st: RE: RE: Error in ivreg2 and ivregress

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

Benjamin,

> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu 
> [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of B Villena
> Sent: 23 July 2011 06:14
> To: Statalist
> Subject: st: RE: Error in ivreg2 and ivregress
> 
> 
> Dear Mark,
> 
> Thank you for your prompt answer
> I updated my versions of ivreg2 and ranktest as you 
> suggested. However, I get a different error code
> invsym(): matrix has missing values
> r(504);
> r(504);
> 
> Regarding repeated instruments, I only notice that ivreg2 
> detects this situation, so I do not believe it is the reason 
> of the error.

Maybe, but you shouldn't do it anyway.  It makes no sense.

> I suspect this problem arises due to high collinearity among 
> instruments (R-squared for regressions of one instrument 
> against the other ones can be as high as 0.9 in my data). 
> Since this R-squared is quite high, I have dropped one of the 
> instruments to make ivreg2 work in this case. Given the high 
> collinearity among instruments, I guess this approach is 
> losing little information. 

Impossible to say without seeing more details.  For example, you could
have very badly scaled data.  If the mean of one variable is 000000s of
times larger than the other, you can have big numerical problems with
matrix inversion.

> Nevertheless, I still wonder whether there may be some other 
> bug in my code or in ivreg2 - ivregress code.

The bug is very unlikely to be in the estimators - the two programs are
independently written.  The problem is in your specification/data.

> The 
> collinearity seems high, but not that high!
> 
> In addition, I follow your suggestion of using ivreg2 to 
> compute a 2SLS estimator in this case. I do obtain a 
> meaningful estimation with 2SLS.

You should probably be able to tell from the first and second stage
output where the problem is.  That was the reason I suggested you try
2SLS.

> However, I strongly prefer 
> LIML in this case since the instruments seem somewhat weak.

You also should try partialling out the exogenous regressors that aren't
of interest (if there are any).

Yours,
Mark
 
> 
> Thanks a lot
> 
> B
> 
> 
> 
> > Subject: st: RE: Error in ivreg2 and ivregress
> > Date: Fri, 22 Jul 2011 23:11:56 +0100
> > From: M.E.Schaffer@hw.ac.uk
> > To: statalist@hsphsun2.harvard.edu
> > 
> > Benjamin,
> > 
> > A few suggestions/thoughts:
> > 
> > 1. Tell us which versions of -ivreg2- and -ranktest- you have 
> > installed. (Use the -which- command.) You may not have the most 
> > up-to-date versions (they should be 3.0.06 and 1.2.03, 
> respectively).
> > 
> > 2. When you say
> > 
> > > In both cases there is a repeated instrument on the exogenous 
> > > instruments list, i.e. z2=z3 in the above example
> > 
> > I don't understand what you mean. If z2 and z3 are 
> identical, then it 
> > makes no sense to include them both since they are 
> perfectly collinear.
> > But -ivreg2- (and, I think -ivregress-) should catch this 
> and drop one 
> > of them.
> > 
> > 3. LIML and 2SLS are very closely related. If this is a data-driven 
> > problem, it will probably be apparent in the 2SLS output. Also, the 
> > collinearity in (2) should be reported in the -ivreg2- 2SLS output. 
> > You can share this output with us if you want.
> > 
> > 4. If it's data related, you might be able to address it by 
> > partialling-out some of the regressors using the -partial- option.
> > 
> > HTH,
> > Mark
> > 
> > > -----Original Message-----
> > > From: owner-statalist@hsphsun2.harvard.edu
> > > [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf 
> Of B Villena
> > > Sent: 22 July 2011 22:42
> > > To: Statalist
> > > Subject: st: Error in ivreg2 and ivregress
> > > 
> > > 
> > > Hello,
> > > 
> > > I'm trying to estimate a series of linear IV model by 
> LIML using the 
> > > user-written command ivreg2.
> > > 
> > > 
> > > 
> > > > ivreg2 y x1 x2 x3 (y1 y2 = z1 z2 z3 z4) if sample==1, 
> liml first 
> > > > savefirst
> > > 
> > > I obtain the following error message
> > > 
> > > 
> > > 
> > > (** On entry to DSTEIN parameter 6 had an illegal value)
> > > _symeigen_la(): 3930 error in LAPACK routine
> > > _symeigen_work(): - function returned error
> > > _symeigensystem(): - function returned error
> > > symeigensystem(): - function returned error
> > > rkstat(): - function returned error
> > > <istmt>: - function returned error
> > > r(3930);
> > > 
> > > 
> > > As far as I know, it seems that LAPACK cannot compute eigenvalues 
> > > needed to obtain LIML estimator
> > > 
> > > 
> > > 
> > > I have tried using Stata 11 command ivregress. 
> > > 
> > > > ivregress liml y x1 x2 x3 (y1 y2 = z1 z2 z3 z4) if sample==1, 
> > > > first
> > > 
> > > 
> > > 
> > > It does not work, either
> > > 
> > > 
> > > 
> > > I get the following error message
> > > 
> > > 
> > > 
> > > estimates post: matrix has missing values r(504); r(504);
> > > 
> > > 
> > > 
> > > In both cases there is a repeated instrument on the exogenous 
> > > instruments list, i.e. z2=z3 in the above example
> > > 
> > > However, this does not seem to be the problem since I 
> have made some 
> > > similar estimations with some repeated
> > > 
> > > instruments and both commands seem to deal with this 
> issue very well 
> > > in other cases.
> > > 
> > > 
> > > 
> > > Has anyone experienced a similar problem? Is there any way to 
> > > circumvent this problem?
> > > 
> > > Is it possible that this problem is related to an 
> unfortunate data 
> > > configuration?
> > > 
> > > 
> > > 
> > > Best regards,
> > > 
> > > 
> > > 
> > > Benjamin
> > > *
> > > * 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/ 		 	
>    		  
> *
> *   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/