Tl;dr: the second half of the post conflates ability bias with signalling.
You’d think you could also do this simply by paying people purely on the basis of their educations (and perhaps seniority), like a government or a union shop where ability isn’t relevant. The problem is you’d still need to fix that more able people more often attend, and conditional on going more often finish.
That is the obvious explanation, and is completely consistent with everyday experience. Sure, I could waltz through a PhD no problem, but I’m not getting paid nearly as much today as I would with a PhD.
And yes, I do get paid more than my former classmates who would not be able to handle a PhD. So that speaks to nonzero ability bias (n=1). On the other hand, the difference in pay between “has an undergrad degree and could handle grad” vs “has an undergrad degree and could not handle grad” is presumably way smaller than the difference between “has an undergrad degree and could handle grad” vs “has a grad degree”.
If you want to hypothesize a world where ability bias is actually zero, then yes, you’d have to turn to the weird scenarios in this post. But you don’t need any of that to hypothesize a world where ability bias looks like zero—is statistically indistinguishable from zero. For that, you just need a world where ability bias is rounding error on top of the main effect, i.e. signalling, and all the statistical effect from ability bias gets masked by that larger effect.
In particular, if you measure ability bias by looking at earnings after controlling for education level, then you do not need to “fix that more able people more often attend, and conditional on going more often finish” in order to find statistically zero ability bias effect. The effect from “more able people more often attend” etc would be signalling, not ability bias—as the earlier part of the post defines them. Those effects go away when we control for education; that’s the whole point of controlling for education. If the large effect is signalling, then of course we’re not going to find a large effect when we look for ability bias separate from signalling!
Now, I certainly agree that Berkeley professors arguing for the effectiveness of education should be viewed with an awful lot of suspicion. But maybe rather than just dropping a-priori anvils, look at the data? It’s entirely plausible that ability bias effects are statistically indistinguishable from zero, but this post doesn’t really provide much evidence toward that question one way or the other.
Tl;dr: the second half of the post conflates ability bias with signalling.
That is the obvious explanation, and is completely consistent with everyday experience. Sure, I could waltz through a PhD no problem, but I’m not getting paid nearly as much today as I would with a PhD.
And yes, I do get paid more than my former classmates who would not be able to handle a PhD. So that speaks to nonzero ability bias (n=1). On the other hand, the difference in pay between “has an undergrad degree and could handle grad” vs “has an undergrad degree and could not handle grad” is presumably way smaller than the difference between “has an undergrad degree and could handle grad” vs “has a grad degree”.
If you want to hypothesize a world where ability bias is actually zero, then yes, you’d have to turn to the weird scenarios in this post. But you don’t need any of that to hypothesize a world where ability bias looks like zero—is statistically indistinguishable from zero. For that, you just need a world where ability bias is rounding error on top of the main effect, i.e. signalling, and all the statistical effect from ability bias gets masked by that larger effect.
In particular, if you measure ability bias by looking at earnings after controlling for education level, then you do not need to “fix that more able people more often attend, and conditional on going more often finish” in order to find statistically zero ability bias effect. The effect from “more able people more often attend” etc would be signalling, not ability bias—as the earlier part of the post defines them. Those effects go away when we control for education; that’s the whole point of controlling for education. If the large effect is signalling, then of course we’re not going to find a large effect when we look for ability bias separate from signalling!
Now, I certainly agree that Berkeley professors arguing for the effectiveness of education should be viewed with an awful lot of suspicion. But maybe rather than just dropping a-priori anvils, look at the data? It’s entirely plausible that ability bias effects are statistically indistinguishable from zero, but this post doesn’t really provide much evidence toward that question one way or the other.