[firethief]

I'd rather not install an app to read that, but presumably they are measuring who is better educated in math. I would expect the correlation to be mostly explained by different ethnic groups in the location the study was conducted having different access to education, cultural values, and expectations placed on them.

It would go a long way toward distinguishing between those two possibilities if a computer judged the faces - find the eigenface(s) of different mathematical ability levels and see if there's a lot of correlation between those and ethnicity, or if features alone are the predictor.

Even if it is features alone though, I would expect a major involvement of education and self-fulfilling expectations: take a starting state where there is no correlation between facial features and math aptitude. Everyone would have their own beliefs about what faces are good at math (humans always see patterns). Some of those beliefs would happen to be similar, so people with certain features would be steered more toward math. As time passes, the bias becomes more legitimate and more self-perpetuating. The end result is a socially-imposed link between particular visible traits and ability in math.

[yummyfajitas]

You don't need to install an app, if you do the right pattern of clicks you can download it. But here is an ungated preprint (wasn't available when I first downloaded the paper): http://www.ereuben.net/research/StereotypesWomensCareer.pdf

They could indeed be observing that "asians are good at math" stereotypes are a useful predictor, the study doesn't really give enough info to determine that one way or the other.