Approximate Bayesian computation (ABC) is now an established technique for statistical inference in the form of a simulator, and approximates the likelihood at a parameter θ by simulating auxiliary data sets x and evaluating the distance of x from the true data y.
Synthetic likelihood is a related approach that uses simulated auxiliary data sets to contract a Gaussian approximation to the likelihood. However, these approaches are not computationally feasible in cases where using the simulator for each θ is very expensive.
This talk investigates two alternative strategies for inference in such a situation. The first is delayed acceptance ABC-SMC (arxiv.org/abs/1708.02230), in which a cheap simulator is used to rule out parts of the parameter space that are not worth exploring. The second is bootstrapped synthetic likelihood (arxiv.org/abs/1711.05825), which uses the bootstrap to cheaply estimate the synthetic likelihood.
We also examine a synthetic likelihood approximation that is constructed, using the bag of little bootstraps, from subsampled data sets. Applications to stochastic differential equation models and doubly intractable distributions will be presented.