Skip to content
This repository has been archived by the owner on Jun 13, 2023. It is now read-only.

parameter diagnostics indicate MLE estimates hitting bounds #11

Open
jockongit opened this issue Oct 20, 2016 · 3 comments
Open

parameter diagnostics indicate MLE estimates hitting bounds #11

jockongit opened this issue Oct 20, 2016 · 3 comments

Comments

@jockongit
Copy link

Hi Jim and all

I've got a situation whereby checking the diagnostics related to parameter estimation shows me that the following parameters are hitting their (upper) bounds in the case of some species. Might you have any advice on 1) what to try in order to avoid the problem and 2) what the implications might be?

"logetaE1"
"logetaE2"
"logetaO2"

The upper bounds for each of those seems to have a default value of 3.3396581.

The command used to check the diagnostics table:
print( Opt$diagnostics[, c('Param','Lower', 'MLE', 'Upper', 'final_gradient')] ))

And the instructions that led me to check them: "...print the diagnostics generated during parameter estimation, and I confirm that (1) no parameter is hitting an upper or lower bound and (2) the final gradient for each fixed-effect is close to zero."

Thanks,

Jock

@James-Thorson
Copy link
Contributor

logeta is the reciprocal of the log-standard deviation for each spatial (O) or spatio-temporal (E) term (1 refers to encounter probability; 2 refers to positive catch rate). When it hits the upper bound, it is essentially driving the SD for that term to zero (i.e., no spatial or spatio-temporal variation in a given model component).

You can do two things:

  1. Turn off the corresponding spatial/spatio-temporal component via FieldConfig. This can of course be automated in the R script
  2. Increase the upper bound by, corresponding to a lower minimum SD, and then not worry about it. If the spatial/spatio-temporal term has a MargSigma below say 0.001, then the maximum likelihood estimate is essentially "turning it off". This can cause problems with estimating standard errors sometimes, but if it doesn't then I think this option is fine.

I include strong advice about not hitting bounds in the markdown document, but it applies more to the range parameters logkappa.

@jockongit
Copy link
Author

Thanks Jim, appreciate the advice. I don't recall coming across MargSigma - could you perhaps point to where I would find it?

@James-Thorson
Copy link
Contributor

That should be in the report 'Obj$report()', or it might be SigmaMarg I
can't remember

On Oct 20, 2016 2:26 PM, "jockongit" [email protected] wrote:

Thanks Jim, appreciate the advice. I don't recall coming across MargSigma

  • could you perhaps point to where I would find it?


You are receiving this because you commented.
Reply to this email directly, view it on GitHub
#11 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AHnqTeGXck3NjEXpymkriFWiQlxDx4dZks5q19yagaJpZM4KcbA_
.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants