Constructing ABC summary statistics: semi-automatic ABC
Paul Fearnhead, Dennis Prangle · Nature Precedings · 2011
Abstract To analyze complex datasets efficiently, ABC algorithms require well-chosen low dimensional summary statistics of the data. We present a method to construct appropriate summary statistics for ABC in a semi-automatic manner. Previous attempts at this have been based around constructing statistics that are sufficient (or approximately sufficient). However, in most real applications it is difficult to know if low-dimensional sufficient statistics exist, or how to approximate them if they do. We take an alternative approach of aiming for summary statistics which will enable inference about certain parameters of interest to be as accurate as possible. This talk describes our method, the underpinning theory, and promising empirical results on a range of applications.