multi-season models and seasonal variables

[emma]

Hi,
I had a question about modeling seasonal habitat covariates in a multi season model. I have 3 seasons of data with 3 surveys per season and am trying to model changes in occupancy and extinction in relation to habitat variables that also change by season. These aren’t site based covariates since they are not consistent across all 3 seasons but they’re not exactly survey based covariates either since they were consistent within each season and weren’t measured on every survey occasion.
So for example…

Survey data for seasons 1, 2, 3
Site 1 000 001 111
Site 2 110 --- 000

Habitat data for seasons 1,2,3
Site 1 2 .5 1
Site 2 .4 - 1

I found an earlier posting that suggested that these types of seasonal covariates could be modeled as survey based if I entered them in the first 3 columns and filled out the remaining columns with zeros.
Like this….

Site 1 2.51 000 000
Site 2 .4-1 000 000

But this doesn’t seem to work even with different combinations of entering zeros and dashes in the remaining 6 columns. I guess this is because the missing surveys no longer correspond correctly between the two matrices. The program does tell me I’m missing survey data and I get crazy AIC values like 4000000000000.

I was wondering if it would be appropriate to simply replicate the habitat value across all 3 surveys (since it was the same during all 3 occasions) so they could be entered as normal survey type variables and the matrix would still match up.
Like this…

Site 1 222 .5.5.5 111
Site 2 .4.4.4 --- 111

I get more reasonable AIC values when I do this, but I’m not totally sure what this is really doing.
Sorry for the lengthy posting. Any help or comments would be very much appreciated!
Thanks,
-Emma

[darryl]

Hi Emma,
You should set up your data with the habitat covariate for each year entered as a seperate site specific covariate, eg. Hab1, Hab2, HAb3, ...
Then in your design matrix instead of having a full column with just the one covaraite name, you have the sprcific name of teh covariate that relates to just that season, eg.
a1 a2
psi1 1 Hab1
psi2 1 Hab2
psi3 1 Hab3

For colonization you do have to put some additinal thought into things in terms of is it the covariate value at the start or the end of the period that's important, ie
c1 c2
eps1 1 Hab1
eps2 1 Hab2

or

c1 c2
eps1 1 Hab2
eps2 1 Hab3


Cheers
Darryl

[jhines]

Hi Emma,

Darryl's suggestion should do the trick, but I'm wondering if there is a problem with your data and/or covariates. The first suggestion you listed, entering the covariate as a survey covariate and adding columns of zeros after the last season, should also work. If you'd send me your latest backup zip-file in your project folder, I'd be happy to see why it didn't work for you.

Jim

[emma]

Thank you both for your quick responses!

I messed around with a bunch of sub-sets of my data and I think I figured out the problem. Either of the suggested methods seems to work as long as I get rid of any sites that are missing all of the surveys (and associated habitat data) for any one of the seasons.
So for example, a site with detection history of 000 110 0-- with habitat 1 2 3 is fine because it includes some info for all parameters, but a history of 000 --- 110 with habitat 1 - 3 would have to be dropped because the habitat changes with every season and the model doesn’t have any info for season 2. Same goes for missing season 1 or 3.

It makes sense that this would be the case, but unfortunately over ½ my sites are missing one of the three seasons of data…any ideas on a way around this problem short of dropping all these sites (I’d only have about 25 remaining)? Is it even feasible to try building these highly parameterized multi-season models with such a small sample size?
Thanks again for your help!

-Emma

p.s Jim- if you don’t think this was the problem I would definitely be very grateful if you would take a look at my project file as suggested and I will send it over.

[jhines]

Emma,

If you were modeling detection as a function of the covariate and have missing covariate values which match the missing detection data, it should work because the detection probabilities are not computed for surveys with missing data. However, modeling extinction or colonization would cause a problem because those parameters are used even when there are no detections.

A couple of possibilities which may or may not make sense to you are to use the an average value for the missing covariate data, or perhaps the value of a nearby site.

Jim

[emma]

Hi,
That makes sense. I don't think averaging or estimating the covariates will work in my case, but I'll give all of this some thought.
Thanks for the help!
-Emma