Learn it → Apply it → Save as Skill
A trend that holds in every subgroup can reverse once the subgroups are combined. When that happens, the arithmetic has not broken. Something about how the groups are sized, and what is steering people or cases into one group over another, is doing more work than the top-line number lets on.
The Dr. Wallace's Favorite Mental Models is designed to turn useful ways of thinking into reusable tools. Each model gives you a fast visual overview, deeper interactive learning, practical application exercises and an AI-ready Markdown skill you can save for future use.
Instead of merely reading about mental models, you can gradually build your own AI skills library: a collection of portable reasoning skills that can help ChatGPT, Claude, Grok and other AI assistants apply the mental models you value.
Whenever two groups are folded into one summary number, the combined result depends not only on each group's rate but on how many observations came from each group. A subgroup that wins every individual comparison can still lose overall, if it was weighted toward the harder cases.
Treatment A beats Treatment B on small stones, and beats it again on large stones.
Treatment B looks better overall, because doctors used Treatment A far more often on the harder cases.
Important: the reversal is almost always caused by unequal group sizes meeting unequal difficulty, not by an error in either set of numbers.
The paradox is most likely to appear when you are comparing two groups across categories that differ in size, such as departments, time periods, hospitals, sales territories or customer segments.
The check is a habit, repeated on purpose whenever a comparison between two groups is about to drive a real decision.
Break the combined number back into its natural subgroups instead of trusting the headline figure alone.
Check whether the winner within each subgroup matches the winner in the combined total.
Look at how many observations sit in each subgroup. Unequal sizes are almost always the hidden driver of a reversal.
Decide which comparison should actually guide the decision, the subgroup rates or the weighted total, based on what you are really trying to measure.
Open a use case to see what the model changes.
A rep can have a higher close rate on enterprise deals and a higher close rate on small deals than a colleague, yet a lower close rate overall, if more of their deals sat in the harder enterprise pool. Compare rates within each deal size before ranking reps on a blended number.
Open surgery outperformed a newer, less invasive procedure on both small kidney stones and large kidney stones in a real 1986 comparison, yet the newer procedure appeared better in the combined total, because surgeons reserved open surgery for the harder large-stone cases. See the worked example below.
Age-adjusted comparisons across regions can reverse once you account for the fact that one region simply has an older population. Always check whether an unadjusted headline rate is being distorted by a difference in the underlying age or risk mix.
An overall admission rate can favor men while most individual departments favor women, if women applied disproportionately to more competitive departments. The department-level numbers and the combined number can tell opposite stories at once. See History and Accuracy below.
A daily-updated A/B test dashboard can show the new version winning every single day, while the pooled weekly total favors the old version, if traffic mix shifted across days. Check the day-by-day breakdown before declaring a winner from the cumulative total.
Outperforming in every calendar year does not guarantee a higher compounded total if contribution timing and amounts differ between accounts. Compare dollar-weighted growth, not only year-by-year percentage wins.
This is a real, published dataset used to teach the method, drawn from Charig et al. (1986) and the reanalysis in Julious and Mullee (1994), both in the BMJ.
Expand for the actual sequence of who found what, and when the modern label was attached.
Not exactly. The statistician George Udny Yule described the same reversal in contingency tables as early as 1903. Edward H. Simpson gave it a more thorough mathematical treatment in his 1951 paper "The Interpretation of Interaction in Contingency Tables," published in the Journal of the Royal Statistical Society, without claiming to have found a new phenomenon. The statistician Colin R. Blyth proposed calling it "Simpson's Paradox" in a 1972 paper, and the name stuck. Many statisticians still call it the Yule-Simpson effect for this reason.
In 1973, the University of California, Berkeley was examined for gender bias in graduate admissions after men were admitted at roughly 44% and women at roughly 35% overall. Statisticians Peter Bickel, Eugene Hammel and J. William O'Connell reanalyzed the data department by department in a 1975 Science paper, "Sex Bias in Graduate Admissions: Data from Berkeley," and found that most individual departments showed a small bias in favor of women, or no bias at all. Women had applied in greater numbers to departments that were more competitive for everyone, which dragged down their combined admission rate.
They are closely related but not identical. Simpson's paradox is the specific, visible symptom that appears when a confounding variable is strong enough to fully reverse the direction of a comparison, not merely shrink or inflate it. Every case of the paradox involves a lurking variable correlated with both the grouping and the outcome, such as stone size or department selectivity, but not every confound is severe enough to flip the sign of the result.
Combined totals get quoted casually in headlines and dashboards. Look for the natural subgroups before repeating a conclusion built on one blended figure.
Sometimes the combined total genuinely is the number that matters, such as total company revenue. The paradox is a prompt to check which number answers your real question, not an automatic instruction to split the data.
A reversal almost always comes from unequal group sizes meeting unequal difficulty or unequal rates. If the groups were the same size, the paradox could not appear.
Slicing a dataset into smaller and smaller categories eventually leaves too few observations per cell to mean anything. Choose natural, meaningful categories rather than slicing until you get the answer you want.
The Berkeley case shows that a purely structural, statistical pattern can produce a reversed headline number without anyone intending bias. The paradox alone does not tell you whether real bias is also present; it tells you to look closer.
Daily win-rate dashboards, cohort reports and A/B test summaries are common places for an unnoticed reversal to quietly steer a decision.
you specifically care about the total outcome, such as total revenue or total patients treated, rather than a rate within any one segment.
the subgroup split is arbitrary, was chosen after already seeing the result, or does not correspond to any real difference between cases.
the direction of the reversal lines up with a variable you already have reason to believe is causally relevant, such as case severity or applicant selectivity.
This browser tool structures your thinking. It does not send your information anywhere or perform AI analysis.
Save as Skill converts the model into portable Markdown instructions that teach an AI when and how to check for Simpson's paradox, what questions to ask, what evidence to seek, what mistakes to avoid, and how to structure its analysis.
Save the skills you find useful and organize them into your own reusable library. The Markdown is intentionally model-neutral, so it can be used wherever an AI system accepts uploaded files, project knowledge, custom instructions or reusable text context. Platform capabilities differ, so this is a portable skill file rather than a claim of one-click installation.
It is not a logical contradiction. Every number involved is correct; nothing is miscalculated. It earns the name "paradox" because the result feels impossible at first glance: how can something lose in every direct comparison yet win overall? The resolution is that a weighted average is not the same operation as comparing two rates, and the weights themselves carry information.
It can happen with any number of subgroups and is not limited to two categories being compared. The same mechanism, unequal group sizes interacting with unequal rates, can produce a reversal across three age brackets, five regions, or any other partition, and in principle a reversal can even flip back again at a further level of subgrouping.
They are related but answer different questions. "Correlation is not causation" warns that an association between two variables might not mean one causes the other. Simpson's paradox is more specific: it shows that the very sign of an association, positive or negative, can depend entirely on which variables you choose to hold constant while comparing. It is a reason a causal claim can look true or false purely based on how the data was grouped.