meshed R package

meshed, available on CRAN, is the R package which implements:

• Bayesian spatial regression with latent Meshed Gaussian Processes (MGPs)
• for multivariate data, coregionalization/spatial factor models on latent MGPs
• for non-Gaussian outcomes, sampling via the Langevin methods of the SiMPA paper
• gridding and parameter expansion from the GriPS article.

The main function is spmeshed which can do Bayesian spatial regression in any combination of the following:

• data at irregular spatial locations;
• spatiotemporal data;
• multivariate outcomes;
• spatial misalignment of multivariate outcomes;
• spatial or spatiotemporal factor models;
• outcomes of different types (currently: Gaussian, Poisson, Binomial, Beta).

Shiny app for R package meshed

This simple Shiny app generates bivariate spatial data on a grid, then fits a cubic MGP using a Gibbs sampler (if both outputs are chosen Gaussian) or via SiMPA/Langevin if not. This app runs meshed::spmeshed on 8 cores on an AMD Ryzen 9 5950X CPU

• MCMC restarts from scratch each time a parameter is changed; none of the simulation parameters are sent to spmeshed
• expected MCMC runtime is about 10 seconds with n=3000 and 5000 iterations, and about 30 seconds with n=5000 and 10000 MCMC iterations
• error variance = nugget for Gaussian outcomes, scale = overdispersion parameter for sampling Negative Binomial outcomes. Error variance/scale are ignored if not Gaussian or Negative Binomial
• spatial decay parameters are used in a coregionalization model in which each margin has an exponential covariance