Where are the critical habitats for a species, given our knowledge of the important environments and the communities in which it lives?
Can we interpret it on a transparent scale, e.g. the one we would observe by standard methods?
How well can we know it? The uncertainty?
PBGJAM offers abundance-weighted habitat scores (AWHS) for current conditions and future scenarios. It is important not to mis-interpret a map from these pages as a prediction of where a species is now or where a species will be in the future. Rather, they are habitats weighted by current abundance of species, based on variables used in model fitting and other species in the community. They are conditioned on a specific scenario (e.g., ‘current conditions’ or an emissions scenario).
AWHS transfer formal predictive distributions from Generalized Joint Attribute Modeling (GJAM), with full uncertainty from parameters, model, and observations to mapped scenarios. For this reason, it is fine to refer to them as “predictive distributions”—again, that’s the formal name for them. Unlike standard predictions, AWHS are based on abundance, incorporating sample effort for observed data. They are not presence-only or presence/absence. The model incorporates the joint relationships between species, because there is mutual dependence between species. The units (dimension) of a AWHS is species abundance—the same units used for observations.
Still, the AWHS is not a prediction of where the species is or will be, because it cannot incorporate all of the variables (and interactions between them) that control the complexity of species distribution and abundance in the real world, including how populations move through landscapes as conditions change. For this reason, maps are accompanied by the sensitivity to variables used in model fitting (this is GJAM’s sensitivity plot). These variables were selected based on accumulated knowledge of ecologists and on fit to the data. The maps of AWHS are a habitat score, based on variables in the model.
AWHS offers a unique perspective on habitat suitability that incorporates species abundance, the full community dependence, with full uncertainty, in transparent (interpretable) units—the same as the observations themselves.
Here are some common questions:
“Species distribution models use presence-only data to produce habitat suitability based on environmental layers at species localities too! It also does climate projections.”
Due to the fact that most species are cryptic, understanding distribution and abundance relies heavily on estimates of uncertainty. Models of presence-only data do not produce species distribution maps that have valid probability interpretations. One of the most recent treatments of this problem is here. GJAM is based on a likelihood for the observations that fully accommodates sample effort and the joint relationships between species that leads to probability-based uncertainty.
“Why is a joint distribution of species important?”
There are at least two reasons. First, the sensitivity of species presence-absence or abundance to climate, soils, or land use is a product of the analysis. These sensitivity coefficients only have valid uncertainty estimates if they allow for the fact that species depend on one another.
Second, the joint distribution improves our ability to ‘conditionally predict’ the abundances of species based on knowledge of others.
“There are now several joint species distribution models, what makes GJAM different?”
There are at least two important innovations with GJAM that allow us to quantify abundance. First, it allows us to analyze communities containing species measured in different ways. Alternative models are limited to a specific species group, because they depend on uniformity in data collection. Examples include trees on inventory plots, birds from point counts, or insects from pitfall traps. In each case, there are multiple species that can be analyzed jointly, because all species are measured in the same way. However, real communities do not consist of a single species group—birds interact with plants and insects. GJAM allows us to jointly model species measured on different scales.
Second, GJAM provides estimates on the observation scale. Generalized linear models (GLMs) require a non-linear link function to accommodate zeros in the data. This results in estimates that do not have transparent interpretation. GJAM avoids non-linear transformations, thus providing estimates on the scales that observations are made (e.g., plot area, number of trap nights, …).
“How can a suitable habitat be predicted far from locations where a species has been observed.”
The concept of ‘suitable habitat’ from an analysis like this is always qualified by the variables used as predictors in the model. We use predictor variables that are known to affect distribution and abundance (e.g., climate, soils, land cover). However, there are many important variables that cannot be measured. These include fine-scale microclimate and soil data, trophic interactions (e.g., competitive exclusion), and biogeographic history. A species may be absent from ‘suitable habitat’. Our analysis helps identify where those locations are that could be suitable based on the variables that could be quantified. For example, the barred owl that has recently invaded the Pacific Northwest and is out-competing the endangered northern spotted owl.