On Tue, 20 Apr 2010, Summers, Peter wrote: I'm going to defer on that one for now, but address the next point. 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

Allin, Many thanks for your prompt replies. Your point about the ill-conditioned instrument matrix makes perfect sense (I'd wondered why the info criteria were showing "1.#INF00"). I've tested the latest cvs with the following model X = {const black ed76 smsa76 south76 exp76} Z = {const black age76 smsa66 south66 nearc4b} The weak instrument test is still sensitive to the order of the instruments -- when it is there, only the minimum eigenvalue is reported. That would seem to work as a "flag" that something's wrong. If I change the instrument list to Z = {const black age76 smsa76 south76 momed daded}, all three tests (Hausman, Sargan and weak instruments) are reported, although the latter starts off with Weak instrument test - Cragg-Donald minimum eigenvalue = 195.731 Critical values for TSLS bias relative to OLS: bias 5% 10% 20% 30% value 0.00 0.00 0.00 0.00 (I'm not familiar with this form of the test, so maybe those 0's are what they're supposed to be). I'm still not sure about the df in the Hausman test though. In the last model it seems to me that this should be 2, not 1. There's no issue of collinearity that I can see (though rcond(Z'Z) = 5.6e-006), and we're asking about the exogeneity of 2 variables. I don't want to belabor this -- I know there are lots of ways of computing the test, but so far I haven't found one that uses (# endog variables)-1 (though I see how that could/should happen with collinear variables). I'm willing to be enlightened or persuaded though... PS -----Original Message----- From: gretl-users-bounces(a)lists.wfu.edu [mailto:gretl-users-bounces@lists.wfu.edu] On Behalf Of Allin Cottrell Sent: Tuesday, April 20, 2010 11:10 PM To: Gretl list Subject: Re: [Gretl-users] tsls tests On Tue, 20 Apr 2010, Allin Cottrell wrote: Ready to answer this one: It depends. The degrees of freedom equals the difference between the number of regressors in the unrestricted and restricted regressions, which may or may not equal the 'k' you're talking about. In the model you describe one "hat variable" is dropped from the the unrestricted regression due to exact collinearity. As for the presence/absence of the weak instruments test depending on the ordering of the variables, clearly that shouldn't happen. I've committed a change to CVS which I think ensures that you'll never see broken "weak instruments" output if the matrix of instruments is too close to rank deficiency. We should probably print a warning in such a case but I haven't put that into CVS yet. Allin Cottrell _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users

On Wed, 21 Apr 2010, Allin Cottrell wrote: You still have perfect collinearity in the relevant respect. The issue is that in the Verbeek schooling dataset the experience values are "fake" (as is quite often the case in labor-market datasets); they are calculated numbers representing "potential experience", as in exp76 = age76 - ed76 - 6 This means that when you run the unrestricted regression for the Hausman test one of the "hat"-variables gets dropped automatically, leaving just one degree of freedom for the test. Allin

On Wed, 21 Apr 2010, Summers, Peter wrote: Are you seeing those with the current snapshot from CVS? Allin

On Wed, 21 Apr 2010, Summers, Peter wrote: OK, that's right enough. The critical values are in fact zero in that region according to the Stock-Yogo tables. Allin