How to deal with multiple samples within sites.

181 views
Skip to first unread message

Don Segundo

unread,
Sep 22, 2015, 3:50:29 PM9/22/15
to unmarked
Hello All,

I'm trying to obtain density estimates with detection probabilities from  point count data using Unmarked however I cant seem to find how to deal with multiple survey locations within sites.  Oddly, all the examples used in the webinar and elsewhere seem to be designed with only one survey location per site. My sampling design includes 221 randomized point count locations at 185 sites.  To avoid psuedoreplication I cant consider multiple survey locations within sites as independent. Can Unmarked deal with this type of study design and if so can you reference where I can find some sample coding?

Don

Jeffrey Royle

unread,
Sep 22, 2015, 9:02:10 PM9/22/15
to unma...@googlegroups.com
hi Don,
 
 What is the difference between a site and a point count location?  If a site has 2 point count locations in it and you treat them as independent then why not just call this "2 sites"?

 If the variation among sites is only due to fixed effects like regression variables then you can model that directly using any of the functions in unmarked. 
 If you're interested in modeling random site effects to account for the grouping of non-independent points within a site, then you probably have to use BUGS to fit such models (examples of this are given in the forthcoming book Applied Hierarchical Models by Marc Kery and myself).

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.

Don Segundo

unread,
Sep 23, 2015, 8:39:02 PM9/23/15
to unmarked
Andy,

I think you misunderstood, I have multiple survey locations at each site and so each survey location within a site is not independent from any others at that site.  I cant treat them as independent because they are spatially auto correlated and so I don't want unmarked to generate density estimates for each point count location rather I want estimates for each individual site.

Based on what Ive seen on the posts among this group it appears that commenters regularly suggest stacking to deal with this issue, i.e. considering each survey location as an independent point which may work depending on the type of questions your trying to answer but if your asking broader scale questions or questions related to discrete patches of habitat stacking does not seem like an appropriate way to deal with this.

Please consider adding a way to account for this in future versions of unmarked because it's value would be greatly improved for many researchers.

Don

Dan Linden

unread,
Sep 24, 2015, 12:17:49 AM9/24/15
to unmarked
Hi Don,

What are the general objectives of the study?  How are you defining sites/surveys, etc.?  What kind of detection data do you have?  It is certainly best to design your sampling with the intended modeling framework and study objectives in mind.  If it is too late for that, sometimes pragmatic decisions are necessary that require simplification, making or potentially violating some assumptions, or discarding some data.

With 185 sites and 221 point counts, you have a very small proportion of sites with point count replicates.  What is the purpose of the replicates?  Depending on the complexity of your model and the covariates included, your inferences using 221 independent point counts (despite violations of independence) vs. 185 independent point counts (1 per site, with some counts discarded) are not going to be all that different.  You should try fitting both the original data assuming independence and the reduced data with 1 point count per site and report back with what you find.

Don Segundo

unread,
Sep 24, 2015, 3:47:17 PM9/24/15
to unmarked
Hello Dan,

The objective of my study is to evaluate the population size and what landscape features influence the density of conservation concern birds using  restored habitat patches.  My detection data includes multiple visits to each point count location where I conducted surveys within and between years. All samples were randomly chosen within the restored patches and patches above a certain size had more than 1 sample location chosen to try to get more accurate density estimates at those patches.

I'm aware of choosing the design to fit the modeling framework and pragmatic decisions, etc. I did not design my study with this particular modelling framework in mind, I intended to use program distance to estimate densities and use those densities in a traditional glm frame work however Im tired of using program distance and wanted to give n-mixture models a try.  In addition, I am also interested in analyzing questions related to species richness  (i.e.Zipkin et al. 2010).  Your right that in my case the number of samples with multiple point count replicates is relatively small and it might not make a difference.  However, in many other studies where having multiple point count replicates (or other types of surveys) within sites is important to get accurate data having a way to deal with this in unmarked could only make it better. I'm assuming there is something about the underlying statistics of n-mixture modelling that make adding this feature difficult however I doubt that it is impossible.

I will try to fit both models and see what the differences are between them. 

Thanks,

Don

Dan Linden

unread,
Sep 24, 2015, 11:02:50 PM9/24/15
to unmarked
Well, the problem is that random site effects would need to be integrated out of the likelihood when fitting the models using MLE, and unmarked models are already dealing with integration of latent variables (true occurrence or abundance), so trying to account for additional unstructured variation would be tricky.  A ton of work for little gain, under the types of data scenarios that are most common.  It is certainly not impossible - see the "lme4" package or Program MARK.  But it is not a straightforward problem that can simply be added.  If you really care about multiple dimensions of random effects, a Bayesian approach is usually necessary (as used in Zipkin et al. 2010).

Also, sometimes a little pseudoreplication is not so bad.  By including the replicates as independent points, you could actually have a better representation of the uncertainty for density relationships with covariates.  Un-modeled heterogeneity is the concern, but again, it's just a model!

Jeffrey Royle

unread,
Sep 24, 2015, 11:11:20 PM9/24/15
to unma...@googlegroups.com
For what it's worth, we have a simple example of computing a Poisson/lognormal type of N-mixture model on the Royle and Dorazio (2008) book (p. 147, panel 4.5, see also equation 4.4.5). It would be very easy to modify this to fit a normal random effect that is shared for 2 or more sites to account for the non-independence.


On Thu, Sep 24, 2015 at 11:02 PM, Dan Linden <danl...@gmail.com> wrote:
Well, the problem is that random site effects would need to be integrated out of the likelihood when fitting the models using MLE, and unmarked models are already dealing with integration of latent variables (true occurrence or abundance), so trying to account for additional unstructured variation would be tricky.  A ton of work for little gain, under the types of data scenarios that are most common.  It is certainly not impossible - see the "lme4" package or Program MARK.  But it is not a straightforward problem that can simply be added.  If you really care about multiple dimensions of random effects, a Bayesian approach is usually necessary (as used in Zipkin et al. 2010).

Also, sometimes a little pseudoreplication is not so bad.  By including the replicates as independent points, you could actually have a better representation of the uncertainty for density relationships with covariates.  Un-modeled heterogeneity is the concern, but again, it's just a model!

--
Reply all
Reply to author
Forward
0 new messages