Random Forests for Benefit Transfer

Benefit transfer (BT) has evolved as the dominant valuation method for environmental benefit-cost analyses, including those required of US federal agencies. Yet even best-practice approaches for BT based on meta-regression models (MRMs) typically exhibit poor predictive fit and out-of-sample precision. This article introduces random forests (RFs) for nonparametric estimation of MRMs and construction of BT predictions. We compare the performance of different RF models to current best-practice approaches for BT. We find that forest-based models substantially improve the out-of-sample accuracy of welfare predictions and tighten confidence intervals of predicted benefits for stipulated policy scenarios. The best performers reside within the family of local linear forests (LLFs), a hybrid approach that combines elements of RFs and locally weighted regression. Results suggest that this new approach has the potential to substantially improve BT accuracy for environmental policymaking without sacrificing theoretical properties, while simultaneously reducing econometric and computational difficulties relative to leading alternatives.