Numerical audit workbench

Check definitions, calculate differences and write only what the data support.

Try this: 20/100 versus 30/100 differ by ten percentage points, a 50% relative rise. The mean of 2, 3, 3, 4, 18 is six; its median is three. Both descriptions can be correct while answering different questions.

Calculations run in your browser; drafts are separate by language. The workbench needs JavaScript, but the checks remain readable without it.

1. Claim, population and definitions

Ask who was counted, when and under which definition. Numerator and denominator must refer to compatible units. Impressions are not unique people; respondents are not all residents.

Original claim, source URL and checking dateTarget population, recruitment, missingness and period

2. Rate change: B relative to AA event countA total countB event countB total count

Percentage-point change = (B rate − A rate) × 100; relative change = (B rate − A rate) / A rate × 100. Relative change is undefined when A is zero. No confidence interval or significance test is computed; dependence, clustering and design require separate treatment.

3. Center and spreadValues in one unit (commas or spaces; no thousands separators)

Mean is sum divided by count. Median is the sorted middle, averaging the middle pair for even counts. Range is maximum minus minimum. Investigate an outlier rather than deleting it without a reason.

4. Checks beyond arithmetic

Are populations, periods and definitions comparable?
Could missing or unreached people differ?
What groups, spread or extremes does the average hide?
How uncertain is the difference, and is causation overstated?

Gaps, comparability and data still neededRevised conclusion: what is established and what is inferred?