The 38% figure: a correction on a published report
You are a statistical analyst in an EU agency. Ten days ago you published, with a senior colleague, Marc, the agency's flagship report on participation in cross-border programmes. While preparing next quarter's dataset, you discover that the underlying data double-count certain umbrella organisations that appear under several national registries. Correcting it would lower the headline figure from 38% to 34%. The report has already been cited in a Commission impact assessment, quoted in two national newspapers, and used by your Director in a speech. Marc, a twenty-year veteran who was the report's lead author, says the difference is “inside the margin of error, statistically and politically”, and asks you to wait until next year's edition. Your head of unit is seconded abroad for a fortnight; her deputy is on leave. A journalist has emailed asking for the raw data behind the 38% figure and for “any corrections you may have issued”.
1. What do you do in the first week, and who do you tell before you are certain?
Read the worked answer
The dangerous middle position is to keep verifying quietly and let the correction window stretch into months. You need certainty about the number and urgency about the process - in that order, but with a deadline on the first.
Verify properly, fast. Reproduce the calculation independently, quantify the effect exactly (38% to 34%), check whether the double-counting affects only the headline or also the sub-indicators, and check whether the report's methodology note already disclosed the limitation. Set yourself a fixed window - days, not weeks.
Inform someone with authority now, not after certainty. Your head of unit is away; the deputy is on leave. That does not mean nobody: the head of unit's own line manager or the Director's office, in writing, with a status of “under verification, not yet published”. Do not send it to a distribution list - send it to one named person and copy one other.
Preserve everything: versions of the dataset, the calculation script, the correspondence. If this becomes an audit or Ombudsman question, reconstructing it later will look like concealment.
Do not reply to the journalist yet - but do not let the query sit unanswered either. Tell the press office a request has arrived and that a factual position will be available shortly. The worst sequence is a journalist discovering a correction you knew about and did not disclose.
Do not agree with Marc to wait. You can agree to a short joint verification and to him leading the technical write-up. Those are different things.
2. Marc asks you to wait until next year's edition. How do you handle the conversation and the relationship?
Read the worked answer
Marc is not the obstacle to be overcome; he is the co-author who will make the correction credible or make it look contested. But the relationship does not set the standard.
Talk to him privately and early, before anyone else's version reaches him. Factual, non-accusatory, over the evidence rather than a conclusion: “here is what the dataset does, here is what it changes”.
Split the two questions. “Is this an error?” is verifiable and not a matter of opinion. “What do we do and when?” is a decision that must be taken within an agreed timeframe, not delegated indefinitely. Say that clearly and kindly.
Answer the margin-of-error argument properly, because it sounds plausible. Sampling uncertainty around a survey estimate is a different phenomenon from a definitional error in the base data; the first is disclosed, the second is not. Even if the effect were smaller, the deciding factor is not the size of the revision but the citation trail - an impact assessment, two newspapers, a Director's speech. A figure cited by others carries an obligation to correct it in their hands, not merely in the next edition.
Give him ownership. Offer him the lead in drafting the correction note and the technical annex. He has twenty years of standing; use it.
If he refuses to engage, do not go around him to the Director as a first step. Send him a short written summary of the verification and your proposed action with a date, and tell him you will need to raise it if there is no agreement by then. Then do so, factually.
3. It is confirmed. How do you correct the report, and what do you do about the Commission impact assessment that cites the 38% figure?
Read the worked answer
Correction, not silence. The cost of a correction is a news cycle; the cost of a discovered correction is the agency's credibility and the Commission's evidence base, and the second is larger and permanent.
Publish a proper erratum. A dated correction notice on the report's landing page: what changed, why, which tables are affected, the original figure shown alongside the corrected one. Do not quietly reissue the PDF; a silent revision is worse than the error.
Get it signed off through the quality function - the head of unit or her line manager - so the correction is institutional, not one analyst's initiative.
Tell the Commission service that cited it, in writing, before they find out. Corrected figure, explanation, link to the erratum, and an offer to brief. Do this proactively even though it is uncomfortable: a service that discovers a revision will not cite the agency again.
Brief the Director before the speech is repeated or quoted further. One page: what happened, what is published, what to say if asked.
Press: answer factually and briefly, refer to the erratum, and do not speculate about consequences for the impact assessment - that is the Commission's to assess. The press office handles it; you supply the facts.
Why not wait: statistical principles of impartiality, objectivity and transparency, and the agency's own data policy, all point the same way - users must be told. What would change the answer: if the effect were genuinely immaterial and affected no classification, it would still belong in the methodology note, not withheld.