Random Forest Adjustment for Approximate Bayesian Computation

We propose a novel method for regression adjustment in approximate Bayesian computation to help improve the accuracy and computational efficiency of the posterior inference. The proposed method uses random forest regression to model the connection between summary statistics and the parameters of interest. Compared with existing approaches, the proposed method bypasses the need of preselection of summary statistics in the model, and is capable of capturing the potential nonlinear relationship between the parameters of interest and summary statistics. We also introduce a measure to quantify the importance of each summary statistic used in the model. We study the asymptotic properties of the proposed estimator and show that it has an excellent finite-sample numerical performance via two simulation examples and an application to a population genetic study.