Space-time
Modelling
Using Independence and Generalised Estimating Equations
Richard Chandler (University College London)
16/02/10
This talk
will consider the analysis of large space-time datasets, which arise
frequently in applications such as climatology. Often, the goal in
analysing such a dataset is to investigate the relationship between one
or more response variables and a (sometimes large) number of
covariates. It is natural to study such relationships using
regression-based methods. In this context we focus in particular on the
problem of computationally efficient model selection when faced with a
space-time dataset of hundreds of thousands, or millions, of
observations. A simple procedure is to model the temporal dependence
explicitly using a generalised autoregressive structure, and then to
adjust for the spatial dependence when calculating standard errors. An
alternative is to use generalised estimating equations (GEEs), which
account explicitly for the spatial dependence in the estimation. We
show that in general the two procedures may not estimate the same
quantities, and develop tests that aim to diagnose when this problem is
present. In addition, an easily calculated adjustment to the
'independence' log-likelihood function will be described. This
adjustment accounts for spatial dependence without having to model it
explicitly. The work is motivated by, and illustrated using, a case
study involving European windspeed.
Back to the seminar
homepage
LSE Home Page |
Departmental
Home Page | Kalogeropoulos
Home
Page
[Last modified: Jan. 18th 2010 by Kostas
Kalogeropoulos]