Municipal records teach you early that a total is a story somebody chose to tell. A clerk adds up the permits, the complaints, the repairs, and the sum goes into a report. Nobody asks what the pieces looked like before they were added. Simpson's paradox is what happens when somebody does ask, and the pieces disagree with the sum.
What the paradox says
The statistical version is plain enough. An association between two variables in a population can appear, disappear, or reverse when that population is divided into subpopulations. Two things may look positively linked overall and yet be independent, or even negatively linked, inside every subgroup. The Stanford Encyclopedia of Philosophy notes that the paradox is unproblematic from the standpoint of mathematics and probability theory, yet many people still find it surprising.
The surprise has a source. We expect an overall rate to land somewhere among the group rates and to keep their ordering. But a pooled rate is a weighted average, and the weights are the group sizes. When the sizes differ, the ordering can flip. Nothing is miscounted. The arithmetic is simply doing what it always does.
Worked examples
Simpson's own 1951 illustration involved 52 patients. Overall, treatment and control both succeeded 50% of the time. Among men, treatment won, about 61% to 57%. Among women, treatment won again, about 44% to 40%. Better in each sex, no better in the pool, because the patients were spread unevenly between the arms.
A restaurant version, described by Science ABC, is easier to picture. The Orchard was preferred by 450 of 500 diners, or 90%. The Bistro got 420 of 500, or 84%. Yet The Bistro led among younger diners, 81.5% to 80%, and among older diners, 94% to 92.5%. The difference sat in who ate where, not in how either place was judged within an age group.
Why it happens
Look for a third variable that is unevenly distributed across the groups being compared. In the trial it was sex. In the restaurants it was age. Once the groups differ on that hidden factor in different proportions, the pooled figure starts reporting the makeup of the crowd instead of the thing you meant to measure.
The mechanism is modest. Suppose one option beats its rival in every subgroup, but most of its cases fall in the subgroup where rates run low. Its overall rate can sink below the rival's. Group sizes do the work. No single number is wrong.
A weekend of problem sets
The cleanest demonstration comes from the Brilliant wiki. On Saturday you solve 7 of 8 problems, and your friend solves 2 of 2. On Sunday you solve 1 of 2, and your friend solves 5 of 8. Your friend is more accurate on both days. Over the weekend, though, you solved 8 of 10 and your friend 7 of 10.
Read the schedule and the reversal explains itself. Your attempts piled up on the easy day. Your friend's piled up on the hard one. The totals compare workloads as much as skill.
Which total should you trust?
There is no universal answer. If you want to know what happens to a whole population under a policy, the combined figure may be the relevant one, so long as the mix stays the same. If you want to know whether an option is better for a person of a given type, the breakdown usually says more.
Causal questions need more than choosing a level. Researchers in epidemiology, decision theory, and causal inference use tools such as causal diagrams to decide which variables to control. A casual reader should treat any claim that leans on a hidden variable as open until that variable has been considered.
Questions to ask and red flags to watch
Ask whether the groups are of similar size and composition. If one option was tried mostly on easy cases and the other mostly on hard ones, the pooled comparison is suspect. Ask whether a third variable differs between groups, and whether the pattern holds inside every subpopulation or only in the combined data.
Be wary of a headline that compares pooled rates with no breakdown, especially when subgroup sizes go unstated. The missing table is itself a reason to ask for the table.
Limits of what is known
The sources here do not give a procedure for deciding which level of analysis is causally correct, and they do not describe methods for adjusting or stratifying data. They also do not estimate how often the paradox turns up in published research, and the fuller Berkeley admissions and COVID-19 discussions are not covered. Those are the places to look next, with a more specialized text in hand.
Sources
Marcus Adeyemi-Holt is a house writer for Deeperpectives, where he examines how ordinary routines reveal larger patterns in work, memory, and place. He draws on years of reporting on municipal records and oral histories to trace the quiet forces that shape everyday life.
