Local Estimation of Heterogeneous Treatment Effects on Risk Aversion with Generalized Random Forests

Abstract

This thesis assesses the capabilities of the Generalized Random Forests (GRF) algorithm as an estimator for Conditional Average Treatment Effects (CATEs) on risk aversion. The estimating proficiency of this novel method is quantified in an exponential utility and normally distributed lotteries context by conducting a simulation experiment which reproduces a simplified version of a modified Becker-DeGroot-Marschak auction. The limitations of this approach are explored and the GRF-LML estimator is developed in order to make the GRF method applicable to a broader set of decision models. The assumptions under which this new estimator is consistent and asymptotically normal are derived and the possible weaknesses of the approach explored. Finally, a detailed accounting of the development process within the grf library is presented, as well as an exploration of the technical hurdles one needs to overcome in order to program a custom statistical forest.