Michael Blum and Olivier François, along with Katalin Csillery, just released an R package entitled abc. (I am surprised the name was not already registered!) Its aim is obviously to implement ABC approximations for Bayesian inference:
Description The ’abc’ package provides various functions for parameter estimation and model selection in an ABC framework. Three main functions are available: (i) ’abc’ implements several ABC inference algorithms, (ii) ’cv4abc’ is a cross-validation tool to evaluate the quality of the estimation and help the choice of tolerance rate, and (iii) ’postpr’ implements model selection in an ABC setting. All these functions are accompanied by appropriate summary and plotting functions.
The core abc function starts from simulated samples (from the prior and from the sampling distribution) and elaborates on the standard hard-thresholding found in the basic ABC algorithm. The extensions use nonparametric perspectives defended by Blum and Francois that I think are appropriate in this setting. Other major functions include a cross-validation procedure for selecting the threshold and an application that computes posterior probabilities of models under competition, using the conglomerate of summary statistics across models. (As in our paper with Jean-Marie Cornuet, Aude Grelaud, and Jean-Michel Marin.) I have not had time yet to experiment with the package, however I can testify the manual is well-written!
Filed under: R, Statistics, University life Tagged: ABC, approximation, likelihood-free methods