall simply presented at a review as a result — teaches middle management that silence pays, and the lesson travels fast. Presenting the drop as a success to the oversight authority is very likely to produce the opposite of the intended effect: the fall is then read as a question put to the management system, and the ensuing exchange generally turns on the tests described here. A volume is a measurement; a measurement turned into a target ceases to be a measurement.
Third test: where the reports come from
A reporting system does not have one rate, it has as many as it has populations. Related to a hundred flight crew members, a hundred cabin crew members, a hundred certifying mechanics or a hundred operations officers, throughput can be compared between bases, between fleets, between functions and between contractual statuses. It is those comparisons, and not the network average, that carry usable information — the average is precisely the quantity through which a pocket of silence passes unseen.
Two readings emerge. Dispersion first: a network in which all the bases are falling together tells a different story from a network in which one drops away on its own. A collapsed pocket is almost always local and dated — a disciplinary case that ended badly and that everyone has heard about, a change of management, a meeting in which someone was named in front of the others. The date is found by overlaying the local curve on the calendar of staff movements, and what follows is handled by interview, not by pivot table.
Concentration next: in many systems, a handful of regular reporters produces a considerable share of the volume. The counter then becomes sensitive to individual departures and falls without any safety trend having moved. A broad, flat base of reporters is worth more than a high volume carried by a handful of people; the number of distinct reporters, related to headcount, deserves to be tracked in its own right alongside the number of reports.
That leaves the blind spots, almost always the same ones. The cabin, where a rate close to zero never means that nothing is happening there, but that the channel is not being used — often because it has never sent anything back, sometimes because the form is only accessible from a ground workstation. Staff on temporary assignment contracts, on wet-lease crewing or in subcontracted maintenance, for whom the protection attached to a report remains vague: they do not always know which system to file in, nor who will read it, nor what their contracting employer will learn from it. Yet their events are exactly the ones the operator sees through no other channel.
What delay, length and feedback tell you
Three shape indicators complete the picture, and they are available in any system without particular development. The median delay between the event and its filing lengthens when trust falls — people write after seeking advice, or after waiting to see whether the matter would resurface — and it also lengthens, quite simply, when the form has become heavier; the two causes are told apart by dating the latest version of the tool. The median length of the free-text narrative falls when reporters switch to the defensive minimum: the fact, nothing but the fact, no analysis. The share of reports in which the reporter names himself as the author of the deviation is the most sensitive of the three, because it demands the most trust.

A fourth indicator precedes all of them, and it is the only one the organisation controls entirely: the share of reports that have received personal and substantive feedback within the announced deadline. Not an automatic acknowledgement of receipt, but a reply that says what has been done, or why nothing has been. When that rate falls, reporting falls afterwards, with a lag corresponding to the time needed for the message “it is pointless” to make the rounds of a base. Tracking it amounts to holding a leading indicator of the volume to come, and therefore to being able to act before the drop rather than commenting on it at a review six months later.
The right reflex. Never present the counter on its own. Three figures fit on a safety review slide: the rate related to activity, the coverage rate matched against flight data, and the share of reports that implicate their author; dispersion by base and by contractual status goes in an annex, together with the number of distinct reporters. And if a target has to appear somewhere on the dashboard, let it bear on the feedback delay to the reporter, never on the number of reports received.
Taken end to end, these indicators are not read in isolation: none of them settles the matter alone, it is their concordance that decides between the two hypotheses. Two signatures then stand out unambiguously.
| Indicator tracked | Genuine improvement | Under-reporting |
|---|---|---|
| Reporting rate related to activity | falling | falling |
| Rate of exceedances detected by flight data | falling in step | stable or rising |
| Coverage rate of matched events | stable | falling |
| Share of mandatory reports | stable | rising in relative terms |
| Share of reports implicating their author | stable | markedly receding |
| Dispersion between bases, fleets and statuses | tight, shared fall | one or two collapsed pockets |
| Number of distinct reporters | stable | falling faster than the volume |
| Median filing delay | stable | lengthening |
| Share of reports that received feedback | maintained | already falling for several months |
| Location of the fall in the taxonomy | one or two families, dated | diffuse, with no attributable date |
When the fall is real, it can be proved
Nothing forbids a rate to fall for good reasons, and that is established positively rather than by default. A genuine improvement is almost always local, dated and attributable: a specific family of events becomes rarer from the month in which a modification, a new procedure, a redrawn chart or a change of configuration entered service, and the fall is found again in the flight data for the classes those data see. Three concordant signatures, a named mechanism, a before and an after. That is an argument one can defend in front of an auditor.
