B1724
Title: Box-ABC for likelihood-free inference
Authors: Elena Bortolato - University of Padova (Italy) [presenting]
Laura Ventura - University of Padova (Italy)
Abstract: Among the determining factors for accurately approximating posterior distributions in Approximate Bayesian Computation (ABC), the choice of a suitably small threshold $\epsilon$, for bounding the distance between observed and simulated data, plays a crucial role. This decision also prescribes the computational effort of the procedure. In fact, the value of $\epsilon$ is inversely related to the necessary number of simulations to be performed for obtaining sufficiently large Monte Carlo draws from the posterior. Furthermore, the process of tuning the tolerance may be time demanding. We propose to modify ABC algorithms, by defining acceptance rules that circumvent the use of distance functions and the choice of threshold parameters. The method implicitly makes use of a pseudo-likelihood that inherits some desirable properties from confidence distributions. We study the asymptotic behaviour of the methodology and the computational efficiency in different regimes.