seminari Prof. Berger

[mortera@uniroma3.it (Julia Mortera)]

                      AVVISO DI SEMINARIO
Presso la sede del  Dipartimento di Economia, Universita' 
di Roma Tre (Via Ostiense, 139), il Professor
                     JAMES O. BERGER

Arts and Sciences Professor of Statistics, Duke  University, USA
terra' i seguenti seminari:

1. Objective Bayesian Analysis of Spatially Correlated Data
              J. BERGER, V. DE OLIVEIRA and B. SANSO
        Lunedi' 11 Ottobre ore 16.30, Aula 6

2. Space-time Bayesian modeling of ozonesondes 
                  J. BERGER and J. LEE
        Martedi' 12 Ottobre ore 11.30, Aula 6

                            Abstract 1
Gaussian random field models for spatial data are typically specified
in terms of a linear mean structure (as in linear models), and a 
covariance structure. While specification of noninformative priors 
for the linear parameters is straightforward, such specification 
for the covariance parameters is quite problematical. Indeed, the 
commonly used `constant' noninformative prior for the covariance 
parameters typically yields an improper posterior distribution. This 
surprising, and not previously recognized, fact was the motivation 
for this investigation.
The most commonly used alternative to the constant prior in statistical
inference is the Jeffreys prior, which we derive for the spatial model.
Surprisingly, it happens that the Jeffreys prior also yields an 
improper posterior for this model. We finally thus derive the 
reference prior for the spatial model, and show that it solves the
problem, yielding a proper posterior distribution. It is hence the
recommended default prior for automatic use.

                            Abstract 2
There has been considerable interest among meteorologists in modeling 
ozone levels as a function of altitude. Not only are environmental 
effects highly altitude-dependent, but understanding ozone altitude 
profiles can be important in understanding the ozone depletion process. 
Each ozonesonde, as it ascends, generates measurements that can be fit 
to spatial curves. These curves must also be viewed as time series, 
since ozone levels are known to have considerable seasonal variation. 
Thus the problem is that of space-time analysis of thousands of 
ozonesonde curves.

The spatial components of the process were found to be well fit by 
mixture models. It was then possible to incorporate a time component by 
the simple device of letting the parameters of the mixture model be 
time-dependent, allowing a fully Bayesian times series analysis of the 
data. This was carried out via a dynamic linear model, formulated on 
the basis of a mixture of scientific knowledge about the seasonalities 
and data snooping. 
________________________________________________________
Tutti gli interessati sono invitati a partecipare 


Julia Mortera
Dipartimento di Economia
Universita' di Roma 3
Via Ostiense 139
00154 Roma- ITALY
tel 39-6-57374206
fax 39-6-57374093
email: mortera@uniroma3.it