Short answer: Simpson’s paradox occurs when a trend found in separate groups reverses after the groups are combined. The arithmetic can be perfectly correct. Unequal group sizes and a third variable change the weights in the overall result, so the combined percentage answers a different question from the subgroup percentages.
This matters whenever a dashboard, headline, or experiment compresses many kinds of people or cases into one number. A total can be useful, but it can also hide how the observations were distributed.
Simpson’s paradox in one sentence
A comparison favors A within every relevant subgroup, yet the aggregated data favor B—or the reverse. The reversal often looks impossible until you inspect the denominators.
The phenomenon is easier to understand after reviewing the basics of data analysis. A rate always combines a numerator with a denominator. When the denominators differ sharply across groups, the combined rate gives those groups different weights.
A worked example: two treatments and two clinics
Suppose two clinics treat patients with different initial risk levels. Treatment Y has a higher recovery rate in both clinics, but Treatment X appears better overall.
| Clinic group | Treatment X | X rate | Treatment Y | Y rate |
|---|---|---|---|---|
| Lower-risk cases | 81 of 90 | 90% | 19 of 20 | 95% |
| Higher-risk cases | 8 of 20 | 40% | 45 of 90 | 50% |
| Combined | 89 of 110 | 80.9% | 64 of 110 | 58.2% |
Within lower-risk cases, Y leads by 5 percentage points. Within higher-risk cases, Y leads by 10 points. Nevertheless, X leads by more than 22 points in the combined table.
Where did the reversal come from?
Treatment X was used mostly for lower-risk patients, who had high recovery rates under either treatment. Treatment Y was used mostly for higher-risk patients. The overall figures therefore compare two different mixtures of patients.
The calculation is:
Combined rate = total successes ÷ total observations.
It can also be written as:
Combined rate = Σ (subgroup weight × subgroup rate).
For X, lower-risk cases receive a weight of 90/110. For Y, the same group receives a weight of only 20/110. Those unequal weights, not a mathematical error, produce the reversal. The lesson builds directly on the idea of the arithmetic mean, especially a weighted mean.
The hidden role of a third variable
The grouping variable may be a confounder when it is related both to which option a case receives and to the outcome. In the clinic example, initial risk relates to treatment allocation and recovery. Ignoring risk mixes two different comparisons.
Still, analysts should not split data by every available label. A subgroup must be meaningful for the process and the question. Grouping by an effect of the treatment, for example, can introduce a new bias. Sound analysis combines mathematics with subject knowledge.
E. H. Simpson’s 1951 paper examined interaction in contingency tables. The idea became widely known through later examples, including the 1973 University of California, Berkeley graduate admissions data. The aggregate figures suggested one pattern, while department-level comparisons revealed a different structure. Readers can consult Simpson’s original paper and the Berkeley admissions study.
A practical checklist for detecting Simpson’s paradox
- Define the comparison. State the options, outcome, population, and time period.
- Keep the counts. Record numerators and denominators, not only percentages.
- Inspect the overall result. Calculate the difference and note its direction.
- Choose plausible groups. Consider baseline risk, department, location, age band, device, difficulty, or another pre-existing factor.
- Recalculate within groups. Use the same outcome definition everywhere.
- Compare the weights. Ask whether A and B contain different proportions of easy and difficult cases.
- Plot the results. A dot plot or small-multiple chart often exposes the reversal quickly.
- Explain the mechanism. A reversal alone does not identify a cause.
- Report both views. Present totals, subgroup rates, sample sizes, and the interpretation.
For a broader workflow, see data analysis and data science. When a metric guides a business decision, the same discipline also improves the use of key statistics.
Which result should you trust?
Neither level wins automatically. The overall result describes the observed mixture. A subgroup result describes a conditional comparison. The right level depends on the question.
If a hospital needs to predict total bed demand for its actual patient mix, an aggregate figure may be useful. If researchers want to compare treatments for patients with similar initial risk, stratified or adjusted estimates are more relevant. A causal claim requires additional assumptions and an appropriate study design; a table alone cannot prove what would happen if treatment assignments changed.
That distinction also explains why “control for everything” is poor advice. Analysts need a causal or substantive reason for including a variable. Otherwise, adjustment can remove part of the effect of interest or create selection bias.
Common mistakes
- Looking only at percentages. A percentage without its denominator hides the weight.
- Calling every subgroup difference a paradox. Simpson’s paradox involves a meaningful reversal, not ordinary variation.
- Assuming the aggregate is dishonest. It may answer a legitimate but different question.
- Assuming the subgroup result is causal. Stratification does not replace randomization or careful causal reasoning.
- Choosing groups after seeing the answer. This invites cherry-picking.
- Ignoring uncertainty. Small groups can show unstable rates, so confidence intervals or other uncertainty measures may matter.
- Using inconsistent definitions. Success, population, and time window must match across groups.
Where reversals appear
Possible settings include medical outcomes, university admissions, sports averages, loan approvals, advertising conversion rates, school performance, and branch-level sales. For example, every device category may improve its conversion rate while the company’s total conversion rate falls. A shift toward the device with a lower baseline rate can outweigh each within-device improvement.
That scenario is not automatically proof of Simpson’s paradox. Verify that the direction actually reverses and that the change in group weights explains it. A teaching case from the Journal of Statistics Education, indexed by ERIC, offers additional classroom practice.
Quick quiz
- What defines Simpson’s paradox?
Answer: A relevant relationship reverses when groups are combined or separated. Explanation: the key feature is direction, not merely a smaller effect. - Why do denominators matter?
Answer: They determine each subgroup’s weight in the combined rate. - Which treatment wins within both clinic groups?
Answer: Treatment Y, at 95% versus 90% and 50% versus 40%. - Which treatment wins overall?
Answer: Treatment X, at 80.9% versus 58.2%. - Was the overall calculation wrong?
Answer: No. It compared different mixtures of lower- and higher-risk patients. - Does a reversal prove confounding?
Answer: No. It is a signal to investigate group composition and the data-generating process. - Should analysts publish only the subgroup table?
Answer: Usually no. Transparent reporting shows both levels and explains which question each answers. - What is the safest first check?
Answer: Recalculate the counts and rates, then compare subgroup sizes and allocation. - Can the paradox affect averages?
Answer: Yes. Unequal weights can reverse comparisons of averages as well as proportions.
Frequently asked questions
Is Simpson’s paradox the same as a spurious correlation?
No. A spurious correlation is a broader idea: an association can appear because of coincidence or another variable. Simpson’s paradox specifically concerns a reversal caused by aggregating or separating groups.
Can software detect it automatically?
Software can scan candidate groupings, but it cannot decide which variables are scientifically or ethically appropriate. Human judgment remains essential.
How should a dashboard handle it?
Show counts, allow meaningful segmentation, flag major composition shifts, and keep definitions consistent. A short note about changing weights can prevent a confident but misleading interpretation.
Conclusion
Simpson’s paradox is a warning against treating one summary number as the whole story. Check the counts, compare plausible subgroups, and ask what changed in the mixture. When analysts report both aggregate and conditional results, readers can see not only what the numbers say, but also why they say it.
Read the Indonesian companion article: Paradoks Simpson: Mengapa Kesimpulan Data Bisa Berbalik.
