Boosted Regression Trees for ecological modeling

This is a brief tutorial to accompany a set of functions that we have written to facilitate fitting BRT (boosted regression tree) models in R . This tutorial is a modified version of the tutorial accompaniying Elith, Leathwick and Hastie’s article in Journal of Animal Ecology. It has been adjusted to match the implementation of these functions in the ’dismo’ package. The gbm* functions in the dismo package extend functions in the ’gbm’ package by Greg Ridgeway. The goal of our functions is to make the functions in the ’gbm’ package easier to apply to ecological data, and to enhance interpretation. The tutorial is aimed at helping you to learn the mechanics of how to use the functions and to develop a BRT model in R . It does not explain what a BRT model is for that, see the references at the end of the tutorial, and the documentation of the gbm package. For an example application with similar data as in this tutorial, see Elith et al., 2008. The gbm functions in ’dismo’ are as follows: 1. gbm.step Fits a gbm model to one or more response variables, using cross-validation to estimate the optimal number of trees. This requires use of the utility functions roc, calibration and calc.deviance. 2. gbm.fixed, gbm.holdout Alternative functions for fitting gbm models, implementing options provided in the gbm package. 3. gbm.simplify Code to perform backwards elimination of variables, to drop those that give no evidence of improving predictive performance. 4. gbm.plot Plots the partial dependence of the response on one or more predictors. 5. gbm.plot.fits Plots the fitted values from a gbm object returned by any of the model fitting options. This can give a more reliable guide to the shape of the fitted surface than can be obtained from the individual functions, particularly when predictor variables are correlated and/or samples are unevenly distributed in environmental space. 6. gbm.interactions Tests whether interactions have been detected and modelled, and reports the relative strength of these. Results can be visualised with gbm.perspec