One of the more surprising concepts in statistical analysis is Simpson’s Paradox. It occurs when a relationship observed in aggregated (or marginal) data reverses direction after the data are separated into meaningful subgroups (or conditional data).
In simple terms, the overall data may suggest that Option 1 performs better than Option 2. However, when the same data are examined within each relevant category, Option 2 may actually outperform Option 1 in every subgroup. The apparent contradiction arises because the groups being combined are not represented equally, allowing an underlying factor to distort the overall result.
This phenomenon is particularly important in categorical data analysis, where outcomes are often summarized in contingency tables. As Alan Agresti notes in his well-known text on categorical methods, Simpson’s Paradox occurs when the marginal association has a different direction from the conditional association.
A simple everyday example involves choosing the fastest route to work. Imagine that over several months you record whether two different commuting routes get you to work in less than 30 minutes. Looking only at the overall data in the figure below, Route B appears to be the better choice (see dashed line). However, after separating the trips by time of departure, a different picture emerges: Route A actually performs better whether you leave before or after the morning rush begins.
How can this happen? In this example, the departure time acts as a lurking variable. Route A may have been used much more frequently during the peak rush-hour period, while Route B was used primarily before traffic increased. Since traffic volume is closely related to departure time, combining all of the observations masks the underlying relationship and produces a misleading overall conclusion.
Why does this matter? Because decisions are often made using summary statistics or aggregated dashboards. If important explanatory variables—such as departure time, operating environment, product type, customer segment, or production batch—are ignored, the overall conclusion may not accurately reflect what is happening within the individual groups. As a result, organizations may direct resources toward the wrong corrective actions or make incorrect comparisons between alternatives.
The lesson is not that aggregated data are “wrong,” but rather that they may be incomplete. Before drawing conclusions, analysts should ask whether there are natural subgroups or confounding variables that should be considered. Examining both the overall data and the conditional relationships can help ensure that the statistical story is both accurate and meaningful.
Simpson’s Paradox serves as a valuable reminder that in statistics, context matters. Looking one level deeper into the data can often prevent a misleading conclusion and lead to better-informed decisions.