In an article on mises.org Bob Murphy writes about the "signaling" model of education:
In this approach, the reason an employer will offer a higher salary to, say, a Harvard physics major with a 4.0 GPA than he would to someone with a 2.5 GPA and a literature degree from a community college isn't that the job will require knowledge of Maxwell's equations for electromagnetism. Rather, Spence's signaling model says that the type of person who chooses to major in physics at Harvard, and can graduate with a 4.0, is also the type of person who will do well in many jobs. So it's not that college (and graduate school) actually endow students with useful skills, it's just that they provide mental hoops for students to jump through and thereby exhibit their intrinsic abilities to employers.
To the extent that the Spence model is accurate, getting formal degrees is an arms race of sorts.
If everyone could agree to spend fewer years in school, then everybody would be better off. However, no individual student can benefit from unilaterally skipping college, because our society is currently stuck in the equilibrium in which productive individuals signal their talent by getting degrees from prestigious colleges.
Let me help Murphy insofar as I can. The problem with the signaling theory appears to me to be this. Let Smith the physics guy be productive at physics with 100 points of efficiency. And let Jones the literature guy be productive at literature with merely 60 points of efficiency.
It seems to me that their skills are not fully transferable to, say, accounting. We have to discount their efficiency for that occupation somehow. Let's suppose that we discount equally by 50%. Of course, this is too generous a favor to the theory, e.g., because accounting may be easy to master for both Smith and Jones, such that even Jones can become quite proficient with it which he could never do with physics. Anyway, Smith is 50 points as efficient at accounting, and Jones is 30 points as efficient at accounting.
(We may interpret these values as competence: 100 points out of possible 150, write it as 100/150; the others might be 60/90, 50/50, and 30/50. If we want to accommodate the fact that accounting is "easy," then change the last two sets of numbers to, say, 40/40 and 30/40.)
Clearly, on this understanding the signaling model breaks down. For employers apparently look at nothing but education. But if they do so, then, though Smith is only 20 points more efficient that Jones, he will be paid the equivalent of 40 points more. That is, Jones is only 20 points worse than Smith, but he is 40 points cheaper than Smith and may therefore be a better candidate for the accounting position. The profit from employing this factor of production, Smith, is too low (and may even be a loss) compared to the alternative, Jones. In simple terms, Smith is "overeducated."
So, if the employer does not look and discriminate beyond the education history, then he will miss opportunities to hire people like Jones and in so doing will lose in the competition to those entrepreneurs who do look.
P.S. Another way to approach this problem, of course, is to look at unpredictable creative entrepreneurship of people to market themselves without a college degree and succeed anyway.