Re: st: How to avoid a break down when convergence is not achieved?

["Benjamin M. Craig, Ph.D." <craig@pharmacy.arizona.edu>]

I found -capture- to be useful. However, if the ML completes the maximum 
number of iterations without a break, _rc still equals 0. -capture- 
avoids the breakdown, but _rc is a poor indicator of non convergence.

Instead, ML returns e(converge), which equals 1 if convergence is 
achieved, otherwise 0. Therefore, if _rc+(e(converge)==0) is greater 
than zero, the estimation failed to render the intended results.

The combination of these two pieces of information has worked very well 
for my purposes.  Thanks for all the feedback.

Ben

Tim Wade wrote:

> Ben, check out the combination of -capture- and -noisily-, or the
> nostop option of the -do- command, these might help.
>
> On 1/31/06, Benjamin M. Craig, Ph.D. <craig@pharmacy.arizona.edu> wrote:
>
> > For those who use ML or any difficult optimum estimator, convergence can
> > break down for multiple reasons (e.g., poor initial conditions or
> > multicollinearity in a bootstrap sample). When it does, stata stops.
> >
> > Is there a way of convincing stata to make a note and keep going?
> >
> > In other words, can ML report an error in a manner which can be
> > identified without disabling the batch code?
> >
> > For example, suppose I created a do loop through a five sets of initial
> > conditions such that if one set of conditions does not converge, the
> > others might. Or, to see whether each set arrives at the same optimum.
> > Currently this is not possible, because when "convergence not achieved;"
> > r(430), the batch stops running. This is a real annoyance for lazy
> > people like me who would like to run code over nights, weekends, etc.,
> > instead of setting up appropriate initial conditions.
> >
> > Cheers,
> > Ben

*
*   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/