Colquhoun.AI
← All guides

Research Guides

Keep Denominators With Rates Before Averaging Research

· Colquhoun.shop editorial

Two page groups can have perfectly accurate percentages and still produce a misleading combined headline. The problem is often not the source numbers but the choice to give a small group the same influence as a much larger one. Keep the underlying counts beside each rate before summarizing website research.

This guide provides a reproducible arithmetic check, not a claim about any analytics provider's metric definition. It does not establish causation, statistical significance or a new Colquhoun.AI feature.

1. Decide what the average should describe

“What is the average of our two page-group rates?” and “What fraction of all eligible events succeeded?” are different questions. The first treats each group as one observation. The second treats each eligible event as one observation. Label the intended question before choosing a spreadsheet formula.

NIST's weighted-mean reference defines a weighted mean as the sum of each value multiplied by its weight, divided by the sum of the weights. Choosing defensible weights remains the analyst's responsibility. For the pooled event-rate example below, the weights are the groups' eligible-event counts.

2. Recalculate a fictional report

Suppose a fictional publisher measures completed downloads out of eligible download attempts. Group A records 10 completions from 100 attempts, or 10%. Group B records 90 completions from 900 attempts, also 10%. Either averaging method happens to return 10%; this agreement does not prove the methods are interchangeable.

Now use a second fixture: Group A records 50 completions from 100 attempts, or 50%, while Group B still records 90 from 900, or 10%. The simple average of the two group rates is 30%. The pooled rate is 140 completions divided by 1,000 attempts, or 14%.

The equivalent denominator-weighted calculation is (100 × 0.50 + 900 × 0.10) ÷ (100 + 900) = 0.14. Neither result needs to be concealed. The 30% figure answers an equal-group question; it does not describe the fraction of all attempts completed in this fixture.

3. Keep an aggregation ledger

Prefer the original counts over reconstructing them from rounded percentages. A displayed 14% does not reveal whether the source recorded 140 out of 1,000 or a nearby ratio rounded for presentation. Keep display rounding separate from the calculation record.

4. Reject incompatible pooling

Before adding counts, establish that the groups are mutually exclusive for the observation being counted. A visitor who appears in both device groups cannot automatically be counted twice in a combined unique-person denominator. Likewise, sessions, impressions and users are not interchangeable weights.

A group with zero eligible events has no defined rate under numerator divided by denominator. Mark that condition explicitly rather than silently assigning a zero-percent performance result. Missing counts also are not zero counts. Request the source totals or report that the pooled figure cannot be established.

5. Publish the calculation boundary

A useful review note reads: “For these two nonoverlapping groups in the stated period, 140 of 1,000 eligible attempts completed, yielding 14%. The equal-group average is 30% and is not our overall event rate.” Include the counts so another reviewer can reproduce the result.

Finally, inspect each group before interpreting a change in the combined rate. A different mix of large and small groups can change the aggregate without demonstrating an improvement within either group. Correct aggregation makes the headline auditable; it does not, by itself, explain why the outcome changed.

This guide describes the built-in Web Reader and Keyword Density behavior verified for this publication. Available data depends on the page response; these tools do not measure visits or search rankings.