Simpson's Paradox

Simpson’s paradox describes a reversal or disappearance of an association when data are aggregated over subgroups. For example, a treatment can show a higher success rate than a comparison within each severity group, yet a lower overall rate if the treated group contains more high-risk cases. Differences in the weights assigned to subgroups can change the aggregate. The pattern is an arithmetic property of the rates; it does not by itself say which comparison answers the decision question. The Berkeley graduate-admissions analysis by Bickel, Hammel, and O’Connell is a classic case. Aggregate acceptance rates appeared to favor male applicants, but department-level analysis showed much of the overall difference reflected application patterns across departments with different admission rates. The authors analyzed department-level data and discussed the limits of drawing a discrimination conclusion from aggregate proportions. This is not resolved simply by declaring that aggregate or stratified data are always correct. The relevant comparison depends on the question and causal structure. When a reversal appears, inspect denominators, subgroup composition, and the process assigning treatment or exposure. Consider whether the subgroup variable is a confounder, mediator, collider, or simply descriptive. Conditioning on a variable can reduce, create, or distort bias depending on the data-generating process. Report both aggregate and stratified estimates, explain their weights, and use study design or causal assumptions to choose an estimand. The paradox warns against inferring a causal story from a single summary table.