Investigate disagreements systematically and prevent a preferred number from replacing the evidence trail.
Contradictions are often measurement information, not editorial defects. Two credible figures can differ because they cover different populations, periods, units, adjustments or release vintages. ONS defines coherence and comparability in terms of similarity across sources and comparability across time and domains; that makes method differences part of quality, not clutter [1]. OECD's metadata guidance separates conceptual, methodological and quality metadata, and defines a revision as a change to an already released statistic [2]. A contradiction log preserves each claim in its original construction, tests the dimensions of divergence, and records whether the conflict was reconciled, bounded or left open.
Record claims before explanations
Create one row per claim with exact value, unit, population, geography, reference period, release date, vintage, adjustment, source, method and claim type. Add the precise sentence or table location that supports it. Do not normalize currencies, annualize periods or select a preferred release in the evidence column. Those are transformations and should create derived rows with explicit formulas and provenance.
Pair apparently conflicting rows under a contradiction ID. Mark the dimension that first triggered the conflict: value, direction, timing, definition or status. A positive and negative growth rate is not automatically a contradiction if one is seasonally adjusted month-on-month and the other is unadjusted year-on-year. Likewise, a forecast and an observed result can disagree in outcome but are different claim types. The log prevents category errors before asking why numbers diverge.
Test a fixed divergence checklist
Compare scope, unit of analysis, inclusion rules, denominator, reference period, publication vintage, seasonal or inflation adjustment, weighting, imputation, currency, gross-versus-net treatment and sponsor. Then inspect common upstream data. Sources can present different outputs from one dataset because they apply different filters; other sources can present similar outputs despite genuinely independent methods. Preserve both facts.
Assign a disposition only after the checklist: compatible after transformation, different questions, superseded vintage, probable error, unresolved method difference, or genuinely conflicting observation. A superseded estimate is not deleted. Link it to the revision and note which decisions used the earlier vintage. OECD's definition makes revisions changes to released statistics, so the history is analytically relevant rather than embarrassing debris [2]. If a source corrects an error, distinguish that from a planned revision.
Decide with ranges and branch rules
Hypothetical example: Source A reports a $120 million market using supplier revenue; Source B reports $210 million using customer spending including integration services. Both are methodologically disclosed. The log marks different value, denominator and inclusion rule. It does not average them to $165 million. For a supplier-capacity decision, A is closer to the needed construct. For customer-budget planning, B may be closer. The choice is tied to use, not prestige.
If neither construct matches the decision, carry a range or hold the claim. Write a branch rule: the plan is viable under both figures, viable only under the broader figure, or blocked pending a new observation. This converts disagreement into sensitivity analysis. ONS's fitness-for-purpose framing supports choosing a measure for a stated use while still disclosing comparability limits [1]. Close the log only when the disposition, reasoning, reviewer and date are recorded; unresolved is a valid terminal state.
- Never overwrite an original claim with a normalized value.
- Classify the reason for divergence before choosing a measure.
- Tie any preferred construct to the decision it is fit to inform.
Take it into the meeting
- Preserve competing claims and their original methods.
- Do not average figures that answer different questions.
- Use branch decisions when disagreement remains material.
Sources & boundaries
Source statements are attributed; the decision process is Signal Atlas analysis. Examples marked hypothetical are teaching inputs, not observed outcomes.
- Source methodologies may be too thin to identify the cause of divergence.
- Fitness for one decision does not make a measure generally superior.
- Unresolved contradictions may require holding a conclusion rather than producing a single figure.
- Quality definedOffice for National Statistics · Source publication: 2020-10-08 · Retrieved 2026-09-19
Coherence and comparability are quality dimensions across sources, methods, time and domains. Statistical quality is evaluated as fitness for purpose.
- Data and Metadata Reporting and Presentation HandbookOECD · Source publication: not established · Retrieved 2026-09-19
Reference metadata should describe concepts, production methods and quality dimensions. A data revision is a change to a statistic already released.