Re: [Gretl-users] tsls tests
On Tue, 20 Apr 2010, Summers, Peter wrote:
I've got a couple questions regarding tsls and the automatic
specification tests gretl does.
First, in a model with k (potentially) endogenous regressors,
gretl seems to use k-1 degrees of freedom for the Hausman test
rather than k. My impression is that gretl uses the "variable
addition" form (eg, Greene p. 323), is that right? If so,
shouldn't the the df be k?
I'm going to defer on that one for now, but address the next
point.
The specific model I'm using is from Verbeek's
"schooling" data
set. Log wage is the dependent variable, with "black", "south76"
and "smsa76" as exogenous regressors; education, experience and
its square ("ed76", "exp76", "exp762") are instrumented
regressors. The instruments are age, its square, and a "near
college" dummy (plus the 3 exogenous x's). If I "align" the
regressor and instrument lists like this:
Model 21: TSLS, using observations 1-3010
Dependent variable: lwage76
Instrumented: ed76 exp76 exp762
Instruments: const black smsa76 south76 nearc4a age76 sq_age76
then I don't get the weak instruments test. But if I switch the
order to this:
Instruments: const black smsa76 south76 age76 sq_age76 nearc4a
I get the test, but with all 0's for the critical values:
The trouble is that gretl is not failing gracefully on an
impossible task. The correlation between age76 and sq_age76 is
0.9988, and (probably more relevant) the reciprocal condition
number of Z'Z (where Z is your instrument matrix) is 1.52e-10.
Stata 10 responds to "estat firststage" with a bunch of "not
available" messages. We should find a better criterion (or a
criterion at all) for supressing the first-stage test results
rather than printing garbage.
Allin Cottrell