Our graduate student David Albouy brought his critique of the mortality data underlying Acemoglu, Johnson, and Robinson's "Colonial Origins of Comparative Development" to our Economic History Seminar on Monday. I tried to channel Daron Acemoglu as hard as I could--yet I did not manage to score many points off of David.
I also clearly need to learn more about the "weak instruments" problem...
Posted by DeLong at September 28, 2004 11:22 AM | TrackBackEcon 211, Fall 2004: September 27: David Albouy, UC Berkeley, "The Colonial Origins of Comparative Development: A Reexamination Based on Improved Settler Mortality Data" (Joint with Macroeconomics Seminar).
"I also clearly need to learn more "
This is a professor !
Footnote eight ("The ratios I found in Curtin were slightly different than the ones I found in Curtin")
certainly make it sound like something interesting is going on ...
Weak instruments problem, my pocket summary of ...
Say you have a recursive equations model:
x = A + By + error1
y = C + Dx + error2
This can't be estimated as is, obviously (the errors aren't independent).
So, you take z (=H + iy) as an instrument for Y and write the reduced form:
x = J + Kz
y = M + DKz + error2 +Derror1
(M= C + DJ)
Now start estimating the reduced form for Y by maximum likelihood. You have a perfectly normal likelihood surface more or less everywhere, except ....
At the point where D=0, the likelihood for y no longer depends on K. Therefore, the likelihod function is no longer identified for K, and there is a "ridge" in the likelihood surface.
If z is a good instrument for y, then this doesn't matter all that much; the likelihod surface is dominated by the peak associated with the ML estimate. If, on the other hand, z is poorly correlated with y, then the ridge takes up quite a lot of the density of the likelihood function and you can get bogus estimates.