Wednesday 07 May 2014Copula-based Multivariate Regression models with gamlss Marginals

Mario Cortina Borja - University College London

Plymouth University, PSQ C2 14:00-15:00

Copulas provide a flexible way to model multivariate outcomes by decomposing information from their joint distribution into that from its marginal distributions and that of the chosen copula function which de fines the dependence structure between the marginals. In the conditional case, copulas can include vectors of covariates for both the marginal distributions and the dependence parameters related to the copula function. Estimation is usually carried out by maximum likelihood methods. Often the marginals require characteristics diff erent to the multivariate normal case. In such cases the class of models within the gamlss family of generalised additive models for location, scale and shape distributions provides a wide range of alternatives including 3- and 4-parameter probability models accommodating asymmetric and leptokurtic or platykurtic patterns. Appropriate link functions can be speci ed with diff erent linear predictors on each marginal's covariates. In this talk the possibilities of embedding gamlss marginals in copula-based multivariate regression models will be discussed. Two examples will be presented, on modelling bivariate refractive error data, and on modelling dates of birth and dates in cases of sudden infant death syndrome.

Add to calendar

Add to calendar (.ics)