[Date Prev][Date Next]
[Chronological]
[Thread]
[Top]
seminario Subhashis Ghosal
Il Professor
Subhashis Ghosal
Free University, Amsterdam
terra', presso il Dipartimento di Statistica, Probabilita' e
Statistiche Applicate dell'Universita' "La Sapienza",
un seminario dal titolo:
"Constructing non-informative priors
through approximating sieves"
Venerdi 7 maggio 1999 ore 12:00 - Aula Castellano
Abstract:
Non-informative or default priors for Bayesian analysis has
received a lot of attention in the recent years.
The notion existed since the early days of Laplace under the
name of inverse probability. There are various methods of
construction of such priors, depending on what is meant by
non-informativeness. Among the most popular are the methods
based on invariance, relative entropy and probability
matching.
All these methods, are however, applicable only to
parametric problems. We propose a method which is
applicable to both parametric and non-parametric problems.
Our construction is simple and intuitively appealing.
We begin with an appropriate finite set, called a sieve,
which may be thought of as an approximation to the parameter
space and consider the discrete uniform distribution on it.
The approximation is made finer and finer with the
increasing sample size. The proposed prior is either a weak
limit of these discrete unform distributions, or a mixture
thereof. We show that the first approach leads to Jeffreys'
prior for smooth parametric families. For a wide
class of non-parametric problems including density
estimation, the second approach results in a consistent
posterior.
Tutti gli interessati sono invitati a partecipare.