Aggregation reversal
Simpson's Paradox
A trend that appears within every relevant group can weaken, disappear, or reverse after the groups are combined with different weights.
aggregate rate = sum(group weight x group rate)
The reversal is arithmetic; whether conditioning is appropriate is a causal question. A confounder, mediator, or collider must not be treated interchangeably.
Two treatments retain their within-group rates while the share of easy and difficult cases changes, allowing the aggregate comparison to reverse.
(%)
The animation runs automatically, pauses on the conclusion, and then repeats. The main control changes the scenario rather than scrubbing the timeline.
- CHANGE
- Group-mix imbalance
- WATCH
- aggregate direction
- MEANING
- Two treatments retain their within-group rates while the share of easy and difficult cases changes, allowing the aggregate comparison to reverse.
The denominator mix can overpower every within-group comparison.
Two small-multiple rate panels feed a weighted aggregate scale, keeping numerators, denominators, and group composition visible.
What it actually says
Simpson's paradox occurs because aggregated rates are weighted averages and the weights can differ across the compared populations. A treatment used mainly in difficult cases can look worse overall despite doing better within each severity group.
There is no universal command to prefer the aggregate or the stratified result. The estimand and causal structure determine which comparison answers the question.
"A useful law compresses a pattern. It does not erase the conditions that make the pattern true."
How the idea developed
The modern form emerged through observation, argument, and later refinement. The timeline separates the first insight from the version now used in textbooks and practice.[1]
Yule discusses association reversals in contingency tables.
Edward Simpson analyzes interactions in contingency tables.
Causal diagrams clarify when adjustment creates or removes bias.
How the pattern works
The relation becomes useful only when its mechanism, measurement process, and operating range are visible.
Groups begin with different outcome rates.
Compared populations contain different group proportions.
The dominant group changes the aggregate.
The reversal is arithmetic; whether conditioning is appropriate is a causal question. A confounder, mediator, or collider must not be treated interchangeably.
Where it earns its keep
Applications are strongest when the law changes a decision, measurement, model, or experiment rather than merely providing an analogy.
Stratify by justified risk factors
ApplicationCase mix can reverse hospital or treatment rankings.
Define the target population.
Show counts with percentages
ApplicationRates alone hide their weights.
Use a causal model before adjustment.
Where it stops working
Not every change after stratification is a paradox, and conditioning on a collider or mediator can introduce rather than remove bias.
"Disaggregated results are always true"
Better: Conditioning can answer a different or biased question."Statistics are contradictory"
Better: The results describe different weighted populations.Sources and further reading
Original publications and serious secondary scholarship are prioritized over summaries.
- Simpson - The Interpretation of Interaction in Contingency TablesPrimary 1951 article.https://doi.org/10.1111/j.2517-6161.1951.tb00088.x
- Pearl - Understanding Simpson's ParadoxCausal interpretation.https://doi.org/10.1080/01621459.2013.820457
- Stanford Encyclopedia - Causal ModelsCausal-graph background.https://plato.stanford.edu/entries/causal-models/