Analysis Questions Regarding an Avian Abundance Study

445 views
Skip to first unread message

Russell Stocker

unread,
Jan 25, 2015, 1:43:23 PM1/25/15
to unma...@googlegroups.com
Hi,

I am currently analyzing data from an avian abundance study that was conducted in Pennsylvania.  Below are some of the specifics about the study.

1) The study used 26 different sites that are located across the state of Pennsylvania.  Within each site, there were places (subsites) randomly chosen to count birds.

2) Within each subsite, birds were counted (over a 5 minute window) along with the distance they were away from the surveyor.  The distance was recorded as either being in 0-50 or 50-100 meter bins.

3) The data was collected over two years and each site was visited twice during a year. 

4) Site specific covariates were collected for each visit (wind conditions, time, surveyor, etc.)

There seems to be a couple of different routes that could be taken when doing the analysis with the functions in the unmarked package.  Below I give those and my comments/questions regarding each in terms of this data set.

A) Use the gdistsamp function.

1) Would the analysis need to be done for each year separately since the population would be considered open over a two year period?

2)  Will there be estimation problems?  There are only two distance bins and two visits per year.  These are the bare minimum required for estimation (probability of detection and availability) and the counts per bin per visit are small.

3) Can the model be fit for all 26 sites at one time or would it need to be fit separately for each of the sites?

4) Does the amount of time per site (5 minutes), play a role in the model?

I have fit this model by year several times and the convergence problems are significant. When optim does converge it takes a long time to do so.

B) Using the pcountOpen function.

1) Does it make sense to reformat the data into point counts and ignore the distance bins?
Will this result in a sampling design that can be used with the Dail and Madsen (2011) model? This would take care of the small number of bins and the model allows for an open population.

2) Are there any real advantages to using the distance sampling design versus the point count design, especially when there are only two distance bins?  This study will be conducted in the future and I want to help improve the study's design.

3) 
Can the model be fit for all 26 sites at one time or would it need to be fit separately for each of the sites?

4) Does the amount of time per site (5 minutes), play a role in the model?

Thanks for your help,

Russ

Jeffrey Royle

unread,
Jan 25, 2015, 2:07:14 PM1/25/15
to unma...@googlegroups.com
hi Russ,

 Yes you should use the gdistsamp function if you have distance interval data. If you have a reasonable sample size there should not be convergence problems that can't be resolved by improving starting values and so on.


1) Would the analysis need to be done for each year separately since the population would be considered open over a two year period?
no you should be able to combine both years into one model, this is what gdistsamp is good for (it fits a temporary emigration type of model). But you have 4 total periods (2 within a year) so it's unclear how to properly structure the multi-scale sampling. You might think about stacking the replicate samples within a year or among years or pooling the data from both samples within a year. There is no right answer here you'll have to think about it and try different things.

2)  Will there be estimation problems?  There are only two distance bins and two visits per year.  These are the bare minimum required for estimation (probability of detection and availability) and the counts per bin per visit are small.
I don't see any reason to expect estimation problems if you have sufficient data. If you have too small a data set then you could always have problems no matter what you do.

3) Can the model be fit for all 26 sites at one time or would it need to be fit separately for each of the sites?
I would recommend analyzing the 52 different point count locations as one big data set.

4) Does the amount of time per site (5 minutes), play a role in the model?
not as long as it was the same at all sites.


regards
andy




--
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unmarked+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Russell Stocker

unread,
Jan 27, 2015, 11:32:49 AM1/27/15
to unma...@googlegroups.com
Andy,

Thanks for your response.  I wanted to followup with a couple of additional questions.

