Royle/Nichols model for camera trap data

[matobler]

I am currently analyzing camera trap data for five species of ungulates from four 60-day camera trap surveys, three in study area A and one in study area B. The total number of sites surveyed (for both areas and all years combined) is 147 with 10 survey periods (each lasting 6 days) for each site. There are two different habitat types in both study areas. I use the following possible covariates: habitat, study area, survey (area A year 1, area A year 2, area A year 3, area B year 1). I am less interested in the absolute occupancy values than the effect of the covariates.

Running some first analysis the results look reasonable. For four out of five species the Royle/Nichols models performs much better than the single species, single season models, usually with a deltaACI of >20. I attribute this to the large amount of heterogeneity present in the data (Mh usually is the model of choice for capture-recapture studies with camera traps). Standard error and confidence intervals are relatively large but that seems to be normal for the Royle/Nichols model from what I have read. The results from the Royle/Nichols model would actually answer my original question better if I interpret the lambda as an index for relative abundance (the mean number of individuals/site should be linearly related to density). So if ACI shows a high support for a model with a habitat covariate for lambda and lambda for habitat A is 1.5 and for habitat B is 3.2 then I would conclude that the species shows a clear preference for habitat B (on a population level) and is more abundant in that habitat.

Since I have never seen this kind of analysis applied to camera trap data I wanted to ask if anyone sees any potential problems with using the Royle/Nichols model and if my interpretation of the results seems correct for the question asked.

Thanks,

Mathias

[darryl]

Personally I think you have to be pretty cautious when interpreting lamda in terms of abundance (relative or absolute) as all the information about 'abundance' comes from the observed level of heterogeneity among sites in the probability of detecting the species. There may be factors other than abundance that are contributing to the heterogeneity, but because lambda is the only 'outlet' for heterogeneity in p in the model then it may naively appear that abundance is different. Have you tried including covariates for detection using the non-RN method?

Darryl

[matobler]

I share your concern about other factors contributing to the heterogeneity (animal behavior, camera placement, camera malfunctioning etc.). I share your concern about other factors contributing to the heterogeneity (animal behavior, camera placement, camera malfunctioning etc.). I tried to run the single season model with the different covariates both for psi and p. In many cases the selected covariates are very similar to the ones selected by the RN model. However in some cases the interpretation of the results is somewhat complicated. I would like to illustrate that on an example: one of the species I am working with (white-lippec peccary) is a wide-ranging species so there are many cameras within each groups home range. I am therefore looking at use and not occupancy. Now looking at the capture frequencies this species seems to prefere one habitat but uses both. The single season model would select a model with p(Habitat) over a model with psi(Habitat). This does mak sense since the speces uses both habitats but spends more time in one habitat and is therefore more likely detected there. I belief that unlike for studies where animals are detected visually by an observer, for camera traps set on trails habitat should not have a large impact on detectability except throug a difference in use or abundance. The RN model select the lambda(Habitat) over the r(Habitat) model.

Is there any software that can run contiuous mixture models for occupancy as described in the occupancy book (not abundance models) or is there a reason those were not implemented in Presence? Since the finite mixture models in Presence can't use covariates they are not of much help for my problem. I noticed however, that for several of my datasets the 2 Group, Constant P model performs much better than the best non-mixture model with covariates (with deltaACI up to 20) indicating that there is a lot of heterogeneity in the data.

Thanks for the advice,

Mathias

[darryl]

The lamda(habitat) model may be getting selected for the same reasons as the p(habitat) model, because the individuals are spending more time in the habitat rather than more individuals in that habitat. Recall that when you have 'use' rather than 'occupancy', then detectability is made of 2 components: Pr(in unit at time survey)*Pr(detected given in unit).

There's no software that I know of, but Andy Royle may have R code to do this. Otherwise there's WinBUGS if you want to go down that track.

Darryl