Grouping species and using time covariates in detection analysis

62 views
Skip to first unread message

Katrina Fernald

unread,
May 21, 2021, 4:53:00 PM5/21/21
to distance-sampling
Hello!

Apologies in advance if these questions have been asked in a prior message.

I'm working with bird point count data from 3 sites with control stands and treated stands. I would like to compare community composition and densities of particular species between the control and treatment. There are 10 species that were detected more than 60 times, which is often the minimum recommended for calculating an estimate of detection, but I've also seen at least 80 detections recommended. Which minimum is more widely accepted, or does it depend on the study design?

Either way, this leaves a few dozen species with insufficient detections, so I would like to group them for analysis. From what I understand, these should be species with similar detection probability, but how can I know that without detailed information on call rate, volume, etc? Is it robust enough to use functional traits like ground foraging (which is how I would like to present the results anyway)?

A somewhat unrelated question: anytime I include a covariate related to time in the ds() function, it returns a "no models could be fitted" error. Any insight on that would be much appreciated, here's and example of the code, with mas meaning minutes after sunrise:

birds.hr.time<-ds(birds, key = "hr", transect = "point", formula = ~mas) 

Katrina Fernald
MS student, Forestry
Iowa State University

Stephen Buckland

unread,
May 22, 2021, 6:43:49 AM5/22/21
to Katrina Fernald, distance-sampling

Katrina, you generally need more observations in point transect sampling than in line transect sampling to get comparable precision, hence the suggestion of a larger sample size for point transects.  But these are just cookbook rules.  Whether you can get good estimation with smaller sample sizes depends on many factors.  If you get good fits to your data and the shape of the detection function is plausible, given what you know about your species, I would not apply these rules rigidly.

 

If you use mcds, you can model data from multiple species, and include species as a covariate in your detection function model.  Avoid pooling together species for which the detection process is expected to be quite different – e.g. mostly visual detections for one and mostly aural for another.  But if a group of species all tend to be detected in much the same way, then it should be OK to pool, even if some have e.g. a louder song than others – that will be taken care of by including species as a factor in your model.

 

I’ll leave someone else to respond to your final question …

 

Steve Buckland

--
You received this message because you are subscribed to the Google Groups "distance-sampling" group.
To unsubscribe from this group and stop receiving emails from it, send an email to distance-sampl...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/e9ede235-2a9d-49bc-a5e4-f08ecc7261f5n%40googlegroups.com.

Eric Rexstad

unread,
May 22, 2021, 9:55:56 AM5/22/21
to Katrina Fernald, distance-sampling

Katrina

Regarding fitting detection functions with covariates (whether or not the covariates are related to time), the first thing you should do is perform exploratory analysis of your data; before adding the covariates to the detection function.

The case study we have online, below, shows some exploratory plots

http://examples.distancesampling.org/Distance-covariates/covariates-distill.html

It may be the case that some covariates contribute nothing to the fit of a detection function to your data (the covariate does not influence detectability); or the relationship of detectability and the covariate is highly nonlinear, or there is an outlier in the data.  Any of these situations can result in the inability of the software to fit models containing certain covariates.  You'll need to dig in your data to understand the nature of the challenge you are presenting to the software.

To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/DB9PR06MB75948F103ACB78EE6580629FC5289%40DB9PR06MB7594.eurprd06.prod.outlook.com.
-- 
Eric Rexstad
Centre for Ecological and Environmental Modelling
University of St Andrews
St Andrews is a charity registered in Scotland SC013532

Katrina Fernald

unread,
Jun 4, 2021, 6:53:19 PM6/4/21
to distance-sampling
Thank you for these detailed answers!

Eric, what you're saying about covariates makes sense. It looks like my key only models with adjustments have lower AICs than the models with covariates anyway, so the pooling robustness assumption must be working.

Steve, I'm only using aural detections to eliminate some bias so I've tried fitting the detections function with all species included. After leaving it running much of the day, it hasn't finished. Is there something wrong with this code, or is running ds() with a factor with 61 levels just going to take a long time? If it's the latter, could or should I remove species with few detections? Here is the code:

birds.covs$SppCode<-as.factor(birds.covs$SppCode)
birds.sp<-ds(birds.covs, key=hn, transect="point", formula~`SppCode`)

Thank you again!

Katrina Fernald
MS student, Forestry
Iowa State University


Eric Rexstad

unread,
Jun 5, 2021, 3:58:20 AM6/5/21
to Katrina Fernald, distance-sampling

Katrina

I've not tried to fit a detection function with >10 parameters, so fitting a model with >60 parameters is an adventure.  I'm not optimistic about your chances of success.

You shouldn't expect to produce reliable estimates for species with a small number of detections even using the covariate approach.  Give thought to constraining your analysis to species for which you will have some confidence in the result.  As a middle ground between "all species have same detection function" and "each species has a unique detection function, requiring 60 parameters to estimate", you could take a "guild" approach and merge species that have similar vocal properties, considering the possibility they share a common detection function.

Reply all
Reply to author
Forward
0 new messages