Model selection for geostatistical models

Area of Science:

• Geospatial statistics
• Environmental modeling
• Statistical ecology

Background:

• Spatial correlation is frequently overlooked in geospatial data analysis, potentially leading to inaccurate variable selection.
• Ignoring spatial dependencies can mask the true importance of explanatory variables in statistical models.
• Traditional model selection methods may not adequately account for the unique characteristics of geospatial datasets.

Purpose of the Study:

• To investigate the impact of spatial correlation on model selection for geospatial data.
• To propose and evaluate the Akaike Information Criterion (AIC) for geostatistical model selection.
• To compare geostatistical AIC-based selection with traditional approaches that ignore spatial structure.

Main Methods:

• Heuristic derivation of the Akaike Information Criterion (AIC) for geostatistical models.
• Simulation studies to assess the performance of AIC in geostatistical variable selection.
• Application of AIC and Minimum Description Length (MDL) principles to a lizard abundance dataset.
• Exploration of sampling design effects on covariate selection in geostatistical models.

Main Results:

• Geostatistical AIC-based model selection significantly outperforms traditional methods that ignore spatial correlation.
• The importance of explanatory variables is more accurately identified when spatial correlation is considered.
• The Minimum Description Length (MDL) principle also shows promise for geostatistical variable selection.
• Sampling design influences the selection of explanatory covariates.

Conclusions:

• The Akaike Information Criterion (AIC) is a superior tool for model selection in geostatistical analyses.
• Accounting for spatial correlation is crucial for accurate variable importance assessment in geospatial data.
• The developed methods and provided R software facilitate improved geospatial model selection.