(J$W:) I am perfectly comfortable saying that race and ethnicity are categories that are used (but not useful) as learning tools. Different categories have different predictive power, and race and ethnicity just happen to be categories that are misleadingly uninformative. We expect them to be informative because outward appearances in skin color and body type give the impression that big differences exist, so we spend an irrational amount of energy searching for differences, but in fact race and ethnicity might not be any more informative than categories like shoe size or what you ate for breakfast. (This is not to say that race and ethnicity are completely uninformative categories, just that they are less informative than we tend to unconsciously assume.)
(Bob:) I'd love to know why race and ethnicity are not useful stereotypes. Are you saying this based on some empirical data, or your gut feeling? I truly don't see how they're different from gender in that way.
(Xan:) By the way your footnote 4 is diplomatic, though I don't buy it and I don't think you do either. Race goes into regressions, and the coefficient isn't usually 0. Race and ethnicity can be useful categories whenever we're asking a statistical question that varies by race or ethnicity. You probably just weren't thinking about that sort of question, but sometimes we're interested in things like expected income or the likelihood that someone off the street can give us directions to the metro in English. *Any* observable characteristic over which that quantity varies, becomes statistically useful.
First, I’ll defend my diplomacy (≈bullshit) by pointing out that I only said they’re not as useful as we unconsciously assume. And I’m standing by that. But you guys are daring me to say that they’re not useful at all.
There are certainly statistical correlations (and so predictive power) that go along with race categories (e.g., income, education, marital status). As Xan notes, the coefficients aren’t 0.
The important distinction is that race correlations unlike sex correlations are almost certainly not causal. With race categories, I could get the same or better correlations by using a completely different independent variable like neighborhood of residence. With sex, not so.
Here’s the important thing to understand about race: It has almost no biological meaning. The genetic differences are either so tiny that they’re unnoticeable or totally non-existent. I didn’t read this PBS article, but it probably has good evidence if you’re craving it.
With sexes, on the other hand, I dare you to read this and this and tell me that there are no meaningful psychological or neurological differences. For the longest time I denied that there were, but the weight of the evidence is overwhelming.
Here’s another, more eloquent way of making the point, courtesy of Nathan:
Calculus only gives an approximate answer, depending on your comfortability with the size of epsilon.
Sterotyping merely does the same. The claim is not that the mean size of epsilon is harmful, but that the variance size is.
In general, I think we avoid stereotyping when the variance is high, and feel safer stereotyping when it is low, regardless of the size of the mean.
So we ought to be less comfortable generalizing on race than sex if the variance is much higher, and I bet it is.
