ovaskain
commented
4 years ago Hi,
A model such as: Dens ~ trait1 + trait2+ env1^2 + ... etc.
Just to clarify that HMSC is a hierarchical model where the first layer is how species occurrences depend on covariates, e.g. Dens~envi^2 + .... The second layer asks how these responses depend on traits, so e.g. response to env^2 ~trait1 + trait2.
1.
I am only interested in predicting density where a species is present, so the distribution maps are only used as a "mask" to avoid predicting over a global map. However I am not sure if the community component offered by jSDM is important because in my case NPP is related to species richness, so by adding NPP as a variable I already capture that aspect. How are the biotic interactions between species captured in this case?Such masking is similar to the hurdle model, having first presence-absence model for occurrence, and then abundance model conditional on presence. I have generally found this strategy meaningful for HMSC modelling. Adding a spatial community-level random effect (which relates to biotic interactions) helps making predictions because it allows borrowing information from other species.
1.
Sample size for single species can be quite low (<10 locations/grid cells) compared to a distribution of +500 grid cells. How well does HMSC work with small sample size? And is it a good idea to pool similar species using taxonomy or phylogeny or traits?In the Norberg et al. (2019) comparison (a spatial version of) HMSC was found to have generally the best predictive performance for typically community data with many rare species. But of course which approach is best depends on the data in hand, and especially how the analyses are set up (which can be done in many ways e.g. within HMSC). In my view it is not generally a good idea to pools similar species, but rather to include species traits and phylogenies so that the estimation can borrow from related species (if there is a signal in the data about e.g. related species having similar responses).
fabeit
Thanks to the developers for this useful package and scientifically important contribution. I would like to better understand if HMSC is the right tool for my task.
Briefly, I am trying to create maps of animal species density ( individuals/km2) from distribution maps (presence/absence) and density estimates. The covariates will be species traits (body mass, diet, etc.) and environmental variables ( NPP, temperature, etc. ). A model such as: Dens ~ trait1 + trait2+ env1^2 + ... etc. A few more details:
I am only interested in predicting density where a species is present, so the distribution maps are only used as a "mask" to avoid predicting over a global map. However I am not sure if the community component offered by jSDM is important because in my case NPP is related to species richness, so by adding NPP as a variable I already capture that aspect. How are the biotic interactions between species captured in this case?
Sample size for single species can be quite low (<10 locations/grid cells) compared to a distribution of +500 grid cells. How well does HMSC work with small sample size? And is it a good idea to pool similar species using taxonomy or phylogeny or traits?
Some of these questions are rather complicated to answer but would appreciate any general feedback on the feasibility of HMSC to solve this task. Thanks in advance.