Who's Counting: Testing and Hiring Disparities Need Not Imply Bias
Feb. 6, 2011— -- A claim of prejudicial treatment is the basis for many news stories. The percentage of African-American students at elite universities, the ratio of Hispanic representatives in legislatures, and, just recently, the proportion of women among Wikipedia contributors have all been written about extensively.
Oddly enough, the shape of normal bell-shaped (and other) statistical curves sometimes has unexpected consequences for such situations. This is because even a small divergence between the averages of different population groups is accentuated at the extreme ends of these curves, and these extremes, as a result, often receive a lot of attention.
There are policy inferences, most of them wrong-headed, that have been drawn from this fact, but I certainly don't want to delve into questionable claims. I merely want to clarify a couple of mathematical points.
To illustrate one such point, assume two population groups vary along some dimension - height, for example. Let's also assume that the two groups' heights vary in a normal or bell- shaped manner. Then even if the average height of one group is only slightly greater than the average height of the other, people from the taller group will make up a large majority of the very tall.
Likewise, people from the shorter group will make up a large majority of the very short. This is true even though the bulk of the people from both groups are of roughly average stature. So if group A has a mean height of 5'8" and group B a mean height of 5'7", then (depending on the variability of the heights) perhaps 90% or more of those over 6'3" will be from group A. In general, any differences between two groups will always be greatly accentuated at the extremes.
Let me illustrate with another somewhat idealized case. Many people submit their job applications to a large corporation. Some of these people are Mexican and some are Korean, and the corporation, perhaps unwisely, uses a single test to determine which jobs to offer to whom.
For whatever reasons, let's assume that although the scores of both groups are normally distributed with similar variability, those of the Mexican applicants are only slightly lower on average than those of the Korean applicants. (I chose the direction of the difference at random. The same point holds if the Mexicans' scores are slightly higher.)
