Hi Rodrigo,

I have touched the common denominator of your paper and with respect to your last statement on the possibility of "wrong results", I need three clarifications: (1) while your paper is premised on the hypothesis of many instruments(significant overidentification), the instruments in my paper are very few(just about six) in tandem with my research question which is based on 'assessing the validity of certain channels to the endogenous variable, contingent on the instrumental variables'; (2) the Hausman(prior to the 2SLS adoption) and Sargan-OIR(after the 2SLS) tests have null hypotheses that are rejected and unrejected respectively(testament to good model specification);
(3)the example you ask me to follow( rho=0.3 and K=5) has fewer instruments than the number employed in my paper, and this confuses me a bit. Did I get something wrong?

Thanks

Interesting result.

In the standard OLS framework moving from normal SE to robust ones usually implies higher SE's and therefore reducing t-stats, which also 'reduces promising-results'. However, in 2SLS framework one of the main problem is the bias of the estimator, which of course affects the estimate of the robust SE's, leading to a dirty t-stat. So, let's move to the bias issue.

I did explore it years ago (http://www.bcentral.cl/eng/studies/working-papers/pdf/dtbc464.pdf). If you check the Monte Carlo section in the paper, you will see in Tables 2 to 5 that 2SLS-bias increases with both endogeneity (rho) and the number of instruments (K), this is of course a standard results. But moving to Tables 6 to 9, I did robust inference which is the main point of your question. So, let's assume that your empirical problem implies a low 2SLS-bias as in the case rho=0.3 and K=5, you could see that for all designs considered there is not a significance distorsion in the rejection frequency meaning that for 'almost-unbiased' estimators robust SE's shouldn't affect your inference. It could be the case that your 2SLS estimator is biased, meaning the 'promising-results' could be
'wrong-results'.

Maybe you could add to the discussion bootstrapping SE and reporting coeff (non-robust SE) [robust SE or bootstrap SE], adding in text the standard 'on the one hand... on the other hand ".

Best, Rodrigo.

2012/6/1 Allin Cottrell <cottrell@wfu.edu>

On Fri, 1 Jun 2012, Anutechia Asongu wrote:If in tests for heteroskedasticity and autocorrelation the null is not rejected, one could argue that HAC estimation is just introducing noise. On the other hand, if the HAC standard errors differ substantially from the "classical" ones, that suggests that the errors are not i.i.d.My 2SLS results without HAC standard errors are promising whereas those with HAC standard errors are not. Hitherto, I have been used to robust HAC standard errors in my estimations. I'll like to enquire if a sound case could be made for presenting results without HAC standard errors.

Allin Cottrell

_______________________________________________

Gretl-users mailing list

Gretl-users@lists.wfu.edu

http://lists.wfu.edu/mailman/listinfo/gretl-users

_______________________________________________

Gretl-users mailing list

Gretl-users@lists.wfu.edu

http://lists.wfu.edu/mailman/listinfo/gretl-users