Some developments of ABC
Seminar Room 1, Newton Institute
Approximate Bayesian computation based on parameter sets simulated under the prior that are accepted or rejected according to whether a simulated dataset resembles the observed data, has become a widely-used tool in population genomic studies since Pritchard et al (1999), and its use is growing in other areas. Developments of the basic idea have involved regression adjustment of the accepted values to mitigate the effects of discrepancies between simulated and observed datasets (Beaumont et al 2002) and embedding the approximation within a Metropolis-Hastings algorithm to create "likelihood-free" MCMC. We review these and more recent developments, for example based on sequential Monte Carlo and various adaptive simulation schemes.