Work

The Linked Importance Sampler Auxiliary Variable Metropolis Hastings Algorithm for Distributions with Intractable Normalising Constants

Year: 2008
Type: report
Abstract: We consider parameter inference for the class of models where the likelihoodfunction is analytically intractable as a result of a complicated normalising con-stant. This means that an MCMC algorithm for drawing from the posterior ofthe parameters of the model would involve evaluating an acceptance ratio con-taining a ratio of unknown normalising constants. We propose to improve onthe recently proposed auxiliary variable MCMC by extending the variable spaceof the auxiliary variable. Whereas the auxiliary variable MCMC may be con-strued as a Metropolis-Hastings algorithm that estimates the acceptance ratioin each iteration using simple importance sample based on only one observa-tion, we show how the algorithm proposed here can be seen as substituting theone-observation simple importance sample with the more efficient linked impor-tance sampler (LIS). While retaining the properties of the Metropolis-Hastingsalgorithm the use of LIS is generally applicable and allows flexibility in tuningthe mixing. In particular we show that the algorithm can be made to work forsome social network models for which Bayesian analysis has not previously beenfeasible. While the auxiliary variable MCMC works for an Ising model it doesnot work for other models, for which we show that we can achieve reasonablemixing by using the tuning features of the proposed algorithm. The modelsmentioned include a social influence model for correlated binary outcomes forpupils in an Australian school class where sociometric data is available, and,(curved) exponential family models for the collaboration network of partnersin a New England law firm.
Language: English
Open Access status: green
OpenAlex ID: W2526560059
Cites: 59
Cited by: 25
Related to: 9