no subject (file transmission)

[Renata Rotondi <reni@iami.mi.cnr.it>]

=======================================================================
=                                                                     =
=                         C.N.R. - I.A.M.I.                           =
=  Istituto per le Applicazioni della Matematica e dell'Informatica   =
=                                                                     =
=======================================================================


Si comunica che il professor James Berger, Department of Statistics,
Purdue University, terra`, presso la sede del C.N.R. (via Ampere
56, Milano) aula A, i seguenti seminari:

	9 Maggio 1996   11.00-12.00
    Recent Developments in Bayesian Model Selection

    10 maggio 1996  14.30-15.30
    Unifying Classical and Bayesian Testing

Abstract: The classical Neyman-Pearson theory of testing hypotheses 
is well developed, but has often been criticized for providing fixed
error probabilities, no matter what data is observed. Common 
alternatives, such as P-values, are data-dependent but have
come under different criticisms, from both frequentists and Bayesians.
The result has been that testing, unlike much of estimation, has
remained a very divisive subject, with frequentists, Bayesians, and
P-valuers all proposing substantially different answers.
Kiefer [J. of the American Statistical Association 72 (1977),
789-827] sought to overcome these problems by developing a
"conditional frequentist" theory, in which the error probabilities
are allowed to depend on the data. The theory never became popular,
however, because of the difficulty of choosing an appropriate
conditional frequentist procedure from among the many possible.
A recent look at the problem has revealed a very natural
and attractive choice for the conditional frequentist error
probabilities, namely the Bayesian posterior probabilities of the
hypotheses. That the Bayesian answers can be given a valid frequentist
interpretation is quite surprising. Even more surprising is that the
Bayesian answers can often be shown to be the optimal frequentist
answer, better than the Neyman-Pearson fixed error probabilities.
A true unification of testing thus appears to now be possible.
The talk will primarily focus on testing in standard normal
situations, such as two-sided normal testing and ANOVA type testing.
For these standard situations, the dual "frequentist-Bayesian" tests
will be presented and compared with the classical tests. If time
permits, discussion of sequential testing will also be undertaken,
and it will be shown how the new tests surprisingly ignore the
"stopping rule", while maintaining a frequentist interpretation.

Con l'invito ad intervenire, La prego di dare la piu` ampia diffusione
al presente annuncio.


                                      Il Direttore dello IAMI

                                      Eugenio Regazzini

=======================================================================
Per informazioni rivolgersi a:
Renata Rotondi
Istituto per le Applicazioni della Matematica e dell'Informatica
Via Ampere, 56
20131 Milano
reni@iami.mi.cnr.it
=======================================================================