Statistical comparison of additive regression tree methods on ecological grassland data

Abstract

Additive tree methods are widely used in ecology. To date most ecologists have used boosted regression tree (BRT) methods. However, Bayesian additive regression tree (BART) models may offer advantages to ecologists previously unexamined.Here we test whether BART has some benefits over the widely used BRT method. To do this we use two grassland data and 13 hydroclimatic and land use predictor variables. The dataset contained data from a period of drought as well as during a recovery phase after the drought. The response variable was the trend in the Enhanced Vegetation Index (EVI), which is an remotely sensed indicator of grassland degradation and recovery.The settable parameters of both methods (BRT and BART) were varied to compare the performance of each method. BRT and BART models were evaluated using three prediction error statistics; root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R2). The best models across the two methods were assessed by inspecting the relative importance of predictor variables and the prediction error statistics.BRT and BART models exhibited similar variable selection abilities, but the BART method generated models with similar or more favourable prediction error statistics than the BRT method (BART explained an additional 10.17% to 11.92% of the variation than BRT models). Our results indicate that BARTs may be more effective at modelling ecological data than BRTs.BARTs also had shorter run times, more reasonable defaults in its software implementation, and greater functionality of said software implementation, beyond model building and prediction functions. Ecologists using additive regression approaches may benefit from using BART approaches and we suggest their use alongside more commonly used BRT methods in ecological studies.