Incident Investigation
5 Whys vs Fishbone vs Fault Tree: Choosing a Root Cause Method
Use 5 Whys when a failure runs down one chain and a short meeting can resolve it. Use a fishbone diagram when causes are many and spread across people, method, machine, material, measurement, and environment. Use fault tree analysis when the outcome is severe, several conditions had to combine, and the logic has to be shown to someone else.
By Matthew Hart
CEO, Soter
Written for safety professionals selecting an investigation method. Method choice and the adequacy of any corrective action remain decisions for the investigating team.
Three methods, one question
All three methods answer the same question: what has to change so this does not happen again. They differ in the shape of reasoning they impose. 5 Whys assumes a chain. A fishbone assumes a spread. A fault tree assumes a combination. Picking the wrong shape for the incident in front of you is the most common way an investigation ends with a corrective action nobody believes in.
OSHA's incident investigation guidance makes the underlying point without naming a method: investigations should reach the system failures behind the immediate cause. The method is the instrument you use to get there.
One incident, analysed three ways
Take a real-shaped event. In aisle 4 of a distribution centre, a forklift reversing towards the dispatch bay struck a racking upright. The beam was replaced the same evening. Nobody was hurt, and a picker had walked that aisle two minutes earlier.
5 Whys: following the chain
The team runs the chain in twenty minutes with the driver, the supervisor, and the shift lead.
- Why did the truck strike the upright? The driver reversed with no clear line of sight.
- Why was the line of sight blocked? The load on the forks sat above eye level.
- Why was the load that high? Two pallets were carried together to save a return trip.
- Why were two pallets carried together? The dispatch cut-off was ten minutes away and a trailer had arrived late.
- Why did the schedule pressure reach the forks? Nothing in the pick process flags a double-stacked load, and no one is asked to approve one.
That is a usable result. The corrective action writes itself: a check in the pick process, and a rule about travelling with the load trailing when it blocks the view. It took twenty minutes and no specialist.
Now notice what the chain had no room for. The aisle had no rack protection at floor level. The reverse alarm was inaudible over the baler running fifteen metres away. Neither of those caused the strike, and both of them determine whether the next strike breaks a beam or kills a picker. A single chain cannot hold parallel contributors, so it drops them.
Fishbone: opening the field
The same event on a fishbone forces the team through six categories before it converges on anything. The diagram ranks nothing; it exists to stop the room from settling on the first plausible answer.
| Category | Contributors surfaced |
|---|---|
| People | Driver certified and current; picker in the aisle with no task requiring them to be there |
| Method | No approval gate for double-stacked loads; no rule on which direction to travel when the view is blocked |
| Machine | Reverse alarm at a fixed volume; no camera or proximity system fitted |
| Material | Pallet height inconsistent between two suppliers, so a double stack is sometimes above eye level and sometimes not |
| Measurement | Near misses in the aisle reported verbally and not logged, so no trend was visible |
| Environment | Baler noise masking the alarm; no floor markings or barrier separating pedestrians from truck routes |
The fishbone found the two contributors the chain dropped, plus one nobody had raised: the same double stack is a hazard with one supplier's pallets and harmless with the other's. It also produced no ranking at all. Fourteen boxes on a whiteboard is a starting point for evidence gathering, and it settles nothing on its own.
Fault tree: showing the combination
A fault tree starts from a defined top event and works backwards through AND and OR gates. Here the top event that matters is not the beam strike; it is a pedestrian struck in aisle 4. The logic reads: a truck reverses into the pedestrian route AND the driver cannot see the route AND the pedestrian is present AND the pedestrian is not warned. Each of those branches then opens further. The driver cannot see the route when the load is above eye level OR the mirror is misaligned. The pedestrian is not warned when the alarm is masked by noise OR the pedestrian is wearing hearing protection OR the alarm has failed.
The value is the AND gates. They show which single control breaks the whole path. A physical barrier separating the pedestrian route removes one required condition and stops the top event regardless of everything else in the tree, which puts the barrier ahead of the driver briefing on any honest priority list. That reasoning survives the meeting in a form someone else can audit, which is the property the other two methods lack.
