Statistics Seminars 2001-2002,
Department of Economics, Pompeu Fabra University
Schedule.
Thursday, October 18, 12:00, room 20.179, Jaume I building.
Anna Espinal
A two-step estimator for censored linear models with measurement
errors
on covariates.
ABSTRACT: In linear regression models when covariates are contaminated
with
measurement errors the Least Square estimates of the coefficients are
biased.
Some specialized methods taking into account for measurement errors have
been
largely studied (e.g. Fuller 1987). In this paper we propose to adapt
some of
these methods when the dependent variable of the linear model may be
censored. A
two-step estimator giving consistent estimates of the regression
coefficients,
emerges using the method proposed by Schneider & Weissfeld (1986). The
standard
error of the two-step estimator is obtained using Bootstrap methods.
Several
simulations are displayed to show the behavior of the two-step estimator.
Tuesday, October 23, 17:30, room 20.237, Jaume I building.
Forrest Young (University of North Carolina at Chapel Hill)
and Pedro Valero-Mora (Universitat de Valencia)
Seeing Your Data:
Dynamic, Highly Interactive Statistical Analysis with ViSta, the Visual
Statistics System
Abstract:
ViSta - The Visual Statistics System - is ready to help you "Have fun
seeing what your data
seem to say". ViSta's fun and playful approach to data analysis helps
you see what your data
seem to say, and to test the truthfulness of what you think
you've seen.
ViSta's very high interaction dynamic graphics show multiple animated
views of the data which
the user can play with. The graphics and their interactivity are
designed to augment the
user's visual intuition, and to provide a better understanding of the
data.
ViSta is a free and open system: You can download it for free from
www.visualstats.org. The
code is open to you so that you can add analysis methods of your
choosing.
We will show you what ViSta looks like, and how you use it, including an
example of a visual
Log Linear analysis. We will also show how the Log Linear analysis and
visualization methods
were added to ViSta.
Thursday, November 8, 17:30, room 20.183
Nick Longford (De Montfort University, Leicester, UK)
Examples of multiple imputation in large-scale surveys
Abstract:
Multiple imputation (MI) is a general method for dealing with
incompleteness of the analysed data. Although it has been applied in
a number of survey programmes in the USA, European statistical offices
are reluctant to use it. I will present three examples (proposals),
in which there is no effective alternative to MI, especially for some
more detailed analyses. In the analysis of ILO unemployment in the UK
Labour Force Survey, values are imputed for non-response by `bringing
the last value forward'; MI is a more realistic proposal because it
does not inflate the information. In the analysis of the food
frequency questionnaire in the UK Women's Cohort Study, there is
non-response to individual items, but MI can be used also to deal with
coarse data and respondent's misjudgement. In the Scottish House
Condition Survey, surveyors assess a stratified random sample of
dwellings, and based on their assessments the comprehensive repair
cost (CRC) of the dwelling is calculated. A random sample of
dwellings is assessed a second time to estimate how reliable are
the surveyor's assessments. MI is used to estimate the sampling
variance of the mean CRC for the country, the housing types, and the
country's unitary authorities (small areas). Details of the
implementations of MI in these examples will be given, and the full
potential of MI will be outlined.
Thursday, November 22, 17:30, room 20.237
Nicolas Vayatis (UPF)
Consistent strategies for combining classifiers and their application
to boosting algorithms.
Abstract:
The idea of combining models has shown to be very fruitful in many
statistical applications. Several algorithms, which take advantage of
the whole class of candidate models, like bagging and boosting, have
been developped to achieve efficient model selection on the basis of the
data at hand. However, the control of these algorithms with respect to
the phenomenon of overfitting is not very well understood. In this talk,
we aim at providing elements of a theoretical explanation by presenting
some results showing the existence of consistent strategies in the
spirit of boosting methods. Moreover, we will introduce some computer
simulations to discuss the impact of these results on practical
classification algorithms.
Thursday, November 29, 17:30, room 20.237.
Michael Wolf (UPF)
Subsampling inference in threshold autoregressive models
ABSTRACT:
This paper discusses inference in self exciting threshold
autoregressive (SETAR) models. These models have a wide
variety of applications in economics but there remain
a number of inference problems that have not been resolved
so far.
Of main interest is inference for
the threshold parameter. It is well-known that the asymptotics of
the corresponding estimator strongly depend upon whether the SETAR
model is continuous or not. In the continuous case, the limiting
distribution is normal and standard inference is possible. In the
discontinuous case, the limiting distribution is
non-normal and cannot be estimated consistently;
this makes standard inference impossible. We show valid inference
can be drawn by the use of the subsampling method.
Moreover, the
method can even be extended to situations where it is unknown
whether the model is continuous or not. In this case, also the
inference for the regression parameters of the model becomes
difficult and subsampling can be used advantageously there as
well. In addition, we consider an hypothesis test for the
continuity of the SETAR model.
The good small sample performance
of the methods is illustrated via some simulation studies.
As a byproduct, the paper studies the limiting distribution of
the estimator of the SETAR parameters when the general,
discontinuous model is estimated but the true model happens to be
continuous.
Thursday, December 13, 17:30, room 20.237.
Michael Greenacre (UPF)
Compositional Data Analysis versus Correspondence Analysis
ABSTRACT:
Compositional data analysis (CDA), originating in the natural sciences,
is
usually applied to data with the unit sum constraint, such as weights (or
volumes)
of components in geological or chemical samples. These compositions are
called
profiles in correspondence analysis (CA) terminology. CA originates in
the social
sciences and is usually applied to crosstabulations or other frequency
tables, but
can also be legitimately applied to compositional data. On the other
hand, CDA
can also be applied to crosstabulations and so we have two strong
competitors for
visualizing a table of nonnegative data. Also, for a particular type of
data we
can show that both methods give very similar results.
