An approach ending with a go-around under the stabilization threshold, an aircraft taking off without a scratch, and an internal report of a few lines concluding with a « loss of situational awareness by the crew. » A few months later, the same sequence repeats, on the same airfield, with another crew. The cause-and-effect tree is designed to make this ending impossible: it forces us to go back to where the narrative usually stops, and it provides the reviewer with criteria to assert that a branch was cut too soon.
The no-damage event is the only complete material
An accident destroys part of its own evidence: dislocated airframe, unavailable actors, investigation conducted under official constraints and communication rules. A no-damage event preserves everything. The aircraft is intact and its recordings readable, the crew can be reached within the hour, the company’s configuration that day—crews, maintenance, dispatch, weather, schedule—can be reconstructed piece by piece. It is the only case where both the objective trace and the actors’ reasoning are available.
This wealth has a very short lifespan, and the curve varies depending on the source. Flight parameters, maintenance messages, and technical recordings do not degrade; they are erased all at once, either at the end of the recording loop or on the next flight, depending on the equipment. The actors’ reasoning, on the other hand, degrades continuously: as soon as the first review of a report takes place, memory reorganizes around what is now known about the outcome. What was once uncertainty becomes a sign that should have been noticed.
When something goes wrong but no one gets hurt, it’s the best time to investigate because all the evidence is still there. But memories fade fast, so you need to talk to people right away. The technical data stays the same until it gets erased, but people’s stories change as they try to make sense of what happened.
A Boeing 737 on a short runway in hot weather (35°C, runway 1,800m) had a rejected takeoff at V1. The crew reported “loss of situational awareness.” Three months later, another 737 on the same runway had the same issue. The investigation found that both crews missed a critical maintenance note about brake wear, which wasn’t properly highlighted in the pre-flight briefing.
Hence a counterintuitive rule: the decision to conduct an in-depth analysis is based on the event’s potential, not on its observed severity. The useful criterion is the plausible severity of a very close combination—one more parameter, one fewer barrier. A company that triggers its trees based on damage will only analyze cases where the evidence has already disappeared.
Establishing the facts before drawing the first arrow
The quality of a tree is entirely determined before the first stroke. A node is a fact: something observable, dated, attributed to an actor or a system, and verifiable by a citable source. Three formulations immediately disqualify a node. Judgment—“insufficient coordination,” “poor-quality briefing”—is not a fact but a disguised conclusion, and it prevents going further back. The compound fact—“the crew continued an unstabilized approach in an incomplete configuration”—hides two or three facts that do not share the same antecedents. The negation—“the crew did not announce”—is only acceptable if a prescribed action existed, and in that case, what was done instead must be written.
The material comes from four sources that are not mixed: recorded traces, the documentation applicable at the time of the facts (not the version in force at the time of the analysis), external conditions and the day’s schedule, and finally, interviews. These are conducted separately and never report back to the interviewee what another person said: otherwise, you collect a common narrative, i.e., a single testimony.
The deliverable of this phase is not the tree, but the timeline: the facts aligned on a time axis, each with its source. It reveals things that the tree masks—a delay abnormally long between two actions, a decision made before the information supposedly justifying it was available, two facts that everyone thought were simultaneous but were actually separated by several minutes.
The grammar of the tree: necessity, sufficiency, antecedent
The construction starts from the last fact—the one that closes the sequence—and works backward. At each node, a single question, word for word: what was needed for this fact to occur, and was that alone sufficient? Depending on the answer, three structures—and only three—are obtained.
The sequence: a necessary and sufficient antecedent produces the fact. The conjunction: several antecedents, each necessary, none sufficient alone—dominant structure of no-damage events, and the reason why they remain rare despite frequent deviations. The disjunction: a single antecedent produces several facts, which almost always signals a high-level common cause.
To build a good cause tree, you need to ask two questions at every step: “Was this the only thing that could have caused it?” and “Was this alone enough to cause it?” If the answer is yes to both, it’s a direct cause. If multiple things were needed, it’s a combination. If one thing leads to multiple outcomes, it’s a common cause.
The word “necessary” has a strict and counterfactual meaning here: without this fact, and all other things being equal, the next fact does not occur. The word “sufficient” has an equally strict meaning: this fact alone, in the context of the day, produces the next one. Crossing the two gives a sorting grid that resolves most session disputes.
