Skip to main content


Thursday 19 Feb 2015Statistical Science Seminar: Building a prior distribution for the covariance matrix of a spatial multivariate normal distribution

Sarah Heaps - Newcastle University

H101 15:00-16:00

In many analyses of multi-dimensional data, the dependence
structure of the multivariate normal distribution is used to build
relationships between variables. We discuss how to construct a prior
distribution for the variances and covariances when we wish to
convey substantive prior beliefs, with particular emphasis on
spatial problems where there is typically no natural ordering of the
variables but where, in our prior beliefs, we may well have
different degrees of association between different pairs of
covariances. Convenient conjugate priors are generally too
inflexible and others which address this problem are difficult to
interpret. We discuss an approach in which an interpretable prior is
constructed and then converted to a flexible structure using a Cholesky
decomposition. We apply this to some spatio-temporal data on
monthly rainfall.

Add to calendar

Add to calendar (.ics)