Pennsylvania State University - Department of Statistics
Markov chain Monte Carlo (MCMC) algorithms provide a general recipe for estimating properties of complicated distributions. They are particularly useful in the context of spatial generalized linear mixed models (SGLMMs). SGLMMs based on Gaussian random fields are widely used for spatial models and machine learning. While MCMC has become commonplace in inference for such models, users have to contend with fine-tuning the algorithm for each new model and data set, determining appropriate starting values, deciding whether the algorithm is producing accurate estimates, and determining an appropriate length for the Markov chain. Using recent developments in MCMC theory and analytical approximations to construct provably fast mixing algorithms, I will outline some approaches for automating MCMC algorithms in the context of SGLMMs. While my focus will be on SGLMMs, some of the ideas in this talk apply to MCMC-based inference much more generally.
SGLMMs also often have a large number of random effects that make regression parameters uninterpretable and computing very challenging. Time permitting, I will describe an approach that reduces the number of random effects and makes the regression parameters of the model more readily interpretable.