I feel like a lot of people don’t recognize how incredibly hard it is to check bias from data. The problem is that you have to define “subgroups” in a way that’s inherently arbitrary. This turns out to intersect with Simpson’s paradox in a weird way.

It’s a bit debatable if this is really a “data” problem that could be solved in principle by gathering enough data or more of a “philosophical” problem where we aren’t exactly sure what bias means.