Conversely, a diffuse, undated fall that touches every family except the mandatory ones, that is found nowhere else and that no corrective action can claim, is not a result. It is a symptom, and it is treated as such: field interviews conducted by someone with no disciplinary power, a visible restart of the feedback loop, and an audit of the confidentiality chain — who has access to reports that are not disidentified, for how long, under what traceability, and what happens when a line manager asks for one.
What to check in your own system
Four checks are made in the cold and condition everything else. The denominator, first: does your system record activity at the same granularity and the same time step as the reports? Without that, no rate per base or per fleet can be calculated, and the safety review will remain condemned to comment on raw counts. The taxonomy, next: does it separate the mandatory report from the voluntary one, and does it carry a field indicating that the reporter is the author of the deviation? Adding that field costs infinitely less than a climate survey, and it can be read every month.
Matching with flight data, next: is it written into a procedure, with a named trusted third party and the agreement of staff representatives, or does it rest on an informal arrangement between two people who know each other? In the second case it will survive neither a departure nor an inspection, and no one will dare invoke it at a review. The feedback loop, finally: what deadline is announced to the reporter, who signs the reply, and what proportion of last quarter’s reports actually received one? That last question is put to the system, but the useful answer is found in the crew room.
The grey area does not close for all that, and it is better to own it in front of the authority than to dress it up. No threshold makes it possible to state that a given fall is abnormal: the tests described here produce a convergence of indications, not a proof, and a genuinely safe system served by a genuinely open culture also displays a modest volume. That is exactly why the coverage rate, and not the volume, is the indicator that survives both hypotheses. For anyone preparing for a nominated person role or sitting an interview for a safety post, the question will come up in one form or another — your reports are falling, what do you make of that? — and the expected answer is not an explanation. It is a protocol, with its three tests and the order in which they are run.
These reference points describe a method of analysis; in the event of any divergence, the applicable legislation, the flight data monitoring agreement and the operator’s approved documentation prevail over what is written here.
Questions fréquentes
At what point should a fall become a cause for concern?
No threshold makes sense taken in isolation, because volume depends first of all on exposure. The first thing to do is to normalise by activity, then to look at where the fall sits: a modest decrease concentrated on reports that implicate their author is far more worrying than a steeper decline spread proportionally across every family. It is the location of the fall in the taxonomy, and whether or not it agrees with the flight data, that triggers action, not its size.
Can reports and flight data be matched without breaching the flight data monitoring agreement?
Yes, provided the agreement explicitly allows for it and the matching is carried out by the designated trusted third party. That person works on flight identifiers, returns an aggregate count and discloses no identity, either to the safety function or to management. If the procedure is not written down, the calculation must not be improvised: it must first be added to the agreement, failing which the coverage rate will be contested at the precise moment it would be useful.
What should be done when the cabin or the station reports almost nothing?
A rate close to zero is never interpreted as an absence of events, but as an unused channel. Three causes are checked in order: the accessibility of the form from the means actually available to those populations, the time it takes to fill in, and above all the existence of visible feedback. Publishing a few cases that have been dealt with, naming what has changed, restarts a channel far more surely than any reminder note about the obligation to report.
Is a quantified reporting target really harmful?
Yes, and in both directions. An upward target produces splitting and noise, to the point of degrading the capacity to analyse; a downward target, or a fall presented as a merit at a review, teaches middle management that silence is rewarded. If a target must exist, let it bear on what the organisation genuinely controls: the delay and the rate of feedback to the reporter, whose deterioration always precedes that of the volume.
How should crews on temporary assignment contracts or subcontracted maintenance staff be handled?
The point to settle is that of the counter and of protection: these staff must know in which system to file, what protection attaches to it, and who will answer them. That is written into the assignment contract and into the manual, not into a welcome e-mail. Without that clarity, their reporting rate most often remains well below that of in-house staff, and their events escape the operator through every other channel at the same time.
Is a rise in reports a better sign than a fall?
It is just as ambiguous and is read with the same tests. A new, simpler reporting tool, an internal campaign, a merger of systems or a recent striking event all push the volume up without any safety trend having moved. The rise becomes good news only if the coverage rate matched against flight data also improves, and if the number of distinct reporters increases at the same time as the number of reports.

Leave a Reply
You must be logged in to post a comment.