2013/05/18 by Oliver Ratmann, Anton Camacho, Ratmann, Oliver +5
Computer Science · Mathematics · #FOS: Computer and information sciences #Gaussian Processes and Bayesian Inference #Markov Chains and Monte Carlo Methods #Methodology (stat.ME) #Statistical Methods and Bayesian Inference #stat.ME
paper · pdf · doi:10.48550/arxiv.1305.4283
Videos can be played with Acrobat Reader. Manuscript under review and not accepted
openalex publication_date 2013/05/18 · arxiv created 2014/01/22 · arxiv updated 2014/01/24 · openalex created_date 2016/06/24 · openalex updated_date 2026/07/28
Approximate Bayesian Computation (ABC) methods rely on asymptotic arguments, implying that parameter inference can be systematically biased even when sufficient statistics are available. We propose to construct the ABC accept/reject step from decision theoretic arguments on a suitable auxiliary space. This framework, referred to as ABC*, fully specifies which test statistics to use, how to combine them, how to set the tolerances and how long to simulate in order to obtain accuracy properties on the auxiliary space. Akin to maximum-likelihood indirect inference, regularity conditions establish when the ABC* approximation to the posterior density is accurate on the original parameter space in terms of the Kullback-Leibler divergence and the maximum a posteriori point estimate. Fundamentally, escaping asymptotic arguments requires knowledge of the distribution of test statistics, which we obtain through modelling the distribution of summary values, data points on a summary level. Synthetic examples and an application to time series data of influenza A (H3N2) infections in the Netherlands illustrate ABC* in action.