Research Seminar Series
Statistics Seminars: Random Effects, Mixtures and NPMLE
15 May 2012 16:30 in TBA
There are many situations where a basic simple model is inadequate to explain the key components of variation in a particular setting or dataset. This can arise when there are structured sampling aspects, such as a hierarchy of sampling units, longitudinal observations or missing unobserved covariates at one or more of the sampling levels. Two, now relatively common, possible model extensions are the inclusion of one or more random effects and the use of mixtures of the basic model. There are many links between these approaches, particularly in the use of the EM algorithm for model fitting, and this talk will consider these. I will also discuss the effect of different distributional assumptions for the random effect and consider the use of a random effect with an unspecified distribution using non-parametric maximum likelihood estimation with the R package npmlreg. For simplicity, we will generally confine attention to models for a univariate response, typically in the single parameter exponential family, although multivariate extensions are certainly feasible. This talk will touch on these various ideas and illustrate the models with some applications.
This is a local RSS seminar.
Information about seminars for the current academic year. For information on previous years' seminars please see the seminar archives pages.