I spoke with @Jdforrester-WMF and apparently there's a lot of existing data about people opting in and out of preferences, but it's never really been meaningfully analysed. It might all be junk. We should check that data first to see if it's any good. If not, then we should implement extra data collection.
Presumably, the data lives in the PrefUpdate schema. @Jdforrester-WMF can correct us if that's wrong.