Measurement
Part of Lead network measurement that preserves the state ledger
Lead network measurement: correction or revision when a figure changes
Lead network measurement mistakes need a correction, new data needs a revision. How to tell them apart, label each and republish the affected periods.
A published lead network measurement can change for two different reasons, and the notice must say which. A correction fixes a mistake in what was published. A revision brings in better methods or data that did not exist at first publication. Mixing the labels hides errors behind ordinary updates.
What to take away
- A correction answers a mistake. A revision answers new data or a better method.
- Freeze the affected output and preserve its query before anything is changed.
- Name every period that changed, with the reason, in the notice.
- Independent assurance reviews the repair before it is republished.
Correction or revision
The ONS says revisions should not be confused with the correction of errors: a revision is not published to correct a mistake. The policy covers official statistics, but the distinction helps private teams avoid disguising errors as ordinary updates.
In a lead network report, a correction follows the discovery of a mistake after publication. A revision follows improved methods or additional data that were unavailable when the figure first appeared. A mistake never becomes a revision because the update looks routine.
Six situations, labelled
Five of these six situations are corrections. Only data that genuinely arrived after publication is a revision. Match the situation to a label before writing the notice.
| Situation | Label | Repair |
|---|---|---|
| Every submission was called a lead, hiding quarantine, failed validation, no-match and non-disclosure | Correction | Rebuild the population with each state kept separate |
| Failed, late or disputed records were removed from the denominator without a note | Correction | Publish eligibility, numerator, denominator, exclusions and query version |
| Events arrived after the period closed | Revision | State the new data and the periods it changed |
| A recipient's disclosed events were missing because the event route failed | Correction | Repair the route and republish the affected periods |
| Attribution credit was reported as causal contribution | Correction | Restate the claim as attribution, or support it with a comparison |
| A supplier case study and an England cohort were pooled as one benchmark | Correction | Separate them and keep supplier figures as attributed claims |
The repair routine
- Freeze the affected output and preserve its query.
- Publish the nature of the limitation while the repair is under way.
- Reconcile states across the ledger, then correct the code or data under change control.
- Run independent assurance on the repair.
- Republish the affected periods with a note stating which periods changed.
Step four has a defined owner in the government AQuA Book. The analyst has primary responsibility for verification and validation, and the assurer is responsible for reviewing that work. A repair written and approved by the same person has had no independent review.
A calculation should stop when loss crosses the organisation's approved tolerance, not continue with an unlabelled partial population.
What the notice must say
A notice should let a reader reproduce the change. Publish eligibility, numerator, denominator, period, time zone, exclusions, query version and the revision reason for each affected rate.
Record which states are disproportionately missing. If disclosed events are absent for one recipient, the bias differs from random intake loss. The response might hold a recipient comparison, repair the event route and republish the affected periods with a correction note.
Where the corrected claim treated credit as cause, say so plainly. Attribution rules allocate recorded credit but do not construct a counterfactual, so a causal claim needs a defensible comparison and stated uncertainty.
Benchmarks and baselines after a repair
If the repaired figure sat beside a benchmark, check the comparison again. Geography, service definition, source universe, sample, period and denominator must align. If they do not, keep the results separate.
An internal baseline can still be useful if its query, operating context and version remain stable. Label it as organisational history, name material changes and avoid calling it an industry norm.
Show the repair where readers look
Put freshness, reconciliation and known loss before the performance view, so a corrected period is flagged where readers already look. The dashboard bands are the natural place for that flag. Keep failed records in visible failure queues, because a repair needs those records to restate the period.
Privacy, PECR, security, accessibility, consumer, advertising, commercial and statistical gates remain independent of the repair. CAP's misleading advertising rules also mean an internal chart cannot automatically substantiate a public performance claim, so run advertising checks before a corrected figure is quoted outside the team.
Before you act
- Decide whether the change is a correction or a revision
- Freeze the affected output and preserve its query
- Reconcile states and correct under change control
- Get independent review of the repair
- List every period that changed in the notice
- Record an owner and review trigger for the error
Common questions
Is a late-arriving event a correction?
Not by itself. Data that was unavailable when the figure was published is a revision. If the event route failed to capture it, repairing the route is a correction, and the affected periods are republished with a correction note.
When should a calculation stop?
When loss crosses the organisation's approved tolerance. Continuing with an unlabelled partial population turns a known gap into a hidden one.
Who reviews the repair?
Someone other than the analyst who made it. In the AQuA Book the analyst conducts verification and validation, and the assurer reviews them. The repair is republished only after that review.
