It actually seems pretty complicated here. IIRC the best previous studies just used years in education as a proxy -- not a great one, but an easy number to get,

It actually seems pretty complicated here. IIRC the best previous studies just used years in education as a proxy -- not a great one, but an easy number to get, thus allowing large a sample size. This one seems to do some weird combination of that, plus several smaller datasets which have some kind of test score, all combined & normalized somehow. (Presumably they had some fun conversations with their referees about the statistics...)