Peruzzi M (2024) Inside-out cross-covariance for spatial multivariate data.
https://arxiv.org/abs/2412.12407
When the number of spatial variables is not tiny, the most popular and common model for multivariate analyses is the linear model of coregionalization (LMC) aka spatial factor model. The LMC is tractable, leads to structured sample covariance matrices, and makes some things easy. But also, it’s not a great model in several important ways.
In this article, I introduce Inside-Out Cross-covariance (IOX), which is also tractable, also leads to structured sample covariance matrices, and also makes some things easy. All while resolving a number major issues of LMCs:
- IOX is constructed based on a list of correlation functions (like LMCs)
- These are also the marginal correlations for each outcome (not like LMCs!)
- All the outcome-specific features go through the marginal correlations (not like LMCs!)
- Easy to model outcomes with different smoothness, range, nonstationarity (not like LMCs!)
- Easy to set priors for covariance parameters (not like LMCs!)
- Can do dimension reduction, factor modeling, scalable computations (like LMCs)
R package spiox on github. Code to reproduce paper.
