Changes in occupancy between cycling seasons

[bmattsson]

I am analyzing a dataset of birds captured in mistnets in Costa Rica during three consecutive seasons across three years. Though sites were sampled during multiple days each year, I have consolidated detections across each year to account for the low capture rates. So, essentially, I am looking at three seasons, with three visits (i.e., years) per season. I am principally interested in modeling changes in occupancy between seasons, which are defined by annual patterns in breeding phenology of the avian community and precipitation patterns. I was wondering if it would be legitimate to construct detection histories in the following way (s=season, y=year):

s1.y1 s1.y2 s1.y3 - s2.y1 s2.y2 s2.y3 - s3.y1 s3.y2 s3.y3 - s1.y1 s1.y2 s1.y3

My thinking is that this will allow me to estimate changes in occupancy between seasons 1 & 2, 2 & 3, and 3 & 1, effectively "closing the loop". Of course, detection rates would be estimated twice for season 1, but this seems acceptable to me.

Any thoughts on this would be appreciated.

-Brady

p.s. Each row in the dataset consists of a bird guild-site combination, where 13 guilds are defined by life history characteristics, and a total of 7 sites were visited. Some sites were not visited during some years, and I accounted for these with missing entries in the detection history.

[darryl]

Hi Brady,
What exactly is you're trying to estimate? Is it at individual species level or community level?
Darryl

[bmattsson]

Hi Darryl,

Due to the sparse detections, I pooled detections across species into the 13 ecologically-defined guilds. I tried running some simple models for species-specific detection and occupancy, but the mean detection rate across species was <0.3, which seems problematic for estimating occupancy. Mean guild-level detection (assuming constant occupancy) was 0.4. So, I am therefore trying to model patterns in guild- (or cross-guild) level occupancy.

What I'm really wondering, though, is whether there are any problems with using a set of detection histories for the first and last seasons in the context of multi-season occupancy modeling.

-Brady

[darryl]

So your p-hat=0.4 is for when you've pooled your data across days within each season?

Low p's are only problematic in the sense there require more repeat surveys such that you can reliably separate out detection from the interesting stuff.

What you're proposing does seem to be cheating in some sense, but perhaps it wouldn't be too bad provided you make sure you estimate detection completely separately in the last season compared to the first so the same data doesn't get used twice.

Ideally what you could do is keep the years and seasons in the correct chronological order and use the days as your within-season repeat surveys if p is going to be high enough.

[bmattsson]

Hi Darryl,

Sorry for the delayed response.

Treating individual days as replicate samples resulted in very low detectability estimates, so I pooled data for individual years instead. So, each year*season represents a replicate sample for each guild*site combination.

Thanks for your suggestion regarding separation of the detection estimates. I have now made sure that I am estimating p separately for the repeat season.

-Brady