Sheepskin Effects in the General Social Survey
By Bryan Caplan
A sizable literature on the education premium finds solid evidence of a “sheepskin effect“: diploma years pay noticeably more than other years. When I noticed that the General Social Survey has a DEGREE variable, I saw a golden opportunity to double check the robustness of the result.
As a benchmark, I started by just regressing log earnings (REALRINC) on years of education. The result:
Next, I added DEGREE to the explanatory variables. Degree=0 if you don’t have a high school diploma, =1 if you have only a high school diploma, =2 if you have an associate degree, =3 if you have a bachelor’s degree, and =4 if you have a graduate degree. Result:
Once you control for degrees, every year of education raises earnings not by 11.3%, but by a mere 2.4%. Every degree, in contrast, boosts earnings by a full 25.6%.
To be sure, this set-up implicitly assumes that every step up the degree ladder is equally valuable. What if you relax that assumption? I defined HIGHSCHOOL=1 if you have a high school degree or more, JUNIORCOLLEGE=1 if you have an associate’s degree, BACHELORS=1 if you have a four-year degree, and ADVANCED=1 if you have a graduate degree. Here’s what you get:
In this more flexible set-up, a year of education is only worth 2.2% in extra earnings. But high school graduation gets you an extra 28.1%. Junior college gets you another 23% on top of that. A four-year degree gets you a 45.7% over high school. And an advanced degree gets you another 38.6% on top of a four-year degree.*
Qualitatively, this is just what you’d expect. Quantitatively, however, this result is far more extreme than any other sheepskin effect I can recall. The GSS is a huge, high-quality data set, so you can’t dismiss it as an aberration. The income data is admittedly subpar, because it’s derived from categorical data. But why would that sharply inflate the benefit of degrees relative to mere years of study?
If you’ve got ideas, please share – and show your work.
* For coefficients this large, there’s a modest approximation error – log points and percentage gains start to diverge. I ignore this fact for simplicity.