st: RE: ivreg2 and weak instruments

["Schaffer, Mark E" <M.E.Schaffer@hw.ac.uk>]

Alice,

> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu 
> [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Alice
> Sent: 03 March 2011 11:10
> To: statalist@hsphsun2.harvard.edu
> Subject: st: ivreg2 and weak instruments
> 
> Dear all,
> 
> In a paper that I wrote in 2008, I used the -ivreg2- command 
> to test for weak instruments. I was clustering the standard 
> errors of my IV regression, and so
> -ivreg2- gave me the Kleibergen-Paap statistic that is robust 
> to heteroskedasticity and that I used to test for weak 
> instruments. However, this statistic had to be compared to 
> the Stock and Yogo critical values that were derived under 
> homoskedasticity only.
> 
> I now need to get back to this paper, and I was wondering if, 
> since 2008, any progress had been made regarding the test on 
> weak instruments, and in particular whether new critical 
> values had been derived to be used with the Kleibergen-Paap statistic?

As far as I know, no.

As I understand it, the problem is a tricky one because the robustness
of the K-P statistic to violations of homoskedasticity has a
large-sample justification.  How it behaves on the way to infinity will
depend on the nature of the heteroskedasticity (or autocorrelation, or
whatever).  I think that's why one-size-fits-all critical values for the
K-P stat like those Stock & Yogo derived for the unrobust Cragg-Donald
statistic can't be (or can't be easily?) derived.

We added K-P to ivreg2 anyway because (a) we thought one-size-fits-all
critical values were better than nothing, (b) the K-P statistic for
underidentification (as opposed to weak identification) has a
straightforward asymptotic justification and is perfectly OK, (c) a big
difference between a robust K-P statistic and the unrobust C-D statistic
is a signal to users that using the latter might not be a good idea, and
(d) Jeff Wooldridge had recommended somewhere (sorry, I can't remember
the cite) using a robust first-stage F-stat if heteroskedasticity was
suspected, and the K-P statistic is a generalization of a robust
first-stage F stat to the case of more than one endogenous regress.

Cheers,
Mark

> Many thanks.
> 
> Best wishes, Alice.
> 
> 
> 
>       
> 
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
> 


-- 
Heriot-Watt University is a Scottish charity
registered under charity number SC000278.


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/