Comparing spatial regression to random forests for large environmental data sets.

Area of Science:

• Environmental Science
• Ecology
• Statistical Modeling

Background:

• Large environmental datasets present challenges due to numerous records and covariates.
• Random forests excel with many covariates and nonlinear relationships.
• Spatial regression performs well with many records and spatial autocorrelation.

Purpose of the Study:

• Compare random forests and spatial regression for predicting the macroinvertebrate multimetric index (MMI).
• Evaluate model performance for mapping MMI across the conterminous United States.
• Assess prediction intervals and variability between the two methods.

Main Methods:

• Utilized a dataset of MMI at 1859 stream sites with over 200 landscape covariates.
• Developed a novel transformation procedure for spatial regression, including Box-Cox transformations and handling zero-inflated data.
• Employed cross-validation and a simulation study to compare model performance.

Main Results:

• Spatial regression with transformations achieved cross-validation performance comparable to random forests.
• Both methods demonstrated prediction interval coverage close to nominal.
• Spatial regression yielded narrower and less variable prediction intervals compared to quantile regression forests.

Conclusions:

• Spatial regression, enhanced with transformations, is a viable alternative to random forests for large environmental datasets.
• The choice between methods depends on specific data characteristics and the need for narrow prediction intervals.
• The study provides insights for accurate ecological mapping and uncertainty quantification.