Question about point-specific abundance estimates and ranef()

220 views
Skip to first unread message

Nathaniel Behl

unread,
Feb 22, 2017, 3:15:05 PM2/22/17
to unmarked
Hi all,
I'm currently trying to estimate point-specific abundances for a set of 137 locations where we conducted 3 rounds of point-count surveys, with each round of surveys conducted by a different observer. Because of the multiple visits/observers I have fit all of my models using gdistsamp() and unmarkedFrameGDS(). I have completed the model selection process and settled on one model using a hazard rate key function and including observer as a covariate to detection, that is most strongly supported and fits the data well.

I now want to use this model to estimate the true abundance at each site corrected for imperfect detection. Based on some of my previous research online and through this discussion group, my impression is that to get point-specific abundance estimates from this model I need to use the ranef() function and have essentially done:

top.model <- gdistsamp(~1, ~1, ~observer, data = data, keyfun = "hazard")
re <- ranef(top.model)
re

For all of the sites where I detected birds the resulting estimated abundances seem reasonable and correlate roughly with the number of detected birds at each point, however I also get back a very specific abundance estimate (11.54) for every site where no birds were ever detected. What's more, this abundance estimate is quite a bit higher than many of the sites where I did detect birds (mean abundance at sites with birds detected = ~12, range = 5.6 to 18.8).

I have every reason to believe that sites with no detections truly had no birds (or potentially only very few), and so I feel like the predicted abundance at all of those sites should be 0.

Can anyone explain to me why I'm getting such high abundance estimates at sites with no detections and how to correct this? Am I mistaken about ranef returning point-specific abundance estimates or am I using it incorrectly, or alternatively is there a better method I should be using to estimate point specific abundances?

Any suggestions would be greatly appreciated.

Thanks,
Nathaniel Behl


M.S. student, Benkman Lab
Department of Zoology and Physiology
University of Wyoming

mike worland

unread,
Feb 24, 2017, 6:18:44 PM2/24/17
to unmarked
Nathaniel, I think you can expect to have >0 estimates at zero-count sites using ranef().  My understanding is that ranef() derives estimates from an empirical Bayes procedure, of which the actual count is only one component (see Royle 2004, Royle & Nichols 2003).  I get >0 estimates at zero count sites all the time with pcount and occupancy modeling.  Having said that, it seems like your estimates are definitely too high: I don't think a site with >0 count should have a lower abundance estimate than a site with a zero count?  Maybe it's possible, but I haven't seen it.  One thing that jumps out at me is that your only covariate is observer and you have no availability or abundance covariates.  Both detection and availability could be affected by time, date, and weather variables.  Abundance of course by habitat characteristics, site, etc.  Did you try any of those?

I also noticed a few posts earlier than yours that someone recommended using just distsamp instead of gdistsamp if possible.  Or you could use pcount if you feel okay assuming a closed population across your three surveys (or at least immigration equaling emigration).  Then you would not use the distance estimates.  Distance estimates in point counts can be not only way off but also strongly biased, so it's worth thinking about whether they're the problem. 

Mike Worland  

John Clare

unread,
Feb 26, 2017, 5:43:50 PM2/26/17
to unmarked
Hi Nathaniel,

The ranef point estimate (the mean, at least) cannot be be zero, although the mode can. Derived density estimates (lambda* phi/area) might look more reasonable.  What do the estimates of p/phi look like? It seems likely that either detection or availability actually are quite low (i.e., sparse data), or that there is a lot of unexplained heterogeneity (i.e., highly variable counts across time/space) that might mess things up. I'm guessing that the issue is #2. Your model assumes that the number of birds available for detection at each site during each visit comes from the same mixture of distributions.  The probability of a site having 0 abundance given an expected value of 11 and Poisson variation is about 0.002%. 

John

L. Jay Roberts

unread,
Feb 27, 2017, 1:26:37 PM2/27/17
to unmarked
I recommend using distsamp instead. Yes gdistsamp allows you to incorporate the repeated visits explicitly, but in practice I don't think it's achieving what you hoped. First, the gdistsamp formula is in the order 1) lambda (abundance) 2) phi (availability) and 3) p (detection) and you have put the observer covariate in the detection function, when according to your explanation I would think it should be a factor in availability? 

In practice, unless there is a significant reason why you think availability would change between visits (e.g. temporary or annual migration) then gdistsamp I think has trouble attributing non-detections among the p and phi models. I would expect also that your lack of covariates on lambda has led to difficulty assigning variable abundance values among sites, try changing observer to the availability model and adding a few covariates to abundance. 

If you don't think there are major shifts in availability between visits, and even if there is a large difference in skill level among your observers, then if every site got three visits and three observers then all have roughly identical effort assuming you used standardized field sampling protocols. Therefore gdistsamp is probably not necessary and I'm guessing will continue to give you unsatisfactory results. 
Reply all
Reply to author
Forward
0 new messages