0-1-estimates of Psi in single season single species model

[R.S.]

Hi everyone,
I have started to analyze camera trapping data from 119 trapping sites with 10 "visits" (occasions) each, and a total of 90 detections of my target species spread over 37 sites in my data set. So, not exactly abundant data...
I am using single species single season models.
I am using two covariates to model p, and so far everything looks ok, but I am having trouble with my covariates for Psi. With some covariates I get only 0 and 1 estimates of Psi (and, of course, the convergence and variance-covariance matrix warning). Sometimes, one covariate alone works well, (no warnings and reasonable estimates) but if I add a second one, I get the 0 and 1 estimates. And sometimes, when I add a 3rd covariate, estimates look ok again (no warnings).
None of the continuous covariates I use has an extreme range (otherwise I scaled them in PRESENCE). Is this just a sparse data syndrom, or is there any other reason why some covariates don't work and others do?
In the case where 3 covariates improve the lousy 2-covariate model, can I have any confidence in the results at all, knowing that the simpler model essentially gave no results?
I would greatly appreciate any help, or hint where else to get help (I searched the forum and online book, but maybe I overlooked something). Thanks a lot already!

[cnagy]

In my very very novice opinion it sounds like more of a problem with your covariates than in your study design...if some of the covariates in a particular model are really correlated then the analysis can't pull the effects of the two apart, so you will get wacky betas and SE's even if you looked for your animal pretty thoroughly. Similarly, if you have a noisy covariate effect on psi that will (rightfully) blow up your CI's. This could be a too-few data problem or just the fact you have a complicated system...which i guess is just another data problem in the end.

With your 2 term vs 3 term models issue, try playing with the model set (maybe this is not kosher philosophically). Like,

A+B - errors, bad CI
A+B+C - seems ok

try
A+C
B+C
A
B
etc.
and see what happens. If one covariate is being a jerk, you should be able to isolate it and maybe see why.

Another guess is if p is very low (and i think 10 visits should estimate p fine in most cases, but it still could actually be low), then the psi will be much higher than the naive presence. This can make nearly all of your sites "occupied" and then screw up your attempts to model psi.

You also might standardize all your covars instead of just some.

I don't think there is anything inherently wrong with a 3 parameter model working better than a 2. What did the AIC info say?

-chris

[darryl]

If you're getting the convergence warning, check out the FAQ to see whether it's a problem or not. You should also check that as you add a covariate, the -2log-likelihood value should always get smaller. If it doesn't the current model may have got stuck on a local-maximum; try different starting values.

If all of the above looks ok, then it may just happen that those 2 covariates essentially perfectly align with those sites with at least 1 detection (psi=1), and those with none (psi=0). But then this doesn't explain why things are working 'correctly' with an additional covariate, my guess is that you might be getting stuck on a local maximum.
Cheers
Darryl

[jhines]

Rahel,

I looked at your covariate data and found that for some combinations of the covariates, you only have one site (camera=0, wgt-no-ind=2.36 or wgt-no-ind=1.02) with which to estimate occupancy... and you're trying to model occupancy as a function of another covariate. So, for some combinations of the covariate, you have enough sites to estimate occupancy as a function of the covariate, but others you don't. I'm a little surprised that the other covariate on psi worked, but it could be due to that covariate having more variation.

So, what I think is happening is that when the variation in p is modelled with the two covariates, the sites tend to be grouped into very low or very high occupancy by the 'problem' covariate, giving estimates of zero or one. The likelihood is in range with the other models, so I think that's the best that can be done for this model.

Cheers,

Jim

[R.S.]

Thanks so much everyone for the help, and especially Jim for looking at my data. Looks like I might have to re-think my covariates.
Cheers.