Occupancy Estimation across years

[Sarah Reed Hurteau]

Greetings,
I have been using the occupancy estimation models in MARK for a data set that includes 7 years of avian survey data. We are interested in using mountain range (3-4) and ~6 covariates in the estimation of occupancy (and probability of detection). However, for many of the species there are not enough "presence" records to run this many covariates. So, I am interested in using the Robust Design occupancy models. My question is, do the sites that are sampled have to be the same from year to year, and is there a way to index the result to get an estimate by year? Or would you use year as a covariate? Am I just adding more complexity to my problem of detecting rare species?
Thanks for your help!
Sarah

[darryl]

Hi Sarah,
Can you give us some more details of your study. How many surveys per year? How many sites? What are you trying to divine from the data? etc.

Note that the robust design occupancy model actually estimates more parameters (associated with changes in occupancy through time), so if you already have sparse data then switching to that model is unlikely to help. Likely what you need to do is restrict your candidate models down to those with combinations of up to only 2 or 3 covariates.

Furthermore, recall that the simpler occupancy model assumes there are no changes in occupancy of individual sites for the duration of the data (7 years in your case), or any changes occur at random.

Cheers
Darryl

Greetings,
I have been using the occupancy estimation models in MARK for a data set that includes 7 years of avian survey data. We are interested in using mountain range (3-4) and ~6 covariates in the estimation of occupancy (and probability of detection). However, for many of the species there are not enough "presence" records to run this many covariates. So, I am interested in using the Robust Design occupancy models. My question is, do the sites that are sampled have to be the same from year to year, and is there a way to index the result to get an estimate by year? Or would you use year as a covariate? Am I just adding more complexity to my problem of detecting rare species?
Thanks for your help!
Sarah

[Sarah Reed Hurteau]

Hi Darryl,
There are 3 surveys per year. But the number of sites and site locations change each year (with some sites being sampled in multiple years). This is why I was analyzing each year separately for each species.

Our objective for this preliminary study is to identify patterns in occupancy that can be used in a predictive model of riparian and upland birds in the Great Basin region.

The group who collected the data came up with the set of candidate models that are being tested, but it seems that there are not enough detections to sufficiently model all the covariates they want to include. I was hoping by combining all 7 years together, a detection history of that length would include enough positive detections to be able to model all of the covariates. But I was not sure how to deal with the fact that the sites are not the same among all the years.

Thanks,
Sarah

[darryl]

Sarah,
If most sites are being surveyed in consecutive years then there may be some benefit in using the robust design, I didn't appreciate you were doing it year-by-year at the moment.

However, if most sites are only surveyed in 2 or 3 years over the 7 year period, then trying to use the robust design model is likely to create a whole new set of problems. Do you expect the effect of the covariates to be similar in the different years? One thing you could do is to include the detections histories from all sites in all years to a single file (so the same site in different years would have multiple detection histories being entered), and include 'year' has a covariate. You can then consider a range of models that will allow for annual variation, and interactions with and the year covariate if you expect the effect of teh covariate itself to change over time.

When you say there's few detections, is it a case where if the bird species is detected in 1 survey in a year of a particular site, it also likely to be detected in the other surveys (so detection histories tend to be either 111 or 000), or is it a case usually there is only 1 out of 3 detections at a site. If it's the former then that suggests detectability is high so the sparseness is likely caused by the species being rare, but with the latter then the sparseness is likely caused by low detection rates (and/or rarity) so your results may be somewhat flaky because there isn't enough information with only 3 surveys per year to separate out detectability and occupancy.

It may also be a case that the folks you're working with may have to backtrack on what they reliably will be able to tease out from all of this and have to settle for models with only a couple of covariates at a time. There's nothing to say all your candidate models have to be nested.

Cheers
Darryl

Hi Darryl,
There are 3 surveys per year. But the number of sites and site locations change each year (with some sites being sampled in multiple years). This is why I was analyzing each year separately for each species.

Our objective for this preliminary study is to identify patterns in occupancy that can be used in a predictive model of riparian and upland birds in the Great Basin region.

The group who collected the data came up with the set of candidate models that are being tested, but it seems that there are not enough detections to sufficiently model all the covariates they want to include. I was hoping by combining all 7 years together, a detection history of that length would include enough positive detections to be able to model all of the covariates. But I was not sure how to deal with the fact that the sites are not the same among all the years.

Thanks,
Sarah

[Sarah Reed Hurteau]

Thanks Darryl!
You just confirmed my thinking regarding the robust design.

As far as the detection histories, it is more of the case where there are only 1 out of 3 detections at a site, and lots of records of no detections at all. Often there is one mountain range that has lots of detections while the other two mountain ranges have few to none.

I think we will just have to significantly reduce our candidate model set.
Thanks for your help!
Sarah

[Sarah Reed Hurteau]

Hi Darryl,

How would you code the year as a covariate if all years are stacked into a single encounter history? Do you run this in the regular occupancy model (not robust design)? Can you group by more than one "covariate"? For example, I have been grouping by mountain range, but can I also group by year too?

Thanks,
Sarah

[darryl]

Hi Darryl,

How would you code the year as a covariate if all years are stacked into a single encounter history? Do you run this in the regular occupancy model (not robust design)? Can you group by more than one "covariate"? For example, I have been grouping by mountain range, but can I also group by year too?

Thanks,
Sarah

You could do it either way, either as a series of individual indicator covariates (=1 for year X, 0 otherwise) plus interactions with other covariates if you're going to fit that sort of models. Or you could define 'groups' on the basis of year and mountain range (so group 1 = MR1, year 1; g2 = MR1, year 2; ...). Both approaches should give you the same results, but the group approach will probably give you more flexibility on the fly if there's an interaction with 'year' that you forget to define in the input file.
Darryl