variance-covariance question - have RTFM.

[goodwins]

I am attempting to model constant occupancy with detection as a function of observer in a single season analysis (occupancy(.), p(observer)). I have ten visits to 30 sites, and two observers alternated surveys. The species I am investigating is common, with frequent detections at every site. I get the same error often noted - the variance-covariance matrix issue. I have RTFM (although this error is not, in fact, discussed in the help menus, online book or anywhere other than the FAQs for this forum).

Following your advice, I looked at the following:

1. Identify which real parameter is essentially 0 or 1 and use the ‘fix parameter’ option to constrain it to be that value.

I don’t think any of the parameters are essentially 0 or 1.

2. If using covariates, try a different covariate standardization.

I switched observer covariates from 0 and 1 to 1 and 2, no luck.

3. If using a categorical covariate, try using a different category as the “standard.”

(same as 2, here)

4. Try a simpler model.

Not possible.

5. Try a different parameterization of the same model.

I’m not sure what this would achieve – could you explain a bit further?

[jhines]

Sarah,

Thanks for sending the data and output (off-list). Looking at the data, I noticed that all sites had at least 1 detection. So, I would expect occupancy to be 1.0, and that's what PRESENCE gives. The problem is that it can't get a variance on that estimate and it screws up the whole variance-covariance matrix, giving you the error/warning message. The solution is to fix psi equal to 1.0, and let PRESENCE estimate the detection probabilities. To do this, click on the occupancy tab in the design matrix, then click on the '1' in the 1st row, 1st column, then right-click and select 'delete column' from the menu. (This is important as it deletes the 'beta' parameter associated with psi.) Then, in the 'Setup Run' window, click the 'fix parameters' button and enter a 1 in the box next to 'psi'.

Jim

[elt]

I got the same message for one of my models. There are NO covariates (psi(.), p(.)). Occupancy is certainly not 1 (naive estimate was .09). I only had 46 total observations out of 436 sites. Is this the problem? Is this close enough to 0 to fix it at 0? Any help would be appreciated.

[jhines]

Hi,

The occupancy estimate from PRESENCE should be higher than naive occupancy. Also, if occupancy is zero, no sites are occupied and you have no data. Since this isn't the case, there might be something peculiar about your data. Could you send it to me off-list?

Thanks,

Jim

[darryl]

elt,
You say occupancy was not 1, but was it estimated to be 1? It may be a case you you're data being too sparse for the methods to work. When the data is sparse there's a tendency for psi to be estimated near 1. How many repeat surveys were conducted per site?

[jhines]

Liz,

Thanks for sending me the data. There were 436 sites and 3 surveys for a total of 1305 possible detections. The species was detected at 44 out of 1267 non-missing cells. PRESENCE estimated detection probability (p) at about .035 for the simple model (psi(.),p(.)), and occupancy (psi) = 1.0. The naive estimate of psi was about .10. The reason the estimated psi was so high is that p is extremely low and there are only 3 surveys. The probability that an occupied site goes undetected all 3 times is around 0.90 (1-p)*(1-p)*(1-p). With p this low, it's very difficult to resolve which sites are unoccupied and which sites were occupied but undetected. This is why the standard error of psi is so large. Unfortunately, there is no way to get a better estimate from these data as there is not very much information in the data.

When planning occupancy studies for difficult to detect species, it is suggested that many surveys for each site are conducted in order to help with this issue (big chance of not detecting the species in all surveys).

To make some use of these data, one possibility is to pool this species with another species which has a similar detection probability.

Cheers,

Jim

[elt]

Thanks guys, that's sort of what I expected, with so few observations. Luckily I am analyzing about a billion species, most of which have worked fine, so leaving out the woodpecker, or pooling it, is not a huge deal. Thanks so much for your help.

[darryl]

It's not the ivory bill is it?

As Jim suggests though, you could model it together with (an)other species that you expect to have a similar level of detectability to get something better than 1.

Other option is to go Bayesian and assign an informative prior to p.