My math is pretty solid - socio-economic classes are defined by values in intervals. So guessing the average income for each race would be guessing what interval? It would be guessing middle class for each race! Remember, you stated the proposition that, 'race is a good class indicator' - thus, we don't care about variance in the interval called middle class...
That is why I explicitly removed it from my analysis - if the only information(class-related) you get from a person's skin color is that they are middle-class, then you have effectively received no information. Certainly you can't be suggesting that those in the middle-class are more likely to commit crimes - can you be?