LogL estimation - "Failure to improve Likelihood"

[CharlesLogan]

Greetings:

I am estimating coefficients via the LogL object in Eviews. More often than not, I get the error message:

"Failure to improve Likelihood after X iterations
WARNING: Singular covariance - coefficients are not unique"

However, if I simply re-start the MLE estimation and not changing the coefficients, it subsequently converges.

Is there a way to avoid this error and have the MLE re-start in a similar fashion to my manual method?

Thanks!

[EViews Gareth]

Try tightening the convergence tolerance.

[CharlesLogan]

That did not work. I originally set the convergence threshold to 1e-5 and tightened it to several values leading down to 1e-50 and still get the singular covariance error.

[EViews Glenn]

It sounds as though there is a singularity at the root of the likelihood equation at which we are evaluating the standard errors. In likelihood estimation, this is akin to multicollinearity...

There is a possibility that you may get better results at a different root, but at least numerically, we are finding a singularity. You might want to try to experiment with different starting values.

[CharlesLogan]

When I get these singularity errors, I simply re-execute the "ml" method on my LogL object without changing any of the coefficients or parameters to the ml method.
Oftentimes, the MLE will subsequently converge. It takes a substantial amount of time to sit by the computer and keep re-executing the 'ml' command when the estimation fails. Is there a way to tell Eviews re-execute this 'ml' command upon an error, up to N times? This will be a tremendous time saver. For example, I would like to have Eviews re-estimate the 'ml' command on each LogL object up to 10 times before giving up.

[EViews Glenn]

If the error is a singularity so that you have NAs as standard errors, you could certainly put the estimation inside a loop where you test for NA values and reestimate if encountered. That would be pretty straightforward. You could even modify the coefficients slightly using a random number generator.