Bryan has responded to my post on the sheepskin effect in typical Caplanian style: even if my critic is right, it just means I’m more right. In this case he says that my argument that ability bias explains some of the sheepskin effect is just evidence against the human capital model of education, because ability bias is a competing theory. I don’t have a dog in the bigger fight on what explains the overall education wage premium. It looks to me like the general consensus is behind the human capital theory, but I’m open to being convinced that signaling is an underexplored and underappreciated explanation. But I don’t think the story I’m telling about the sheepskin effect provide evidence against the human capital theory.
I would argue that Bryan is conflating two kinds of ability bias: the ability that endogenously determines the treatment of attending college, and the ability that endogenously determines the treatment of graduation conditional upon attending. It may be that randomly assigning people the treatment of attending college will lead to the exact same average estimated returns to education. This would not be inconsistent with claim that randomly assigning people to drop out of college would greatly reduce the measured sheepskin effect.
So why might ability bias be a bigger deal for graduating than for attending? Because even the dropouts themselves thought that they looked like someone who would complete college until at some point they no longer did. Attending and then dropping out of college, I would argue, most often results from some unpredicted negative shock. This becomes more true the longer someone attends college, because if you’ve spent 3.5 years in college thinking that this represented a good investment, you need a pretty big negative shock to make that no longer the case after the majority of the costs are sunk. This is going to be true as long as there is some signaling sheepskin effect.
I’m not positive this is correct, but I’m thinking of this ability bias as a downward bias on the part of the spline regression measuring the linear effect of years of education for years 12-16 rather than an upward bias on the dummy variable representing 16+ years of education.