Careful prior specification avoids incautious inference for log-Gaussian Cox point processes

TL;DR

This paper emphasizes the importance of careful prior specification in Bayesian log-Gaussian Cox process models, introducing a reparameterized approach with penalized complexity priors to improve interpretability and invariance in ecological spatial analysis.

Contribution

It proposes a reparameterized model with PC priors for hyperparameters, enhancing interpretability and invariance in Bayesian spatial point process inference.

Findings

The model effectively assesses covariate effects on rainforest tree species.

Hyperprior tuning influences inference robustness.

Application demonstrates improved ecological interpretation.

Abstract

Prior specifications for hyperparameters of random fields in Bayesian spatial point process modelling can have a major impact on the statistical inference and the conclusions made. We consider fitting of log-Gaussian Cox processes to spatial point patterns relative to spatial covariate data. From an ecological point of view, an important aim of the analysis is to assess significant associations between the covariates and the point pattern intensity of a given species. This paper introduces the use of a reparameterised model to facilitate meaningful interpretations of the results and how these depend on hyperprior specifications. The model combines a scaled spatially structured field with an unstructured random field, having a common precision parameter. An additional hyperparameter identifies the fraction of variance explained by the spatially structured term and proper scaling makes the analysis invariant to grid resolution. The hyperparameters are assigned penalised complexity priors, which can be tuned intuitively by user-defined scaling parameters. We illustrate the approach analysing covariate effects on point patterns formed by two rainforest tree species in a study plot on Barro Colorado Island, Panama.