CMStatistics 2022: Start Registration
View Submission - CMStatistics
B1619
Title: Robust approximate Bayesian inference Authors:  Nicola Sartori - University of Padova (Italy) [presenting]
Erlis Ruli - University of Padova (Italy)
Laura Ventura - University of Padova (Italy)
Abstract: A method is illustrated that allows the construction of pseudo posterior distributions based on unbiased estimating functions. In particular, such estimating functions are used to construct suitable summary statistics in Approximate Bayesian Computation algorithms. The composite score function is a prominent example of estimating function that can be used in complex models when the likelihood computation is too demanding or when a full model specification could be too strong. The latter case implies weaker model assumptions than a full Bayesian analysis and therefore can lead to a more robust inference. In order to directly address the robustness of the posterior distribution, we propose the use of M-estimating functions instead. The theoretical properties of the corresponding robust posterior distributions are discussed.