Can't use covariates in "2 groups, Constant P" SS model?

[Tony G]

Hello,

I'm working on single season occupancy model for mapping seagrass. I've collected detection/non-detection data at 89 sampling units (50 * 50 m) with an underwater camera (10 random photos in each unit) and using environmental covariates such as depth and substrate type to model occupancy. Preliminary results show the pre-defined "2 groups, Constant P" model is the best by a long way (deltaAIC 45 second ranked model), which is due to there being units that contain a lot of seagrass (p=0.83) and others with only some (p=0.17).

There does not appears to be a way to create a custom "2 groups" single season model in PRESENCE, whereby Site Covariates can be used to model occupancy in the design matrix. Would it be okay to classify the sampling units and use this as a Sampling Covariate in a custom model (1 group)? Or should I just ignore the 2 Group model?

Thanks
Tony

[darryl]

Hi Tony,
Unless Jim has snuck it in there recently, there isn't a '2 groups' type single season model that you can use with covariates. If the 2 groups model that you've tried is that much better then there is probably some variation in detection not being explained by the other models. Have you tried any covariates on detection so far? You could use some classification of the units as a covariate (I'd enter it as a site-specific covariate though as it would be the same value for all surveys) provided that you're not classifying based upon the observed detection-nondetection data otherwise you're using the response variable (detection) to help define a predictor variable which is something you don't really want to be doing (it will fit really well though! ). Otherwise you could include some of the other covariates you have on detection and occupancy to see if that helps.

Cheers
Darryl

[jhines]

Hi Tony,

I assume from your message that you ran a model where detection is a function of seagrass, since you have detection estimates for lots of seagrass and only some seagrass. And, this model did not do as well as the 2-group model?

I did add the ability to do a 2-group model with covariates recently (without mentioning to Darryl). To run this model, select the single-season custom model, click the 'detection' tab in the design matrix window, then click the 'Initial' menu and select 'Heterogeneity'. You should see double the number of p's, one for each group, and an additional tab for pi (prob. of site being in group 1). You could then model occupancy as a function of covariates, but I would not recommend modeling detection as a function of covariates.

Cheers,

Jim

[Tony G]

Hi Darryl,

I did use a visibility index to test if water quality influence detection but could add in some more detection covariates to explain the differences being picked up in the 2 group model, although it appears Jim has been working on a solution to this.

Hi Jim,

I ran a custom "2 group constant P" as you suggested and this matched the pre-defined model, but when I ran a custom Heterogeneity detection model with occupancy covariates the model didn't estimate 2 detection groups (i.e. only estimate one detection). Example output below:

Untransformed Estimates of coefficients for covariates (Beta's)
======================================================================
estimate std.error
A1 psi : -1.248468 0.257555
A2 psi.depth : -0.235005 0.282097
B1 P1[1] : 0.394751 0.126664
B2 P2[1] : 0.394751 0.126664
C1 pi : 0.000000 9.999997
======================================================================

Cheers
Tony

[jhines]

Hi Tony,

I don't see why it isn't working for you. Perhaps, you have sparse data, or a relationship between the covariates and detection and occupancy. Would you mind sending me your data (or latest backup zipfile)?

Jim

[jhines]

Hi Tony,

I ran your data through PRESENCE and found that the heterogeneity models need a little help to get started. By default, all parameters start with the value 0 (on logit scale, =0.5 on real scale). Starting with both p's equal, probably makes the likelihood jump around quite a bit when it starts to vary them in different directions. I used starting values of 0 for the first two parameters (psi intercept and psi-covariate effect), -1 for p1, 1 for p2, and -1 for pi. So, in the initial values input box, enter:

0
0
-1
1
-1

This starts the likelihood off with a low value for p1 (using the -1 init value) and a high value for p2. Even if these are not near what you end up with, the likelihood search function has an easier time if the p's are different, so it is more likely to work.

Jim