Spatially Correlated, Approximate Modelling of Point patterns in R
scampr is an R package that offers a regression-style framework for modelling spatial point patterns (referred to as presence-only data in ecology) using a log-Gaussian Cox Process (LGCP). The package can also be used to fit a binomial model with spatial random effects to binary data (e.g. presence/absence data in ecology), as well as an integrated model that combines the two while permitting shared spatially correlated latent field(s).
Unlike the inhomogeneous Poisson process, LGCPs offer a way to incorporate additional spatial clustering into models by including a Gaussian random field (GRF) to induce additional spatial correlation between observations — effectively acting as a spatially correlated error term. LGCP models are particularly appropriate in instances where clustering arises from missing or unmeasured environmental predictors/phenomena, as opposed to those in which clustering/dispersal is due to interactions between point events.
Fitting LGCP models can be difficult and time consuming and, as a result, limits the ability of researchers to flexibly analyse spatial point pattern data. scampr is a fast, approximate maximum-likelihood approach to fitting LGCP to 2D spatial point patterns, involving a combination of three innovations. First, variational approximation (VA) permits a closed form approximation to the marginalised log-likelihood. Second, fixed rank kriging provides a rank reduced approximation to the large spatial variance-covariance matrices that arise and are otherwise very computationally demanding. Finally, automatic differentiation is used to quickly obtain gradient information for efficient optimization and inference.
Models are fit using scampr() that adopts a simple interface/syntax following the common regression modelling formats used on R, e.g. lm(), glm(). Many of the common S3
functions familiar to users (such as summary, plot, simulate, predict, logLik, AIC, confint and vcov) are also available to scampr models. The package is built upon the
advances of TMB (Kristensen et al., 2016) which enables coding the likelihoods in C++, as well as providing automatic differentiation for easy access to gradient
information — permitting fast optimisation, automated Laplace approximation, and automated estimation of the variance-covariance matrix of parameter estimates in scampr
models.
The package name stands for Spatially Correlated, Approximate Modelling of Point patterns in R however, the verb "scamper" — to run with quick, light steps — perfectly captures the motivation of this package: to give researchers access to complex spatial models that fit quickly and require only a light touch.
As scampr is not on CRAN please install via:
devtools::install_github("ElliotDovers/scampr", dependencies = "Imports", upgrade = "never")