Transcript blp - Rasmusen

G604, BLP Lectures
Spring 2006, 2 March 2006
Eric Rasmusen, [email protected]
1
GMM in canned programs
Just like instrumental variables.
You say, in SAS or STATA or whatever, that you are
estimating
REGRESS: Y X1 X2 X3
and then you say
INSTRUMENT: X2 X3 Z1 Z2 Z3
This uses Z2, Z2, Z3 to instrument for X1 and correct for
heteroskedasticity. The package calculates the
2
variance-covariance weighting matrix for you.
From Wooldridge. Log(wage) is the dependent variable (hourly wage).
The instrument, neare4, is a man’s distance from a 4-year
college when he was 16. An extra year of education is worth 3.7% more for a black
man than a white man.
The difference between 2SLS and GMM is the weighting matrix.
3
GMM Benefit
• 1. You can add all kinds of crazy instruments to
improve your estimates
• 2. But the weighting matrix means that the
additional effect of bad instruments is slight
• 3. And also that if they are correlated with the
other instruments the effect is slight
• (you get heteroskedasticity correction too, but you
can get that in other ways)
4
Dangers of GMM
• The crazy instruments can hurt you in small
samples (e.g., real samples), because of
accidental correlations
• You can do data-mining searches for crazy
instruments
5
Estimating causes of lawyer income
Taxes-paid = Experience, Talent,
GDP/lawyer, Lawsuits/lawyer
The first three variables are for a given lawyer, the
last two are for the prefecture in which he lives.
What is wrong with this regression?
6
Estimating causes of lawyer income
Taxes-paid = Experience, Talent,
GDP/capita, Lawsuits/capita
GDP/capita comes in *negative*. Lawyers in rich
prefectures have lower incomes! Why?
(IV with variables such as hospitals/capita and
cars/capita as instrument)
7
Estimating causes of lawyer income
Taxes-paid = Experience, Talent, Lawyers in the
prefecture, Lawsuits/capita (demand)
Taxes-paid = Number of movie theaters/capita
(IV with variables such as hospitals/capita and
cars/capita as instrument)
8