Bryan Caplan wrote:
> I'll agree there are problems here, but I still think some simple
> regressions bound the answer fairly well.
How do you figure that? Strip away everything from the problem and
suppose that we live in a world where:
wages = bIQ + u1
afqt = cIQ + u2
educ = dIQ + u3.
By assumption, all that's driving all three outcomes is sheer brain
power, but brain power is measured with noise, and that noise and
unobervables affecting educational choices and wages may all be
correlated. A regression of wages on afqt and education could
easily come up with large values for the coefficient on the latter,
and very misleading estimates for the return to afqt. Or, of
course, some other structural model might imply that regressions
such as were provided underestimate the true returns to educations,
and would generally, therefore mis-estimate the returns to cognitive
ability. I don't see any sense in which the regression supplied
bounds the true values. For formal estimates of somewhat related
models, see for instance Cawley, Heckman, and Vytlacil's work, which
specifically demonstrates that estimates such as you supply are very
misleading.
Again, the empirical literature on this question is vast -- it's a
tough problem, and just running OLS with an coarse measure of
cognitive abilility doesn't bound anything. This puzzle is so
important, and tough to "solve," that is has been and remains a
testbed for the latest and greatest econometric techniques, as well
as the target of numerous attempts to find truly exogenous variation
in various places to help identify the causal mechanisms.
Chris Auld (403)220-4098
Economics, University of Calgary <mailto:[EMAIL PROTECTED]>
Calgary, Alberta, Canada <URL:http://jerry.ss.ucalgary.ca/>
