DSAS Colloquium Talk - Wei Zhang
Room: 248
Speaker: Wei Zhang (University of Auckland, New Zealand)
Title: Maximum likelihood estimation for latent multinomial models
Abstract: We investigate the problem of parameter estimation under latent multinomial models, in which observed data is a linear transformation of a latent vector of counts arising from a multinomial distribution with unknown parameters. Currently it relies primarily on Bayesian methods, which involve long computation times and often require expert implementation. In this talk, I will present a novel likelihood-based approach suitable for all models in the class, using likelihoods constructed by the saddlepoint approximation method. We validate the method by applying it to specific models for which exact or approximate likelihoods are available, by comparing it with other estimation approaches, and by simulation. The saddlepoint method consistently gives accurate inference while being considerably faster than Bayesian methods and more general than other alternative estimation approaches. We show the generality of the approach by applying it to two new models for which no existing likelihood-based approach has been proposed, for estimating the prevalence of diabetes from medical records distributed across several partially-reconciled administrative lists, and for estimating cell entries of a multi-way contingency table given subsets of marginals.