We consider a discrete latent variable model for two-way data arrays, whichallows one to simultaneously produce clusters along one of the data dimensions(e.g. exchangeable observational units or features) and contiguous groups, orsegments, along the other (e.g. consecutively ordered times or locations). Themodel relies on a hidden Markov structure but, given its complexity, cannot beestimated by full maximum likelihood. We therefore introduce compositelikelihood methodology based on considering different subsets of the data. Theproposed approach is illustrated by simulation, and with an application togenomic data.
展开▼