Re: st: Flagging estimation errors in -ivreg2-

["Michael S. Hanson" <mshanson@mac.com>]

On Feb 11, 2007, at 3:16 PM, Maarten buis wrote:

> > 	1. Some regressors are dropped [due] to perfect collinearity.  (This
> > colinearity is not trivial, such as two identical regressors or a set
> > of dummies reproducing the constant;  hence such regressions are very
> > difficult to identify ex ante.)
> -_rmcoll- may be a partial answer.
	Thanks for the suggestion, Maarten.  However, -_rmcoll- drops the  
collinear variables (according to -help _rmcoll-), which -ivreg2-  
already does by itself.  What I actually need is to be notified when  
this happens.  -quietly capture ivreg2- doesn't provide any  
indication that collinearity (or any other problem, short of an error  
that stops estimation) has occurred.


> > 	2. The VCV matrix fails a rank condition;  the error message  
> > displayed on the screen when estimated -noisily- is below.
> >
> > "Error: estimated covariance matrix of moment conditions not of  
> > full rank;
> >        overidentification statistic not reported, and standard  
> > errors and
> >        model tests should be interpreted with caution.
> > "Possible causes:
> >        covariance matrix of moment conditions not positive definite
> >        covariance matrix uses too many lags
> >        singleton dummy variable (dummy with one 1 and N-1 0s or  
> > vice versa)"
	After some poking around and some trial-and-error, I have discovered  
that -ivreg2- does (indirectly) provide information in e(scalars) and  
e(macros) that I may be able to exploit.  For example, it appears  
that I can flag the less-than-full rank of e(V) with the following  
code after an -ivreg2- call:

		if e(rankV) < e(rankxx) {
			noi disp "*** Error: deficient rank of VCV matrix."
		}

Also, variables dropped in -ivreg2- due to collinearity are listed in  
e(collin);  otherwise e(collin) is not created.  So I am able to flag  
the existence of e(collin) with the following code:

		if ~missing(e(collin)) {
			noi disp "*** Error: perfect collinearity."
		}

These two code blocks seem to work OK;  they flag the problems in the  
log file rather than the tables, but I can live with that.  (It would  
appear to require some significant re-writing of the table code --  
based on repeated -matlist- calls -- to do otherwise.)

	Alternative approaches are welcome;  I'm sure I'll learn some more  
coding tricks from them.  Thanks!

                                        -- Mike


*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/