Lab 05 / Relationships

When every group wins but the total loses

Can B do better in both groups but worse overall?

7 min synthetic data fixed seed

Adjust the mix of two synthetic groups and see how different weights can reverse an aggregate comparison. Change one control at a time, then open the values under the chart before interpreting the picture.

Interactive synthetic lab
runs in this browser

View the values behind this chart

Loading the deterministic simulation...

SEED

What this display answers

An overall rate is a weighted average of group-specific rates. Two options can use different group weights, so an option that is ahead inside every group can still trail in the aggregate.

The reversal is arithmetic, not proof that the grouped result is causal. Choosing what to condition on requires subject-matter reasoning about how groups arise.

What to notice

  1. At a high novice share for B, confirm that B is five points ahead in both visible groups but behind overall.
  2. Lower B's novice share until the reversal disappears.
  3. Increase the within-group edge. A sufficiently large advantage can overcome the composition difference.

The model behind it

Option A always has 20% novice cases. Option B's mix is controlled. Baseline completion rates are 60% for novice cases and 90% for returning cases.

The model uses fractional expected counts to keep rates exact while weights move.

Where the result stops

A grouping variable can be a confounder, mediator, or collider; stratification is not automatically the correct analysis.

Neither the grouped nor aggregate association establishes a causal effect without a suitable design.