Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: frontier command and iteration problem


From   "David M. Drukker, StataCorp" <ddrukker@stata.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: frontier command and iteration problem
Date   Mon, 21 Jun 2004 12:55:42 -0500

Dev Vencappa <lexdvv@nottingham.ac.uk> wrote that while -frontier- declared
convergence for a given model and dataset combination when run under version
8.0, -frontier- does not declare convergence for that same model and dataset
combination when run under under version 8.2.

In the update of 05nov2003, the -nrtolerance(1e-5)- became the default
convergence criterion.  I suspect that Dev's model/dataset combination
satisfies the old, weaker criterion, but not the -nrtolerance(1e-5)
criterion.

Dev could use the -nonrtolerance- option in Stata 8.2 to verify that that
this is what is causing the difference.

Another possibility is that the default -nrtolerance(1e-5)- is too tight
while, say , -nrtolerance(1e-3)- is more appropriate for this model/dataset
combination.  If loosening the tolerance induces convergence, then Dev will
probably want to see if the stricter -nrtolerance(1e-5) can be applied after
rescaling the variables, as mentioned below.

While it is not possible to make any definite statements without looking at
the Dev's results, I would hazard a guess the some of the standard errors
are missing in the results that Stata 8.0 declared converged.  These missing
standard errors imply that the Hessian is not of full rank, which means that
the parameters of the model are not numerically identified by the dataset.
(See Drukker and Wiggins (2004) for more about this issue.)

I would recommend that Dev rely on the new, stricter criterion.  Of course,
this implies that Dev now has a model/dataset combination that does not
converge.  My experience with frontier models has taught me that a
multiplicative rescaling of the variables is one of the most successful
methods to induce convergence.  These models converge more often when all
the independent variables are on the same scale.  Using different starting
values can also make a difference.

I hope that this helps.

	--David
	  ddrukker@stata.com

References
----------

Drukker, D.M. and Wiggins, V. 2004. "Verifying the solution from a nonlinear
solver: A case study: Comment", American Economic Review, March 2004
(397-399).

*
*   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/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index