The cost is real. A tree of this size takes a facilitator, several hours, and people who can judge whether a branch is complete. For a scuffed rack upright it is more instrument than the question deserves. OSHA lists fault tree analysis at 29 CFR 1910.119(e)(2)(vi) among the accepted methodologies for a process hazard analysis, which is a fair guide to the class of problem it was built for.
Choosing before you start
| Dimension | 5 Whys | Fishbone | Fault tree |
|---|---|---|---|
| Reasoning shape | One causal chain | Categorised spread | Boolean combination |
| Best when | Low severity, one obvious path | Causes unclear or contested | Severe potential, coincident conditions |
| Typical effort | Twenty minutes, three people | An hour, a cross-functional group | Hours to days, a facilitator |
| Output | A chain ending in one system cause | A map of candidate causes | A logic diagram and the combinations that matter |
| Main weakness | Drops parallel contributors | Ranks nothing | Cost, and it needs a well-defined top event |
| Auditable by an outsider | Weakly, the chain reflects the room | Weakly, categories hide the reasoning | Strongly, the gates are explicit |
On a serious event, run them together. The fishbone opens the field, 5 Whys drills the branches that survive the evidence, and a fault tree records the combination logic for the paths that could have ended badly. They share the same evidence; only the structure of the reasoning changes.
Where an investigation is required by rule
For processes covered by the process safety management standard, 29 CFR 1910.119(m) sets specific obligations. An employer investigates each incident that resulted in, or could reasonably have resulted in, a catastrophic release of a highly hazardous chemical. The investigation is initiated as promptly as possible and no later than 48 hours after the incident. A team is established that includes at least one person knowledgeable in the process. A report is prepared covering the date of the incident, the date the investigation began, a description of the incident, the contributing factors, and any recommendations. The employer then establishes a system to address and resolve the findings, with resolutions and corrective actions documented.
Read that list against the three methods. It calls for contributing factors in the plural and for documented resolution, which is a poor fit for a five-line chain and a natural fit for a tree or a well-evidenced fishbone.
Key takeaways
- Match the method to the causal shape: chain, spread, or combination.
- 5 Whys is fast and drops parallel contributors. Use it on low potential severity.
- A fishbone opens the field and ranks nothing; it needs evidence gathering after it.
- A fault tree shows which single control breaks a path, and OSHA names it at 1910.119(e)(2)(vi) as a process hazard analysis methodology.
What none of the three do
A completed analysis changes nothing on the floor. Each method ends with candidate causes, and the work that reduces risk starts after that: selecting a control at the right level, assigning it an owner and a date, implementing it, and verifying afterwards that it holds. OSHA's guidance on hazard prevention and control is explicit about interim controls while a permanent fix is built and about follow-up to confirm effectiveness. Our guide to the hierarchy of controls covers how to rank the candidate.
None of the three establishes fact either. They organise reasoning about evidence that someone else has to collect: statements, photographs, maintenance history, the physical scene. An analysis built on a weak evidence base produces a confident diagram and a wrong answer. And none of them decides whether a proposed control is adequate for your site. That judgement belongs to the investigating team and the competent person accountable for the work.
Separately, investigating an incident is a different duty from recording one. Recordkeeping obligations under 29 CFR Part 1904 run on their own criteria and deadlines regardless of which analysis method you choose.
Where SoterAI fits
An investigator can describe the scene in conversation or upload photos and video from it, and have that information structured into the fields of an incident record. A qualified person reviews the record, adds the context the capture could not see, and confirms it before it closes.
If you are testing that workflow, use an investigation you have already completed. Check whether the captured evidence stays attached to the finding it supports, whether the record holds contributing factors in the plural, and whether a corrective action can carry an owner, a due date, and a later verification. Then check retrieval: a documented resolution is only worth having if it can be produced two years later. The incident record and corrective action plan pages show the fields each is stored in.
Sources
- OSHA: Incident investigationOSHA's guidance on investigating beyond immediate causes to system failures.
- 29 CFR 1910.119: Process safety managementSource for the 1910.119(m) investigation duties and for fault tree analysis at 1910.119(e)(2)(vi).
- OSHA: Hazard prevention and controlSupports the interim-control and follow-up verification points after an analysis.
- 29 CFR 1904.7: General recording criteriaSupports the separation between investigating an incident and recording one.
See how a reported incident becomes a structured, reviewable investigation record. Read the incident investigation workflow.