The Department of Mathematics and Statistics provides excellent undergraduate programs, as well as advanced instruction at the graduate and research levels. Find out more about our degree programs.
The Department this week
- Crowdfunding site launches new era of fundraising at Dal
- Potential unlocked: President Florizone shares Dal's economic impact strategy with Truro and Colchester Chamber of Commerce
- Student life superstars: Meet this year's outstanding Board of Governors Award winners
- 61st Annual Black & Gold Athletic Banquet
- Shell Canada renews support for enhanced learning experiences at Dal
March 26, 2015 3:30 PM
Speaker: William Aeberhard, Postdoctoral Fellow, Department of Mathematics and Statistics, Dalhousie University
Title: Attempt to Estimate the Mixing Distributions of Mixed Poisson Model
Abstract: For fitting purposes and consistent inference, the correct specification of the relation between the variance of the response variable and its mean (known as the variance function) is crucial. For a response consisting of counts, typically involving overdisperion, many models already exist and offer a wide variety of variance functions.
This variety is appealing, yet the data analyst is still faced with the somewhat arbitrary choice of a specification over another, without many means to check the relevance of such a choice. Most of these variance function specifications are in fact the result of the specification of unobservable mixing distributions, whose realizations can be seen as random intercepts (one per observation) appearing in the mean of Poisson distributions, themselves giving rise to the observed counts. A well-known case consists of the mixing variables to be independent and identically distributed as gamma, yielding negative binomial outcomes and a quadratic variance function with constant overdispersion. We propose here a semi-parametric framework encompassing many models for count data where the underlying mixing variables need not be identically distributed and are estimated by empirical likelihood along with the regression coefficients. By doing so, the shape of the variance function is somehow estimated from the data and the data analyst need not worry about its specification. Algorithmic ideas and encouraging simulation results will be presented for this work in progress.