MATLAB Central Newsreader - glmfit iteration limit

glmfit iteration limit

Ketan — Tue, 22 Jul 2008 19:35:04 -0400

Hello

I am getting a warning "iteration limit reached" while
using GLMFIT function.

How can I increase this iteration limit? I looked into the
code and figured it has hard coded iteration limit set to
100 iterations. Can I modify the code and increase this
iteration limit? Is there a reason (with statistical
significance) for restricting iteration limit to 100?

Regards,
Ketan

Re: glmfit iteration limit

Tom Lane — Wed, 23 Jul 2008 13:07:45 -0400

> I am getting a warning "iteration limit reached" while
> using GLMFIT function.
>
> How can I increase this iteration limit? I looked into the
> code and figured it has hard coded iteration limit set to
> 100 iterations. Can I modify the code and increase this
> iteration limit? Is there a reason (with statistical
> significance) for restricting iteration limit to 100?

Ketan, there's no statistical reason for restricting the limit to 100. Feel
free to change it in your file. I'll record an enhancement request that
this be made adjustable in a future release.

One cause for excessive iterations can be complete separation in regression
with a binomial response. If the response is true/false, and if there's a
point p such that x<p is all true and x>p is all false, then the maximum
likelihood is infinite. (The same could be true if a linear combination of
multiple x values has such a point.) You could check for that condition.

-- Tom

Re: glmfit iteration limit

Peter Perkins — Wed, 23 Jul 2008 13:42:25 -0400

Tom Lane wrote:

> One cause for excessive iterations can be complete separation in regression
> with a binomial response.

Just to follow up on what Tom said, 100 iterations is a _lot_ for the IRLS
algorithm that GLMFIT uses. I'm not saying that there aren't cases where you
usefully need more than that, but I suggest first looking that the fit that
you're getting after 100 iterations and seeing if there's anything you should be
paying more atention to in your data. Tom cites the most common example.