Beyond correlation in spatial statistics modeling

TL;DR

This paper presents a novel spatial statistics model that explicitly captures interactions among multiple field components, highlighting the importance of considering multivariate interactions for accurate spatial inference and forecasting.

Contribution

It introduces a new multivariate interaction model for spatial data, addressing theoretical properties and demonstrating the impact on inference and precipitation forecasting.

Findings

Ignoring multivariate interactions can lead to very wrong inferences.

Additional validation statistics help assess interdependence among variables.

Considering multivariate interactions improves precipitation forecast accuracy.

Abstract

We introduce a model for spatial statistics which can account explicitly for interactions among more than two field components at a time. The theoretical aspects of the model are dealt with: cumulant and moment generating functions, spatial consistency and parameter estimation. On the basis of a detailed synthetic example, we show the kind of inference about the (partially observed) spatial field that can be very wrong, if one validates his model by checking only one and two dimensional marginal fit, and covariance function fit. We suggest statistics that can be used additionally for model validation, which help assess interdependence among groups of variables. The implications of considering multivariate interactions for intense daily precipitation forecasting over a small catchment in southeastern Germany (that of the Saalach river) are investigated.