final_program_abstracts[1]

11 IMSC Session Program
Fast inversion of a flexible regression model for multivariate
pollen counts data
Tuesday - Poster Session 3
Michael Salter-Townshend and John Haslett
We introduce a rich class of models π(y|c;θ) for multivariate zero-inflated count data.
We use the recently introduced Integrated Nested Laplace Approximation (INLA)
methodology for fast Bayesian inference on θ given training data D = {(yi, ci) ; i = 1, .
. . , n} and propose a new algorithm for fast inversion as π(c|y f ).
Such models arise in palaeoclimate reconstruction, where y represents a possibly high
dimensional vector of counts of a proxy such as pollen and c represents a low
dimensional climate. In the context of our motivating application, D represents a
modern data set used to calibrate the forward relationship in which (spatially) varying
values of c drive varying vegetative responses, whence variation in the composition of
the pollen rain, reflected in count data in samples of (eg lake) sediment.
Subsequently y f is a vector of counts in a sample taken from a sediment core, thus
reflecting ancient pollen rain and ancient climate – the palaeoclimate. The
methodology applies in principle to palaeoclimate reconstruction from many other
proxy types found in lake and ocean sediment; eg chironamids, diatoms, testate
amoebae. However, the generic issue - the statistical inversion of a multivariate
relationship - is found in many areas of application; e.g. clustering, supervised
classification, medical imaging, oil shale modelling.
The principle novelty of the paper is a new class of multivariate models based on
nested Dirichlet-Multinomial distributions with zero inflation.
Abstracts 101