Simpson's Paradox
Every Simpson’s paradox involves at least three variables:
- the explained (y)
- the observed explanatory (x1)
- the lurking explanatory (x2)
If the effect of the observed explanatory variable on the explained variable changes directions when you account for the lurking explanatory variable, you’ve got a Simpson’s Paradox.
The classic example involves two doctors, A and B. Each has treated 100 patients, and doctor A has cured 90 while doctor B has cured 80. In this case, success rate is the dependent variable, and doctor is the observed explanatory. We conclude that doctor A is the superior doctor because he cured a higher percentage of his patients.
Only, not so fast. There’s another variable lurking in our sample: severity of illness. For simplicity, suppose each illness has one of two severities: low and high. We dig deeper into the data and see that doctor A treated 80 patients with low severity illness and cured 72 (90%) of them. Doctor B treated only 50 patients with a low severity illness and cured 48 (96%) of them. So doctor B actually does better with low-severity illness patients.
And here’s the amazing part: doctor B also does better with severely ill patients! Doctor A only cured 2 out of 20 severely ill patients; doctor B cured 32 out of 50. So doctor B is an all-around better doctor; he cures a lower percentage of his patients only because they are sicker.
In every example I’ve seen, both explanatory variables x1 and x2 are qualitative and each have a demonstrable and crucially, opposite, association with y. Problems arise when dividing the sample by both factors simultaneously yields unequally-sized classes. The unequal group sizes give a different weighted average among the classes of factor #2 for each class in factor #1, which can lead to seriously flawed conclusions.
As fellows at the Brookings Institute point out:
As our society grows more diverse, Simpson’s paradox may make more frequent appearances. Scholars and policy-makers will have to be mindful as they examine long-term changes of social and economic progress. It would be a shame if real progress in these areas was overlooked because of a naïve reliance on single averages.