www.phidot.org

While working in PRESENCE I have run into the issue that my output is coming up with #IND00 and #QNB in the standard error and 95% confidence intervals.

psi 2 C04 : 1.0000 -1.#IND 1.#QNB - 1.#QNB
psi 4 C07 : 1.0000 -1.#IND 1.#QNB - 1.#QNB

Reading up on this issue, it appears I will need to either normalize or scale my covariates. My site covariables normalize perfectly fine, however I cannot get my sample covariable to. Whenever I tell the program to either normalize or scale this data, it does nothing. Is there a different way to get it to normalize?

Thank you in advance.

Hi Devin
If you check your output there's probably a warning about negative standard errors, or the variance-covariance matrix not being invertible. This means the reported SE's (and CIs) shouldn't be trusted, and are likely junk.

Bit hard to tell exactly what the problem might be from the detail in your post, but given some of your occupancy estimates are 1, you probably have some of the beta parameters (or logistic regression coefficients) being estimated as pretty large values. How much data do you have, and how many covariates are you trying to include in your model?
Cheers
Darryl

I have 49 sites with 5 surveys.

My top model currently has 5 psi and 3 p. My models run fine when I use 1 or 2 variables but none of those are significant.

That's a lot of covariates for that sample size. I suspect you're overfitting the model and finding spurious relationships. Do the effect sizes for the covariates make sense biologically? With 49 sites, I wouldn't recommend including any more than 2 covariates (you might get 3 if you're lucky) for occupancy. Otherwise you're just spreading your data too thin, which results in what you're seeing.

Also check the FAQ for the PRESENCE forum. There's some info in there about things you can try when you get warnings.

Hi Devin,

If you have missing data in the covariates, the normalize and scale functions didn't work (until just now as I've fixed and uploaded Presence to the web-server - version number didn't change). If that wasn't the problem, let me know.

Cheers,

Jim

Darryl,

They make sense to me, but I am very new at this. So far everything I have figured out as been from reading the PRESENCE textbook, user manual, etc. so I'm not entirely positive what exactly I am looking at.

Would I be better off doing a PCA first and then trying it out? Or doing this work through RPresence?

Jim,

I will let you know tomorrow. Unfortunately IT has gone home for today and I do not have the authority to update programs.

PCA would only help if your covariates are highly correlated, which I assume you already checked for, and aren't including such pairs of covariates in the same model. There are pros and cons to using PCAs as covariates, especially when they don't have an obvious interpretation. Not a fan of that approach personally, but that's just me.

RPresence wouldn't help, as uses the code from PRESENCE (essentially) anyway, so will hit the same issue.

The beta parameters are also labelled 'Untransformed estimates of ...' in the PRESENCE output. If any have really large absolute values (ie <-10 or >10), then I'd get suspicious of what's going on. Also check you site specific estimates of psi. If a lot of them are either 0 or 1, then that's also an indication of possible overfitting.

If you want to learn more about these methods, I am running a course in Nevada in a few weeks, or in Virginia in October. More info posted in the training/workshops section of the forum.

Cheers
Darryl

Ok, thank you for your help