A0350
Title: Fast implementation of a general importance sampling algorithm for Bayesian nonparametric models with binary responses
Authors: Dennis Christensen - University of Oslo (Norway) [presenting]
Per August Moen - University of Oslo (Norway)
Abstract: Binary response data problems, such as those arising in bioassay, current status data and binary classification, have been an important subfield of Bayesian nonparametrics for the last 50 years. For models based on the Dirichlet process, there exist Markov chain Monte Carlo (MCMC) algorithms given such data. However, for many of the new models developed over the preceding decade, MCMC methods are unavailable when the data comprise both left and right-censored observations. We introduce a new, highly general-importance sampling algorithm which enables posterior inference for any nonparametric model from which a random sample can be generated. Calculating the importance weights is equivalent to computing the permanents of a class of (0,1)-matrices, which we prove can be done in polynomial time. Furthermore, we provide an efficient implementation of the algorithm by optimising memory management and exploiting sparse data structures. This allows the importance sampling algorithm to handle datasets of size up to several thousand observations.