So when should we use the one, and when the other?
To answer this fundamental question, we turn our attention to the basic
principles
on which these methods are founded. CDA is founded on the principle of
subcompositional coherence, which CA does not satisfy. CA is founded on
the
principle of distributional equivalence, which CDA does not satisfy. We
discuss
this issues and also propose a third method which combines the benefits
of both
methods, satisfying both subcompositional coherence and distributional
equivalence.
Friday, January 25, room 20.237
Wolfgang Runggaldier (University of Padova).
Estimation in Term Structure Models via Stochastic
Filtering
ABSTRACT:
Financial models are often described in terms of
quantities/factors that have a theoretical economic meaning,
but cannot be observed directly and so one may want to estimate
these quantities as well as possible parameters in the model
on the basis of observable financial quantities.
Our estimation approach is via stochastic filtering
that has some advantages with respect to more clasical estimation
approaches. We consider two specific setups :
i) estimation of the instantaneous spot rate of interest and of
model parameters on the basis of observations of discretely
compounded rates;
ii) estimation of latent Markovian factors and parameters in
term structure models of the Heath-Jarrow-Morton type.
Based on joint work with Carl Chiarella (School of Finance and Economics,
University of Technology, Sydney)
Monday, February 25, 17:30, room 20.237
Miguel Delgado (Universidad Carlos III)
Distribution free goodness of fit tests for linear processes
ABSTRACT
Monday, March 4, 17:30, room 20.179.
James Davidson (Cardiff Business School)
Moment and memory properties of linear conditional heteroscedasticity
models.
ABSTRACT:
This paper analyses moment and near-epoch dependence properties for the
general class of models in which the conditional variance is a linear
function of squared lags of the process. It is shown how the properties
of
these processes depend independently on the sum and rate of convergence
of
the lag coefficients, the former controlling the existence of moments,
and
the latter the memory of the volatility process. Conditions are derived
for
existence of second and fourth moments, and also for the processes to be
L1-
and L2- near epoch dependent (NED). The geometric convergence cases
(GARCH
and IGARCH) are compared with models having hyperbolic convergence rates,
the FIGARCH, and a newly proposed generalization, the HYGARCH model. The
latter models are applied to Asian exchange rates over the 1997 crisis
period, and are shown to account well for the characteristics of the
data.
Thursday, March 14, 17:30, room 20.237
Pieter Kroonenberg (University of Leiden)
Correspondence analysis of two non-standard types of
tables: tables with a dependence structure and three-way tables.
ABSTRACT:
In this presentation a brief expose will be given of nonsymmetrical
correspondence analysis for tables with a dependence structure as
developed by Lauro and D'Ambra from Naples and three-way correspondence
analysis largely due to Carlier from Toulouse. Several examples,
some from marketing, will be presented to show the usefulness of
the techniques when dealing with moderately sized tables.
Friday, April 19, 12:00, room 20.101.
Lutz Kilian (European Central Bank)
In-sample or out-of-sample tests of predictability:
which one should we use?
ABSTRACT AND PAPER
Tuesday, April 30, 12:00, room ??.
Heinz Neudecker (University of Amsterdam)
Estimation of the noncentrality matrix of a noncentral Wishart
distribution with unit scale matrix. A matrix generalization of Leung's
domination result
ABSTRACT:
The main aim is to estimate the noncentrality matrix of a noncentral
Wishart distribution. The method used is Leung's but generalized to a
matrix loss function. Paralleling Lueng's scalar noncentral Wishart
identity is generalized to become a matrix identity. The concept of
Lwner partial ordering of symmetric matrices is used.
Gilles Blanchard
(Université Paris-Sud, Orsay)
Modelling pattern recognition by "20-question" games, and the
optimality of coarse-to-fine strategies
ABSTRACT:
We study an abstract model giving a general framework for
several actual object recognition algorithms (in images) that have
proven efficient in practice. The general paradigm is this: the set of
"target" objects is recursively divided into finer subsets, thus
forming a hierarchy. For each subset in the hierarchy, an associated test
is constructed: tests corresponding to finer subsets have more power but
have a higher cost in terms of computation power. A detection takes place
whenever all tests in a branch of the hierarchy give a positive answer.
Thursday, June 6, 17:30, room 20.191.
Stéphan Clémenon
(Université Paris X Nanterre - MODALX,
Université Paris 7)
Nonparametric Statistical Inference
for Markov Chains
ABSTRACT:
A wide variety of nonparametric estimation and testing
procedures for a density or a regression function have been suggested and
studied in the context of i.i.d. observations. The purpose of this talk is
to explain when and why some of the estimation procedures that perform well
in the i.i.d. framework, such as non linear wavelet shrinkage, may be
adapted to the case of estimation of the stationary and transition densities
of a Markov chain. We will show under which conditions of stochastic
stability probability and moment inequalities, basic tools for
investigating the performance of nonparametric estimators, can be proved for
additive functionals of a Markov Chain. As an example of application,
results on the near-minimaxity of some specific wavelet estimators will be
stated. We will also show that our argument may lead, in some specific case,
to a generalization of the Bootstrap, that is a popular and efficient tool
to construct confidence intervals and testing procedures.