Custom model with spatial correlation for pond survey data

[rajah]

Hello,

I have a survey dataset that involves approximately 150 ponds that were visited a single time in one season. The primary survey method involved the use of dip net sweeps around the pond margin (>10 sweeps per pond). We detected 5 native amphibians as well as two nonnative species. We are interested in addressing the effects of the nonnatives on distribution of the natives (we do not have repeated observations of nonnative presence, however, so I don't believe a multi species model is possible).

I would like to know if it would be appropriate to use the Hines et al. 2010 model that accounts for observations which are not sampled with replacement and are non-independent. It seems that we could treat individual net sweeps as the observations and then use nonnative presence (and several environmental variables) as covariates.

This seems a bit unconventional for pond survey data, but would there be any fundamental reasons to avoid this approach? Given that each site was visited once, and essentially only one survey method was used, I don't see how other types of occupancy models would be appropriate.

Thank you very much for your time.

Sincerely,
Dan

[murray.efford]

Hello Dan
As the experts are quiet, I’ll jump in with some opinions. There are several levels to your question. Firstly, is it OK to use spatial replication (sweeps) rather than temporal replication? The tentative answer is Yes. Secondly, does the autocorrelation model of Hines et al. 2010 help? That is something to determine by empirical model comparison. (There are subsidiary issues I haven’t thought about: were the sweeps spread evenly around the margin, and what is the effect of variation in pond size?). Thirdly (implicitly), is sampling with replacement an issue? Probably not: see Guillera-Arroita 2011 Methods in Ecology & Evolution 2:401 for correction to Kendall & White on this.
I would be concerned that variation in the number of sweeps will generate artefacts (this happened with transect length in some unpublished simulations I made of the Hines et al. scenario). One solution would be to select a constant number of sweeps (10) at random from the total. There may be better solutions. Note that variation in population size will cause heterogeneity in detection so the Royle-Nichols formulation is appropriate. Estimates of occupancy are still likely to be somewhat biased, but what the heck: the covariate effects will probably be meaningful.
Murray

[darryl]

Hi Dan,
Adding to what Murray has suggested. The Hines et al. model was developed where the sampling is linear in space (or also applies for linear time), so there is a natural first and last survey. That may not be the case when you're going around a pond as the designation would be more arbitrary (where's the end of a circle?). What distance was there between sweeps? Essentially the Hines et al. model was developed to account for a lack of independence between surveys so if there is a reasonable distance between sweeps (relative to the size of the sweep), then it may be ok.

Is the non-native species detected perfectly? If not, that may cause problems with your inference, particularly if the detection of non-native depends on any of the environmental covariates you're going to be considering...

Cheers
Darryl