estimates based on supply-side innovations tend to recover returns to education for a subset of individuals with relatively high returns to education.
Let me try to put that into English.
The conventional wisdom is that when somebody looks at the relationship between years of schooling and earnings, they over-estimate the impact of schooling because of ability bias. That is, people who have more schooling have more ability, and so you might naively attribute all of their higher earnings to more education when in fact much of it may be due to ability. A more sophisticate approach would attempt to correct for ability bias.
Card's paper, referred to in the Goldin-Katz book, summarizes some attempts to address this. But it uses Fancy Econometrics, which makes the results suspicious, and my interpretation of them even more so.
The way I think of it, suppose we have two types of people. A-plus Adrians have lots of ability. Pass-fail Pats have mediocre ability. If we can find a control group of people, where presumably mostly Adrians go to college, and compare it with an experimental group where we have reason to believe that the college attenders include a higher proportion of Pats, then we can get an idea of how much ability bias there is. The experimental group would have a lower college wage premium because of all of the Pats who managed to sneak into college.
In fact, what studies like this show, according to Card, is that the experimental group tends to show a higher earnings/education gradient than the control group. It appears, therefore, that what we have is reverse ability bias! Perhaps people with more ability get more consumption value out of school and actually derive less investment value at the margin. That is, Adrian uses an extra year of schooling to go for a Masters' in art history, while Pat takes courses in writing and computer science at the local community college.
That's the story, assuming I get the gist of the paper, and assuming you believe the results. Are the experiments convincing?
In Card's own study, you can think of the experimental group as consisting of folks who grew up near a four-year college. If that makes it more convenient to attend college, then you should see more Pats among that group attending college. If the college/high school wage premium is higher within that group, this tells you that at the margin Pat gets more out of college than Adrian.
But as I said, Card uses Fancy Econometrics. That means that, as far as I can tell, the higher returns to education in the experimental group could be coming from higher returns to more years of high school. They do not necessarily have anything to do with college at all. I really wish folks would use more descriptive statistics and fewer regressions.