On selection of statistics for approximate Bayesian computing (or the method of moments)