season-specific covariates

[dbuhl]

I am fairly new to using Presences and am having a problem with entering and using season-specific covariates in a multi-season model. I have tried suggestions I found on this forum and still have had no luck. I have three seasons with three surveys each season. I have 10 covariates with 6 of these measured only once for the site and the other four measured once each season. Not every site was surveyed each season, so I have some missing data. Also, the species we are working with is fairly rare so out of 70 total sites, I have anywhere from 5 to 12 sites with detections a season.

I am having problems with the season-specific covariates since there is no sheet to enter them. Do I enter these in as site covariates or survey covariates and how? Then if I want to model occupancy as a function of one of these season-specific covariates, how do I do this? Anything I have tried has failed to converge or gives me an error because of missing values.

When I model occupancy as a function of one of the site-specific covariates, I am also having some convergence issues and am getting SE=0 for the parameter estimates. Is this due to the sparse data, or am I most likely doing something wrong? For this model, I am choosing the second model parameterization and then add a column in the occupancy tab and list my covariate, I am doing nothing in the colonization tab, and I am using the full identity matrix in the detection tab (but it does not matter whether I use constant, seasonal, or full identity, the model still fails). Does this model sound reasonable, or am I more confused than I thought I was?

Thanks.

[jhines]

Seasonal covariates can be entered as survey-covariates, with missing values filling out the right-hand side of the table. For example, if you have 3 seasons with 3 surveys each, there should be 9 columns for the survey covariates. Fill in the first 3 columns of the survey-covariate table and enter '-' for the rest of the columns. Then, in the design matrix for occupancy, add a column to the matrix and enter the covariate name in the new column. For example, the occupancy design matrix would look like this:

Code: Select all

-    a1     a2
psi1  1   covar1
psi2  1   covar1
psi3  1   covar1

If you're having convergence problems, then the best thing to do is simplify the model, so I'd suggest starting with detection constant over surveys (one column in the design matrix), then try seasonal detection, then the full identity. With fewer parameters to estimate, it is more likely for the model to converge.

If you can't get the simplest model to converge, I'd be happy to look at it and make a suggestion. Just email me the latest backup zipfile in your project folder.

Jim

[darryl]

dbuhl,
Which parameterization are you trying to use for the multi-season model? The second and third ones can have problems when you're trying to also include covariates on occupancy because there's an internal constraint that must be satisfied to ensure that the colonization or extinction probabilities (whichever is being derived by PRESENCE) for all sites are between 0 and 1. Check near the end of the output (where the derived parameters are) to see if there's any indication this might be happening.


In terms of entering the season-specific covariates you could do as Jim suggests, or alternatively you could enter them as a series of site-specific covariates. For example, for some covariate X that was measured in each of the 3 years, enter it as 3 site-specific covariates; X_yr1, X_yr2 and X_yr3. You'd then set up your design matrix to look like:

a1 a2
psi1 1 X_yr1
psi2 1 X_yr2
psi3 1 X_yr3


The advantage of setting up the covariates in this manner gives you a bit more flexibility. For example for colonization you could have
b1 b2
gam1 1 X_yr1
gam2 1 X_yr2

or

b1 b2
gam1 1 X_yr2
gam2 1 X_yr3

Depending on whether you think it's the value of the covariate at the start or the end of the period that's important for whether a site gets colonized (recall this is a between year event).

That all said, PRESENCE doesn't like missing values for site/season specific covariates. If you're trying to fit a model with a site/season specific covariate that is missing for some sites in some years, PRESENCE will use a default value for that covariate (I think it's 999) when it's trying to calculate the occupancy, colonization or extinction probabilities (where ever you've placed that covariate into the model), even if you don't have any observations for that site in that year. That may be contributing to your problem. You could test it by specifying a value for those site/year combinations and see if you get different results.


Cheers
Darryl

[jhines]

Hi Deb,

I think Darryl's right about the 2nd and 3rd parameterization having trouble due to the way the non-estimated parameters are computed. In the 1st parameterization, psi2,psi3,... are not estimated as they are computed using gam1,gam2,eps1,eps2... In the 2nd and 3rd parameterization, gam or eps are computed using the seasonal psi's, but this computation involves dividing one estimated parameter by another, resulting, possibly with a gam or eps > 1.0.

Also, it's very important to have covariate values for each site if you're trying to build a model where a parameter depends on the covariate. I use 999 as a default covariate value in hope that it will cause a weird result (along with the error message when you run PRESENCE). What most people do in this case is either delete the sites with missing covariate data (if there aren't too many), or replace the missing covariate data with the mean value of the covariate from other sites.

The other important factor to keep in mind is that if you want to model a parameter as a function of a covariate, there needs to be sufficient data to support that relationship. If a parameter is near zero or one, there isn't much variation in the parameter across sites to estimate a relationship between the parameter and the covariate(s). Also, if there are only a few sites which are contributing information about a parameter, you won't be able to get enough information to estimate a relationship. The estimate of gam is based on unoccupied sites in the previous season. So, if almost all sites are occupied, then there isn't much information available to estimate gam. Same for extinction, except that eps is based on occupied sites in the previous season.

That said, I tried a few models with your data and found that the 1st parameterization converges with seasonal colonization and extinction, season*survey-specific detection probabilities and occupancy as a function of the pctupland covariate. With this model, colonization and extinction are estimated at zero for the 1st interval (between seasons 1 and 2), so it wouldn't make sense to model those using the covariate. This model does give estimates of occupancy for the other two seasons as 'derived' parameters. I'd suggest sticking with the 1st parameterization for this data-set.

With the sparseness of this data-set, I was a little surprised that this model converged. I wouldn't think a model more complicated than this will converge, but feel free to try.

Cheers,

Jim