Re:Google Fires Employee Behind Controversial Diversity Memo
by waterhouse on Tuesday August 08, 2017 @03:21AM
(#14955289)
Also, I'd mention that, even if, say, you have a noisy (presumably cheap) test that you give to everyone, and then you give an absolutely perfect (presumably expensive) test to minorities that fail the first test... It's probably the case that, of the people who failed the first test due to noise but would pass the second test, most of them are just barely above the "passing" bar. Because it seems reasonable to guess that people way above the bar have a smaller chance of failing the noisy test than people just barely above the bar. The policy would then lower the average ability level for hired minorities (though it wouldn't bring it below the "bar").
How large this effect is depends on just how noisy the first test is and on how widely above the bar people's abilities range.
I don't think that works. If the process is optimised for accuracy, then decreasing the false negative rate should increase the the false positive rate, right? Intuitively that's how it would work for most classification tasks. If you can decrease the false negative rate without increasing the false positive rate then you're improving accuracy. The bar must be correlated with the false negative rate.