ruth@utstat.uucp (Ruth Croxford) (02/08/90)
Topic: Monte Carlo Maximum Likelihood in Exponential Families Speaker: Charles Geyer, Department of Statistics, University of Washington Date: TUESDAY, Feb. 13, 4:00 - 5:00 p.m. Place: Room 1078, Sidney Smith Hall, 100 St. George Street, U of T Abstract: In many exponential family problems the likelihood cannot be calculated exactly except in very small problems. These occur notably in spatial statistics, where they are called Markov random fields or Gibbs distributions, and in genetics. In most cases one can construct by the Metropolis algorithm or the Gibbs sampler a Markov chain whose equilibrium distribution is one of those in the exponential family. From one Monte Carlo sample from this Markov chain, one can construct estimates of every distribution in the exponential family, their moments, the whole likelihood function, and the maximum likelihood estimate. All of these converge in an appropriate sense to objects they estimate almost surely along sample paths of the Markov chain. As an application of this method, an autologistic model for obtaining estimates of relatedness of individuals from DNA fingerprints will be described. This is a constrained model with many parameters. Maximum likelihood will be compared with maximum pseudolikelihood and maximum conditional likelihood. --------- Coffee and tea will be served in the De Lury Lounge (SS6006) at 3:30 p.m.