I learned about the Simpson’s paradox fairly recently, and I found it quite disturbing, not because of the mere “paradox” itself, but mainly because I felt it was something I should have known already.
In case you haven’t heard about it, one instance of the paradox is a real-world medical study for comparing the success rate of two treatments for kidney stones (from Wikipedia):
Overall, Treatment B is better because its success rate is 83%, compared to 78% of Treatment A. However, when they split the patients into 2 groups: those with small stones and those with large stones, then Treatment A is better than Treatment B in both subgroups. Paradoxical enough?
Well, it’s not. It turns out that for severe cases (large stones), doctors tend to give the more effective Treatment A, while for milder cases with small stones, they tend to give the inferior Treatment B. Therefore the sum is dominant by group 2 and group 3, while the other groups contribute little to the final sums. So the results can be interpreted more accurately as: when Treatment B is more frequently applied to less severe cases, it can appear to be more effective.
Now, knowing that Treatment and Stone size are not independent, this should not come up as a paradox. In fact, we can visualize the problem as a graphical model like this
All the numbers in the table above can be expressed as conditional probabilities like so:
- Group 1:
- Group 2:
- Group 3:
- Group 4:
For any of us who studied Probability, it is no surprise that the probabilities might turn up-side-down whenever some conditional variables are stripped out of the equations. In this particular case, since S depends on both St and T, the last 2 equations do not bring any new knowledge about S.
So what is this “paradox” about? Isn’t it nothing more than the problem of confounding/lurking variables, something that most people in Probability/Statistics already known? In this particular case, Stone size is the lurking variable that dictates both Treatment and Success, therefore the scientists who designed the experiment should have taken it into account since the beginning. It is well-known among Statistic practitioners that they must try their best to identify and eliminate the effect of any lurking variables in their experiments, or at least keep them fixed, before drawing any meaningful conclusion.
From a slightly different perspective, the paradox can be understood once we understand the human bias of drawing causal relations. Human, perhaps for the sake of survival, constantly look for causal relations and often tend to ignore rates or proportions. Once we conceived something as being causal (Treatment B gives higher success rate than Treatment A in general), which might be wrong, we continue to assume a causal relation and proceed with that assumption in mind. Obviously with this assumption, we will find the success rates for the subgroups of patients to be highly counter-intuitive, or even paradoxical.
In fact, the connection of this paradox to human intuitions is so important that Judea Pearl dedicated a whole section in his book for it. Modern Statistical textbooks and curriculum, however, don’t even mention it. Instead they will generally present the topic along with lurking/confounding variables.
Therefore, if you haven’t heard about this, it is probably for a good reason, or perhaps you are simply too young.