random effects for detection models

[Brett Walker - DNR]

After reading through numerous papers, vignettes, and googlegroup posts, I have a few unresolved questions about the unmarked and ubms packages that I'm hoping folks with more experience in this group can help answer.

For background, I am hoping to use either unmarked or ubms to analyze unreconciled  independent double-observer count data using N-mixture models to estimate the detectability of birds using small, distinct breeding colonies (as well as the effects of a handful of covariates, including distance from the colony, optics used, weather, etc.). Unlike most studies, I'm not interested in estimates of either absolute or relative abundance or trends, rather, for this particular analysis, I'm only interested in quantifying variation in, and factors affecting estimates of detectability (p).

First question: Can I fit random effects (specifically, random intercept terms) specfiically in the detection model in the unmarked  package? Or can I only fit random effects in ubms? Or only in the abundance model (which is what most studies do)? From an earlier post by Marc Kery (on May 15), it sounds like it's now possible to fit random effects using the pcount function in unmarked (e.g., if I recall, his group was using a random intercept for site in the abundance model as well as a random intercept for observer in the detection model), but I wanted to make sure. The reason I ask is that how we collected data means repeated counts conducted on the same morning at the same location are non-independent (my crews conducted six consecutive 5-minute unreconciled independent double-observer counts per visit, with 4-8 visits per year over 3 years at ~43 colonies.

Second question: Can unmarked or ubms fit hierarchical random effects (random effects of visits within colonies)?

Third question: Papers critcizing N-mixture models have expressed concerns about bias in estimates of absolute abundance (N) from N-mixture models, especially when one has too few sites, too few repeat visits, or low detection probability. Based on my  understanding of how N-mixture models work, I assume the same concerns about bias apply to estimates of absolute detection probability (p). Is that correct? In our study, we anticipate high detection probability (maybe 0.80-0.90) and we have numerous repeated counts per visit and per location (as explained above), but that may be irrelevant if any assumptions are violated.

Thanks for all the time and effort the authors have put into developing and improving these models over time and helping researchers use them effectively, it is much appreciated!

Brett

[Ken Kellner]

1. Assuming you are able to use the multinomPois/stan_multinomPois function, then yes, random effects are supported by both unmarked and ubms. Typically I have found ubms does a better job with random effect estimates. It's a good idea to fit the model with both though and make sure you get similar results.

2. It really depends what exact structure you need. If you are able to write out bugs code or describe your situation in more detail I may be able to give you a better answer.

3. I'm not quite sure what you mean by absolute detection probability, maybe someone else on the list can provide better thoughts. I think in general if all you care about is the detection patterns and you have plenty of data you don't need to worry as much about some of the common N-mixture model challenges.

Ken

[Brett Walker - DNR]

Ken,

Thank you for the quick response.

1. I assume the pcount function in unmarked and the stan_pcount function in ubms also currently allow random effects, is that correct ?I am planning to use the pcount function because the data type is repeated counts. With an *unreconciled* indepedent double-observer approach (Riddle et al. 2010), counts from each observer are treated as "repeat visits" rather than double-observer counts. Colonies are sometimes quite large (150 birds or more), so determining which specific, unmarked individuals were and were not counted by each observer during each interval is impossible.

2. I was asked if unmarked or ubms supported hierarchical random effects by a statistician who has not worked with those packages. May I contact you off-list with more details to see what is and isn't possible?

3. By absolute detection probability, I just meant p. I'm interested in understanding the effects of different covariates on p, but I'm also planning on using p as a parameter in future simulations of count data, so I'm hoping to be able to accurately estimate p under different, specific, realistic sets of covariate conditions.

Thanks again,

Brett