Wim Delva, School of Data Science and Computational Thinking, will present the Department of Statistical Science seminar with a talk entitled, "Model calibration as a missing data problem: Approximate Bayesian Computation with MICE". 

Abstract: Model calibration, i.e. estimating the parameters of a model based on data, can be an intricate problem. Standard frequentists and Bayesian statistical inference methods are available when likelihood functions can be formulated, but calibrating models of complex (adaptive) systems often requires a likelihood-free approach. Approximate Bayesian Computation (ABC) is a broad class of simulation-based, likelihood-free methods for model calibration that has gained growing popularity in a wide range of disciplines, including ecology, genetics, epidemiology, and astronomy. In the basic rejection ABC algorithm, parameters are sampled from the prior distribution, and the model output (features) under the sampled parameter values are compared to the target features. An approximate posterior distribution is obtained by only retaining those parameter values for which the output was sufficiently close to the target features. In the limit of the tolerable distance going to zero, this method converges to the correct posterior distribution. However, as the dimensionality of the parameter space increases, the number of simulations necessary to obtain a result with low bias and acceptable precision, become unfeasibly large. We describe MABC, a new ABC scheme, that hinges on the idea that model calibration can be recast as a missing data problem, for which Multivariate Imputation by Chained Equations (MICE) has become the standard methodology. After describing the MABC algorithm, we compare its performance to rejection ABC and an advanced adaptive population Monte Carlo ABC scheme.