Skip to main content

How much do you want to avoid Simpson's paradox?

Simpson's paradox is the reversal of the direction of association between two variables upon conditioning on a third one. For example, a new treatment may appear better than the old one based on the whole set of data, still, the old treatment may seem preferable for both men and women. This 'paradoxical' situation has a large literature. The talk will show that the possibility of Simpson's paradox is a consequence of using the odds ratio (which fails to take into account the allocations into the treatment categories) to determine which treatment is better. Subject to mild assumptions, there is only one way to reading off the better treatment from the data that avoids the paradoxical reversal. Those unhappy with Simpson's paradox, should give up using the odds ratio, at least in observational studies where treatment allocation is informative.