sjSDM - Scalable Joint Species Distribution Modeling
A scalable and fast method for estimating joint Species
Distribution Models (jSDMs) for big community data, including
eDNA data. The package estimates a full (i.e. non-latent) jSDM
with different response distributions (including the
traditional multivariate probit model). The package can perform
variation partitioning (VP) / ANOVA on the fitted models to
separate the contribution of environmental, spatial, and biotic
associations. In addition, the total R-squared can be further
partitioned per species and site to reveal the internal
metacommunity structure, see Leibold et al. (2022)
<doi:10.1111/oik.08618>. The internal structure can then be
regressed against environmental and spatial distinctiveness,
richness, and traits to analyze metacommunity assembly
processes. The package includes support for accounting for
spatial autocorrelation, for grouping-factor random effects in
the environmental formula, and for deep neural networks, which
can replace the linear predictor or sit beside it as one block
of a formula that also holds interpretable linear terms. As
described in Pichler & Hartig (2021)
<doi:10.1111/2041-210X.13687>, scalability is achieved by using
a Monte Carlo approximation of the joint likelihood implemented
via 'torch', which can be run on CPUs or GPUs.