1) I constructed the data frame, used the summary function, and fit the model with the code given below. Does it look like the "sample size" is large enough?  Even when I get convergence, it takes a very long time when using optim  (is there lots of looping going on?)  I find the half-normal distance function to be especially difficult to get convergence.

umf<-unmarkedFrameGDSy=y,yearlySiteCovs=vars,numPrimary=2,dist.breaks=c(0,50,100),survey="point",unitsIn="m")
summary(umf)

unmarkedFrame Object

446 sites
Maximum number of observations per site: 4
Mean number of observations per site: 4
Number of primary survey periods: 2
Number of secondary survey periods: 1
Sites with at least one detection: 398

Tabulation of y observations:
   0    1        2       3    4 <NA>
 909  608  244   19    4    0

Yearly-site-level covariates:
   wind       sky          surveyor 
 0   :338   0   :255     1   :495 
 1   :261   1   :162     2   :132 
 2   :184   2   :312     3   :208 
 3   : 86    4   : 97      4   : 42 
 4   :  6     5   : 50      NA's: 15 
 5   :  1     8   :  1            
 NA's: 16   NA's: 15      

m1<-gdistsamp(~1,~1,~wind+sky,data=umf,keyfun="hazard")

 2) When you talk about stacking, do you mean that we treat the replicates as new sites?  If so then wouldn't this violate the assumption of independence?

Jeffrey Royle

unread,
Jan 27, 2015, 2:44:00 PM1/27/15
to unma...@googlegroups.com
hi Russ,

 your data set looks sufficient to me, I think you should have no trouble fitting models to this.
 the gdistsamp function does take a long time because it's marginalizing over the prior distribution for N by doing the summation (whereas the distsamp function takes advantage of an analytic result to avoid having to do that).

 (2) right, by "stacking"  I mean treating replicates as new sites. I agree this would probably lead to ignoring some dependence but I don't think that is really critical. (might be worth doing a small simulation study to see the effects)

regards
andy

Russell Stocker

unread,
Jan 28, 2015, 1:20:30 PM1/28/15
to unma...@googlegroups.com
Andy,

Thanks for all your help.  I had just a couple of last questions regarding covariates.

1) When modeling the probability of detection, the researchers I am working with are interesting in using time (time of day) as a covariate.  I have seen this done before by other researchers using unmarked, but I question if it makes sense to believe that the relationship between time and the probability of detection would be linear.  Is it typical to use time as a continuous covariate in these types of models?   

2) There is also interest in using the Julian calendar date as a covariate.  Would we want to set this equal to the number of days since the start of the first year of sampling (therefore the second year of sampling would have this covariate be larger than 365) or should we set this equal to the number of days since the start of the current year (reset it to 1 for the start of the second year of sampling)?

Thanks Again,

Russ

Kery Marc

unread,
Jan 28, 2015, 2:37:10 PM1/28/15
to unma...@googlegroups.com
Dear Russ,

re. your questions, it is best if you think of a hierarchical model as the hierarchical distance sampling model as simply a complex kind of generalized linear model (GLM): inside, there is always some linear modeling of covariate effects happening. So (almost .... exactly ....) everything is *just* like with lm() or glm() when you want to express variability in some response by variability of some explanatory variables.

Therefore,
(1) if you think it's not linear, then you can add a quadratic term of time of day. Remember that the modeling in a GLM takes place on the link-transformed scale, so while the model is linear mathematically, it is not usually represented by a straight line (for a continuous covariate). --- You might also break a continuous covariate into classes, effectively making it a factor, to accommodate possible non-linearity.
(2) Usually you would think that, say, April 1 in year 1 is pretty similar to April 1 in year 2; hence, I would start numbering anew.

Kind regards  --- Marc



From: unma...@googlegroups.com [unma...@googlegroups.com] on behalf of Russell Stocker [rsto...@gmail.com]
Sent: 28 January 2015 19:20
To: unma...@googlegroups.com
Subject: Re: [unmarked] Analysis Questions Regarding an Avian Abundance Study

Russell Stocker

unread,
Jan 29, 2015, 10:16:46 AM1/29/15
to unma...@googlegroups.com
Marc,

Thanks for your reply. 

I guess my biggest concern is the ability to check the functional form of the covariates. Having the time covariate included as a polynomial may not be adequate.  I wasn't sure if there were any residual plots that would help to assess this. It would also be interesting to see if plots could be constructed to check the overall fit of the model. 

