You Assumed

Reference path · conclusion first

Simpson's Paradox: Can Every Segment Improve While the Total Gets Worse?

Aggregate rates move with both segment performance and segment weights.

Precise claimAn aggregate rate is a weighted result: every segment can improve while the total falls when the population mix shifts toward a lower-rate segment.

Applies

A synthetic two-period, two-device completion-rate fixture with exact integer counts, fixed segment definitions and a controlled new-period traffic-share slider.

Does not prove

The page does not determine causality, prescribe which variables to condition on, claim segments always outrank aggregates, or generalize the synthetic rates to a real product.

Portable rule

If an aggregate rate changes, then inspect both segment rates and segment weights before attributing the movement to performance.

Evidence reviewed 2026-07-25. Corrections and review policy.

Use as a reference

Declare the discussion context, then copy a stable link. No account or personal data is attached.

aggregation-illusion puzzle

Simpson's Paradox: Can Every Segment Improve While the Total Gets Worse?

Judge an exact product-metric table before the totals are combined, then keep all segment rates fixed and change only the traffic mix to inspect what the aggregate can support.

Audit ledger · one product, two periods; successes and sessions are fully visible.
SegmentOldNewChange
Desktop480 / 60080%340 / 40085%+5 pp ↑
Mobile160 / 40040%270 / 60045%+5 pp ↑
All sessionsHidden until you commit — no rows or counts are missing.
What happened to the combined completion rate?

Take it into the discussion

Need the conclusion, boundary and sources without the challenge?

Open the same page in its conclusion-first reading state, then declare the work context before copying a stable reference.

Use as a reference →