Re: [Gretl-users] Testing the endogeneity of one or more regressors in a pooled OLS model?
On Tue, 12 Nov 2013, Clive Nicholas wrote:
Sven,
Check out IV estimation with the 'tsls' command. May I make the
> unrelated suggestion to also test if the pooling is actually adequate.
Running the panel diagnostics informed me that
Residual variance: 1.35107e+06/(822 - 60) = 1773.06
Joint significance of differing group means:
F(46, 762) = -1.99828e-14 with p-value 1.79769e+308
What version of gretl is this? (That p-value means "NA", as it obviously
has to be if the F-statistic is numerically negative, but I thought we'd
purged all cases of printing NA as if it were a number.)
I can't properly run IV2SLS models in -gretl- because I only have
one IV
and the main model has 96 predictors, so it responds that I don't have
enough instruments.
That sounds like a non-sequitur: for one endogenous variable you need one
additional instrument, regardless of how many exogenous regressors are
included in the model. Have you looked up the syntax of gretl's "tsls"?
It's
tsls <dependent-var> <regressors> ; <instruments>
where you repeat any exogenous regressors as (their own) instruments in
the second list, behind the semicolon. For example
tsls y const x1 x2 n ; const x1 x2 z
where 'n' represents an endogenous regressor and 'z' the associated
instrument.
So, what to do in order to test the endogeneity of my regressor in a
'standard' model?
Well, what sort of test do you have in mind, other than running an IV
model and comparing the results with OLS via a Hausman test? No
residual-based test is available since the residuals are by construction
orthogonal to all the regressors. (You get a Hausman test automatically
when you use the "tsls" command in gretl.)
Allin Cottrell