ABSTRACT

This chapter presents the approximate Bayesian computation (ABC) random forest strategy for model choice and considers first a toy example and, at the end, a human population genetics example. It also presents a solution for conducting ABC model choice and testing that differs from the usual practice in applied fields like population genetics, where the use of Algorithm remains the norm. The chapter considers the massive single nucleotide polymorphism (SNP) dataset already studied in, associated with a Most Recent Common Ancestor (MRCA) population genetic model corresponding to Kingman's coalescent that has been at the core of ABC implementations from their beginning. Subsequent ABC model choice steps based on the selected summaries are detrimental to the quality of the classification once a model is selected by the random forest. The difficulties to learn how to discriminate between models certainly increase when the number of likelihoods in competition gets larger.