annuncio ciclo seminari

[Elena Stanghellini <stanghel>]

Seminari del Dipartimento di Scienze Statistiche 
Universita' di Perugia, Via Pascoli, 1, Perugia.

Il Prof. Tamas Rudas del Department of Statistics, Institute of Sociology,
Eotvos Lorand, University of Budapest terra' i seguenti due seminari nei giorni 
2 e 3 aprile prossimo:


On marginal models for categorical data

These two talks discuss the wide variety of statistical problems where the
marginal modeling approach has been or can be applied, the difficulties
arising from a naive application of marginal models and theoretical
results that lead to the definition of a flexible class of marginal models
with desirable properties. Much of the material presented is joint work
with Wicher Bergsma.


Seminario I: Lunedi' 2 aprile ore 11:00 Aula A3 Dipartimento di Matematica 
e Informatica

There are several common statistical problems, where restrictions on
various marginals of a multivariate distribution are implied by design or
by the research problem at hand. Examples include test-retest problems,
panel studies, graphical models, data fusion, missing data. The talk will
start by reviewing the variety of marginal restrictions that are relevant
for these problems and will ask the question, how well-behaving sets of
marginal restrictions can be defined. It will be illustrated by examples
that the response to his question is not straightforward, as seemingly
innocent restrictions may lead to non-existing models or models with
undesirable properties. 

A careful investigation of the properties of marginal models will be based
on marginal log-linear parameterizations of the distributions on the
contingency table. Marginal log-linear parameters are defined as ordinary
log-linear parameters computed for certain marginals of the table. This
leads to a flexible class of parameterizations of the distribution where
a generalization of the iterative proportional scaling procedure can be
used to reconstruct the original distribution from the parameters.

Seminario II:  Martedi' 3 aprile ore 11:00 Aula A3 Dipartimento di Matematica 
e Informatica

A certain combinatorial property of the marginals and the effects for
which the marginal log-linear parameters are computed, called ordered
decomposability, assures that the resulting parameters are variation
independent. Ordered decomposability is a generalization of the concept of
decomposability of log-linear models. Variation independence of the
components of a parameterization is the key property in determining what
marginal models exist and what kinds of restrictions may be contradicting.

Marginal models are defined by affine restrictions on marginal
log-linear parameters. Conditions for the existence and asymptotic
normality of maximum likelihood estimates under such marginal models will
be given. This implies that the usual fit statistics will have an
asymptotic chi-squared distribution and a formula for the number of
degrees of freedom will also be presented. 


**********************************************
Elena Stanghellini 
Dipartimento di Scienze Statistiche
Via A. Pascoli - C.P. 1315 Succ. 1
06100 Perugia (Italy)

Tel +39 075 5855229 or 5855242
Fax +39 075 43242

email: stanghel@stat.unipg.it
home page: http://www.stat.unipg.it/DSS/elena.html
**********************************************

=================================================================

       Forum per la Societa' Italiana di Statistica

                       AVVERTENZA

     il nuovo servizio viene offerto in via sperimentale
     dal Dipartimento di Scienze Statistiche di Perugia
           con il software Majordomo 1.94.5
       istallato su una Workstation UNIX (AIX 4.3.3)

           Ci scusiamo per eventuali disguidi !

=================================================================