Open nicholasjclark opened 4 months ago
Excellent point here, @AdamCSmithCWS! Yes, I agree we need to account for this, and I like the idea of comparing the accuracy of the two models. Using phylogenetic information to inform the intercepts like we want to do here comes with the expectation that species with close phylogenetic relationships should have similar baseline abundances (of course, after controlling for other factors). Also, species detectability itself may have a phylogenetic signal. E.g., closely related species might share traits that make them more or less detectable. So, including phylogenetic information in our model can help account for these shared traits and potentially improve our predictions (esp. for species with limited data).
In addition, perhaps we can build a multi-level model that includes species-level random effects— influenced by phylogenetic relationships and observer variability (as well as site and year effects).
Great discussion, and I agree with @drhammed here. From a hierarchical model's point of view, phylogenetics and shared traits can affect both the observation model (detectability) and the ecological model (abundance). For example, big butterflies are easily to detect, but maybe they can also cope with climate change better... If we are not moving into this hierarchical approach, the solution, as @drhammed said, is including species-level random effects, but we are not going to be able to disentangle if it is affecting the observation or the abundance.
I am developing a literature review to evaluate the approaches that have been used to leverage phylogenetic or functional relationships to inform trends. However, it has been more frustrating than I first thought since many excellent studies only evaluate it as a post-hoc but do not consider the phylogenies or the functional traits in the process of developing the population estimates. For example. I will try to elaborate a more systematic review to move forward and see how the people approached this problem.
Thanks for the comments all. Yes I agree that detection error will play a role, but given that this isn't typically accounted for in most large-scale monitoring studies I think we'll have to leave that as a lower priority for the big analyses that are planned.
However, that doesn't mean we can't target this question more directly with a smaller-scale set of questions. For example, we could hone in on one bird conservation area and some species of particular interest for which we think species' relationships may be important in these two processes. We could then fit dynamic N-mixture models (which are readily handled by {mvgam
} (see the N-mixture vignette as an example). Since {mvgam
} relies on {mgcv
} to construct smoothing splines, the mrf
basis and all of it's functionality works without modification. We'll have to organize another chat soon or perhaps start expanding on the list of priorities so we can keep track of these ideas.
@drhammed, the current model I'm using includes species-level hierarchical intercepts (using both phylogenetic interecepts and unstructured random effects), so that is taken care of. But let's please leave particulars of model setup for the discussion in issue #5 and keep this thread for understanding what has been done to date in the literature.
From what I recall, the 'Revealing uncertainty in the status of biodiversity change' study was the first to directly embed the phylogeny in the abundance model. I could be wrong though. Quite a few studies did employ a 2 step analysis:
First running a model like this for each species
abundance ~ year + site
The species-level trend from this would then be aggregated in a pgls
When we were writing the paper, we focussed only on looking for phylogenetic signal in the abundance trends, as we were working with lots of compiled datasets without standardised monitoring. Of course this is much less of an issue in data like the NABBS. So you could imagine the actual abundance and abundance change will be informed atleast partially by this phylo structure
Hi @nicholasjclark and @drhammed Has anyone considered how to leverage phylogentic information on abundance (i.e., the phylogenetically informed intercepts) for a multi-species dataset like the BBS, while accounting for variations in detectability among species (detectability == proportion of the true local population that is counted)? I was just looking over the model and thinking about the mean counts on BBS surveys. Given the field methods, these counts can be reliably interpreted as relative-abundance within a single species. That is, the field methods for the BBS (and most modeling approaches) assume that detectability of a given species is effectively constant through time and space, after adjusting for variation among observers. However, this assumption does not hold well across species. I suppose an answer to my question is part of what could come from the comparison of trait-based models to phylogenetic models (e.g., if we compared the accuracy of a phylogenetic approach to modeling mean abundance to an approach based on traits such as body-size and auditory-cue-pitch). So, maybe I just answered my own question :-), but I'll leave this here in case it's helpful.