Simpson's Paradox — Definition, Formula & Example
Simpson's Paradox is what happens when one group does better in every separate category but worse once all the categories are combined. The reversal usually comes from a lurking variable that changes how the groups are split up.
Let be a categorical variable with levels . Simpson's Paradox occurs when for every level , yet for the aggregated data. This is possible because each aggregated rate is a weighted average of the subgroup rates, and and can have different weights.
Key Formula
Where:
- = Success rates within subgroup 1 and subgroup 2
- = Number of cases in each subgroup, which act as the weights
- = Combined success rate across both subgroups
Worked Example
Problem: Two surgeons each perform 100 operations. Dr. A: 18 of 20 easy cases succeed and 32 of 80 hard cases succeed. Dr. B: 68 of 80 easy cases succeed and 7 of 20 hard cases succeed. Who has the better success rate?
Step 1: Compare the surgeons within the easy cases.
Step 2: Compare them within the hard cases. Dr. A wins again.
Step 3: Combine the categories. The comparison flips. Dr. A took mostly hard cases (80 of 100), so the low hard-case rate gets most of the weight in Dr. A's overall rate.
Answer: Dr. A has the higher success rate for both easy and hard cases, but Dr. B has the higher overall rate (75% vs. 50%). Case difficulty is the lurking variable. Dr. A has the higher observed success rate at each difficulty level.
Visualization
Why It Matters
Simpson's Paradox appears in two-way-table questions like those on AP Statistics, where you must explain how a lurking variable reverses a comparison. Medical researchers, college admissions analysts, and sports statisticians check for it before ranking treatments, departments, or players based on combined percentages.
Common Mistakes
Mistake: Trusting the combined percentage without checking how each group is distributed across the categories.
Correction: Look at the subgroup counts. If one group has far more cases in a hard or low-success category, its overall rate is pulled down by weighting rather than worse performance. Compare within each level of the lurking variable.
