Overdispersed Cameratrap Data

[Carolin Klangwald]

Hello everyone, 
I`m a student working towards my Masters Thesis and new to the Distance software. 
I`m trying to analyse camera trap data, which holds some overdispersion. Because of this I am not quite sure how to judge the goodness of fit. I`ve run several different analyses and the Halfnormal-Cosine with truncation seems to work well, and using the adjustment for a Poisson distribution with an overdispersion factor of 2 did result in improved accuracy. However I`m just not sure if I am handeling this appropriately. Does anyone have some suggestions for me or maybe could direct me to some sources? 

Thank you for any feedback, as I`m currently working on my own from home I appreciate any exchange :) 

[Eric Rexstad]

Carolin

You want to assess the fit of your detection function model to the observed detection distances.  If you are using Distance for Windows (I assume that is what you are using), fit can be assessed using the chi-square GOF test (assuming the distances are binned, as is common with camera trapping data):

If using the R package Distance, fit is assessed using the function `gof_ds()`.

The overdispersion addressed with the Poisson distribution is NOT the type of overdispersion associated with camera trapping data

This is relevant to variability in encounter rates between camera stations which relates to spatial clustering of animals through your study area.

Have a read through the paper by Howe et al. (2019) for matters of overdispersion as it relates to camera traps:

Howe, E. J., Buckland, S. T., Després‐Einspenner, M.-L., & Kühl, H. S. (2019). Model selection with overdispersed distance sampling data. Methods in Ecology and Evolution, 10(1), 38–47. doi: 10.1111/2041-210X.13082

and for an implementation of the model selection algorithm discussed in that paper, visit the case study on our website:

http://examples.distancesampling.org/Distance-cameratraps/camera-distill.html#model-selection-adjustments-from-overdispersion

Thank you for any feedback, as I`m currently working on my own from home I appreciate any exchange :)  --
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/c0fc8893-d1db-4d25-9024-9ae3786871b6n%40googlegroups.com.

-- 
Eric Rexstad
Centre for Ecological and Environmental Modelling
University of St Andrews
St Andrews is a charity registered in Scotland SC013532

[PABLO PALENCIA MAYORDOMO]

Really nice reply, Eric!

I come now with another question related to spatial aggregation of animals and overdispersion (in encounter rate).

When using Distance for Windows, how to deal in those situations in which we found overdispersion in encounter rate (between lines or points)?

Using the "unmarked" R package is possible to include negative binomial instead of Poission.  But, I think, that this is not possible in Distance for Windows. Buckland et al 2001  (section 3.6.1) suggested multiplication of the Poisson variance by two if no other approach for estimating variance was available.

Additionally, Buckland et al 2001  (section 4.4), recommended empirical estimation of the sampling variance of n,  as var(n) > E(n) is expected.

Are there other options included in the software? what is the approach that you recommend?

Thank you very much

Best

Pablo

[Eric Rexstad]

Pablo

There is a difference in the way Distance for Windows (and the R package Distance) deals with variability in encounter rate between transects, and the way this is handled using the 'unmarked' software.

The methods developed at St Andrews do not try to model transect-level encounters, so there is not a model (Poisson, negative binomial, ...) involved.  There is no concept of overdispersion in the absence of a model.  As you note in quoting Section 3.6.1 of Buckland et al. (2001), resorting to the Poisson approximation is only done if no other approach is available (see Section 4.7.1).  We recommend, as you note, that an empirical estimate of encounter rate variance be employed.  The veracity of this empirical estimate is strongly influenced by a sufficient number of spatial replicates in the form of transects.  A resampling approach (bootstrap) is usually employed to estimate uncertainty in abundance or density estimates, with the transect being the resampling unit.  Consequently, Buckland et al. (2001) Section 7.2.1 suggested 10-20 (or more) replicate transects be used for a reliable estimate of encounter rate variance.

If you want a more statistical treatment on the subject of estimating encounter rate variance, consult this Biometrics paper:

Fewster, R. M., Buckland, S. T., Burnham, K. P., Borchers, D. L., Jupp, P. E., Laake, J. L., & Thomas., L. (2009). Estimating the encounter rate variance in distance sampling. Biometrics, 65, 225–236. doi: 10.1111/j.1541-0420.2008.01018.x

The "R2" estimator (for line transects) from this paper is used as the default in Distance for Windows.  See Chapter 8 of the Distance for Windows users guide "Advanced analytical variance estimation in CDS"

To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/d6d8c3e0-1174-4de9-b5b9-ea8c09de169an%40googlegroups.com.

[PABLO PALENCIA MAYORDOMO]

Thank you very much for the clarification Eric!

Now I understand :)