Random forests for dichotomous choice contingent valuation

We introduce a novel, fully nonparametric estimation framework to process data from survey-based environmental valuation with a binary, referendum-style choice question, traditionally referred to as Contingent Valuation. Our approach combines the construction of choice probabilities via Random Forests (RFs), and a variant known as Local Linear Forest (LLF), with welfare predictions via common distribution-free estimators. While popular as back-of-envelope alternatives to parametric estimation, these distribution-free methods are poorly suited for the incorporation of observation-specific heterogeneity. In contrast, our RF/LLF Non-Parametric approaches, which we label RFNP and LLFNP, respectively, produce willingness-to-pay (WTP) estimates at the individual level, conditioned on a potentially large set of explanatory variables. Using simulated data, we find that both Forest estimators are robust to nonlinearities in the WTP function and can compete with correctly specified parametric models in terms of asymptotic efficiency. In our empirical application within the context of biodiversity enhancements on open land in the United Kingdom, we show that the RFNP is immune to negative WTP predictions by construction, and produces reasonable and efficient lower bound estimates for individual and sample-aggregated WTP. It can also generate unbounded welfare predictions that allow for long tails in individual WTP, without having to impose this feature on all observations. Our framework is well-suited for numerous extension, and is readily implemented with existing software packages.