The tests that eliminate comfortable branches
A comfortable branch is one that ends on a node that no one will contest and that no one can correct. “Non-application of the procedure,” “lack of vigilance,” “operator error”: these three formulations close more branches in companies than all the others combined. Four tests are enough to reopen them.
| Test | Question asked at the node | What it invalidates |
|---|---|---|
| Necessity | If this fact is removed, does the next fact still occur? | Circumstances promoted to causes because they were present that day. |
| Suffisance | Does this fact alone produce the next, or has a co-factor remained implicit? | Linear sequences that hide a conjunction, and therefore a missing barrier. |
| Substitution | Would a qualified professional, in the same context and with the same information, have acted differently? | Blaming the individual when the answer is “no”: the cause then lies in the context. |
| Stopping point | Does the terminal node point to something the operator can act on? | Branches closed on a state of mind, a character trait or fate. |
After a runway excursion, an investigation concluded with “pilot error.” The substitution test revealed that 80% of pilots in the same situation would have also missed the wet runway warning due to a poorly designed checklist. The real cause was the checklist design, not the pilot.
The substitution test is the least understood. It does not exonerate anyone; it shifts the question. If qualified professionals placed in the same context would have mostly done the same thing, the context is the cause, and the measure must focus on it. If the answer is clearly “no,” it remains to explain why that actor had different information, workload, or training that day—which is still a system question.
Trap: the hindsight bias contaminates the tree through vocabulary, not conclusions. As soon as a node contains “should have,” “did not see,” or “neglected,” the known outcome has slipped into the description of the facts, and the entire branch above becomes useless: you are no longer tracing the antecedents of a fact, you are building a case. Second trap, less visible: the tree built in plenary with an event actor in the room becomes a negotiation, and the branches that are inconvenient are politely cut, by consensus.
Good reflex: have the tree reviewed by someone who did not participate in its construction, giving them only the timeline and the sources—not the report, not the conclusions. Ask them to reconstruct the first two levels alone. If they differ from yours, it is not they who are wrong: your nodes contain judgments that only the original group can decode. Also impose the writing rule: one node, one sentence, one subject, one action verb, one date, one source.
From the last node to the corrective action
A correct tree rarely produces a single cause, and always several points of action. Each actionable terminal node—and only that one—leads to a measure. The quality of the measure is assessed on a known and rarely applied scale: eliminate the hazard, physically prevent the sequence, detect and alert, prescribe by procedure, and as a last resort, inform or train. An analysis where all actions fall into the last two steps has not gone high enough: it asks the actors to absorb a problem that the system sends them intact.
Three mentions make an action verifiable: a designated responsible party by function, a deadline, and an effectiveness criterion defined before implementation. The latter is the one that is skipped. “Reminder to crews” has no criterion, so there is no way to know if the event can recur; “removal of the double possible interpretation in the approach chart, verified by cross-checking” has one. The effectiveness review is part of the analysis: as long as it has not taken place, the file is not closed, it is pending. This is an area where practitioner feedback is more valuable than form templates.
What a reviewer will look for in your tree
Before transmitting, read your tree backward: start from each terminal node and go back to the event, saying the chain out loud. Any link that requires an “and then of course” is a missing link. Then count your terminal nodes: if there is only one, you have probably built a sequence where there was a conjunction.

Finally, check in your documentation three points that vary greatly from one operator to another: who decides that an event will be the subject of an in-depth analysis, who validates the corrective actions and within what deadline, who can close the file after an effectiveness review. Also look at how your manual links safety analysis and possible disciplinary action: it is the clause that determines whether your colleagues will talk to you next time.
The framework changes with the size of the operator. In a small structure, the person conducting the analysis often flew, programmed, or maintained the aircraft in question: they are both investigator and possible antecedent. The solution is not to abandon the tree but to make this conflict visible—name in the file the nodes where the analyst is a party, and entrust the review of those branches to someone external. A large structure has the opposite problem: the analyst does not know the airfield or local habits and takes for a fact what is only the official description of the work. A node written according to the procedure instead of what actually happened produces a coherent but false tree: it describes a flight that did not take place, and the measures it generates correct a behavior that no one had.
There remains a gray area that the method does not resolve. The tree goes back as long as it finds facts, and it regularly ends up designating a program trade-off, an equipment choice, or a sizing decision—decisions made outside the safety manager’s scope. The method requires writing these nodes; it does not say who will carry them. Knowing how to formulate this level without turning it into a trial is what is evaluated in a candidate for a flight safety position, far more than mastery of the formalism, and it is what is worked on in training programs dedicated to the safety management system.
This article describes a method, not an obligation: the applicable texts, the operator’s approved documentation, and the procedures of its safety management system prevail in all circumstances over what is written here.

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