Russ

Kery Marc

unread,
Jan 29, 2015, 10:33:24 AM1/29/15
to unma...@googlegroups.com
Dear Russ,

yes, you can look at the residuals which are given in a table of site by distance class, I guess. I think that we should look more at residuals also for these hierarchical models.

Of course, you could also run a parboot GOF test and if your data passes that, assume that it's fine, or do more model selection and then let AIC tell you the best structures.

Kind regards  --  Marc


Sent: 29 January 2015 16:16

Russell Stocker

unread,
Feb 2, 2015, 10:56:00 PM2/2/15
to unma...@googlegroups.com
Hi,

After fitting the model, I would like to estimate the abundance for a certain area whose point transects are a subset of the entire sample.  So far I have the following:

m1<-gdistsamp(lambdaformula = ~1, phiformula = ~1, pformula = ~time +  julian+sky+wind, data = umf, keyfun = "hazard")
re<-ranef(m1)
cis<-confint(re)

I believe this produces 95% confidence intervals for each individual point transect.  How would I construct a 95% confidence interval for the sum of the abundances for the first 10 sites?  Would we need an estimate of the posterior distribution of the sum and then use the quantiles to get the correct interval?  Should these be called confidence or credible intervals?

Thanks,

Russ

Kery Marc

unread,
Feb 3, 2015, 4:17:09 AM2/3/15
to unma...@googlegroups.com
Dear Russ,

I think you should bootstrap to get that uncertainty interval.
(1) select a sample of size n from your data set of size (i.e., number of sites) n, WITH REPLACEMENT, fit the model, do any model selection you deem necessary, come up with the estimate of the total N for the first 10 sites and save it
(2) repeat this a large number of times, e.g., 1000
(3) The desired 95% interval is given by the appropriate percentiles

I think this would be called a frequentist interval, i.e., CI.

Kind regards  ---  Marc

______________________________________________________________
 
Marc Kéry
Tel. ++41 41 462 97 93
marc...@vogelwarte.ch
www.vogelwarte.ch
 
Swiss Ornithological Institute | Seerose 1 | CH-6204 Sempach | Switzerland
______________________________________________________________
 
*** Introduction to Bayesian statistical modeling: Kéry (2010), Introduction to WinBUGS for Ecologists, Academic Press; see www.mbr-pwrc.usgs.gov/pubanalysis/kerybook 
*** Book on Bayesian statistical modeling: Kéry & Schaub (2012), Bayesian Population Analysis using WinBUGS, Academic Press; see
www.vogelwarte.ch/bpa
*** Upcoming workshops: www.phidot.org/forum/viewforum.php?f=8


Sent: 03 February 2015 04:56

Russell Stocker

unread,
Feb 9, 2015, 11:11:10 PM2/9/15
to unma...@googlegroups.com
Hello,

I want to thank everyone for their helpful responses.  After talking to my collaborator about the analysis results, I had a couple of more questions regarding the models.   As a reminder, the data was collected using a distance sampling design where each point transect was visited twice a year and there was a total of two years.

Questions:

1) The model being fit using the gdistsamp function is for multiple year data where each point transect is visited once per year.  What is it about the model that makes it inappropriate to use multiple year data where each point transect is visited multiple times within a year, but it is appropriate to use for multiple year data where each point transect is visited only once per year?  What assumptions are being violated?  It seems that there is temporal correlation in each situation.

2) If we where to use the pcountOpen function for the entire data set and ignore the distance information then would the model associated with pcountOpen be valid or is there something inherit in the distance sampling design that when the distance information is ignored leads to the assumptions for the model used in the pcountOpen function being violated?

3) Would it be more appropriate it for us to use the model proposed in the paper entitled "An open population hierarchical distance sampling model" by Sollmann et al. (2015) that is to appear in Ecology.  Andy Royle and Richard Chandler are co-authors and this paper seems to have a similar feel to that of Dail and Madsen (2011).

Thanks again for everyone's help,

Russ
Reply all
Reply to author
Forward
0 new messages