my mistake. since nlminb is minimizing, it should be +Inf ( so that the
likelihood
is large ) as you pointed out. Note that this approach is a heuristic and
may not work all the time.
On Mon, Oct 21, 2013 at 3:01 AM, Steven LeBlanc <ores...@gmail.com> wrote:
>
> On Oct 20, 2013, at 9:54 PM, Mark Leeds <marklee...@gmail.com> wrote:
>
> > Bill: I didn't look at the code but I think the OP means that during the
> nlminb algorithm,
> > the variance covariance parameters hit values such that the covariance
> matrix estimate becomes negative definite.
>
> Yes, that is what I meant.
>
> >
> > Again, I didn't look at the details but one way to deal with this is to
> have the likelihood
> > function return -Inf whenever the covariance matrix becomes not positive
> definite. so, the
> > likelihood should check for positive definiteness first before it
> actually calculates anything.
> > If PD is not true, the -Inf value should push nlminb towards values that
> obtain a positive definite matrix. But there could be something more subtle
> going on that I'm not understanding. I don't know even what algorithm
> nlminb is using ( probably quasi-newton ) but this is one thing the OP
> could try.
>
> I tried this at your suggestion. nlminb() seems to hang at -Inf, but Inf
> works splendidly. Thanks much!
>
>
[[alternative HTML version deleted]]
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
