B0930
Title: The Yin-Yang method for merging parallel MCMC output, with application to Austrian stroke data
Authors: Sylvia Fruehwirth-Schnatter - WU Vienna University of Economics and Business (Austria)
Alexandra Posekany - University of Technology Vienna (Austria) [presenting]
Abstract: In many applied fields like economics and medical statistics large amounts of data are available, leading to a growing interest of applied Bayesian researchers and machine learners in approaches which split big data into subsets, performing inference independently in parallel and then merge these outputs. Often these data are too large for a single analysis due to the computational burden. This led to the development of approaches for combining parallelly obtained inference results, e.g. samples from posterior distributions, and subsequently obtaining a joint result which recovers the full posterior distribution and resulting posterior estimators or decisions. We propose the Yin-Yang sampler, a mathematically well-founded approach merging two samples from posterior distributions inferring different data partitions. Correcting for reusing the same prior for each subset instead of only once for the full data set is the key notion of our method. Sequential usage of Yin-Yang sampling steps retains the joint posterior from separate subsamples' posteriors for any given number of reasonably large subsets which have to contain enough information for sound inference results. To demonstrate our approach, we provide several simulation studies. In addition, we add a medical statistics application using the Austrian stroke unit data set which contains observations of more than 550 variables for over 130000 patients.
Journal