Multi-site Missing Data

[EmmaRigby]

Hello,
I've rtfm several times and have searched in the forums but am still unsure about how to proceed so thought now would be the time to ask for help!

I have data from 6 bat roosts over a 5 year period (115 sampling occasions for 683 bats in total). I would like to split the data into two seasons for each year (spring-summer) and (autumn) with an interval over winter when the bats are hibernating, giving me an encounter history for each bat looking something like this:

00AA000000 1 1 0;
00A0BG0000 1 0 1;

I've also got information on age and sex of each bat so have entered them as individual covariates.

The problem is that because the data is real (and therefore not ideal!) some roosts have been sampled a lot more than others, and although each roost has been sampled at least once per year when I split the years into seasons not every roost has been sampled in every season.

I read something in the forum about how I can code missing data in the encounter history with a '.' but am unsure about how to go about this - how would MARK know which roost I am referring to when the sites visited are in no particular order and only one site is visited at a time? I also presume that I would need to constrain the parameters for the data coded as '.' to 0 when I am running models?

Any help would be much appreciated!

[sbonner]

Hi Emma,

This looks like a multistate model to me. You have individuals that can move between different sites (states) and sampling in each of the states. One simple way to handle the missing data would be to fix the capture probabilities for the site-by-year-by-season combinations with no sampling to 0.

If you have good information on the amount of effort than you could even build a model that uses the number of visits to the site as a predictor of the capture probability. Very simply, you might assume that the capture probability for site i in season j is:
p_ij=\beta x_ij
where x_ij is the number of visits made to site i in season j. More complicated models might allow \beta to vary by site, by year, by season, or some combination of the above.

Hope this helps.

Cheers,

Simon

--
Simon Bonner
Statistics and Actuarial Science
Simon Fraser University

[jlaake]

One thing that may not be clear from Simon's post is that you can't use the dot notation as you described for a multi-state model because all you know is that you didn't sample certain states and you don't know the state for each animal so you can't assign a . because it could have been in one of the states that was sampled but it wasn't seen.

So as Simon suggested you need to set the p=0 for the occasions at which the states were not sampled. In the ch you leave the entries as a 0 if they were not seen.

--jeff

[EmmaRigby]

Hi,
Thank you both for getting back to me. I thought I might have to go down the route of constraining parameters so had already tried that, and after lots of churning away by the computer I was informed that the general models could not reach convergence. Do you have any idea of how I might be able to get around that?
Thanks