Invited speakers (confirmed)
We welcome the following invited speakers:
Alan Agresti (University of Florida) [abstract]
http://www.stat.ufl.edu/~aa/
Jan-Paul Fox (Twente University) [abstract]
http://users.edte.utwente.nl/Fox/
Jerome Friedman (Stanford University) [abstract]
www-stat.stanford.edu/~jhf/
Theo Stijnen (Leiden University Medical Centre) [abstract]
http://www.lumc.nl/3020/
Gerhard Tutz (Ludwig-Maximilian University Munich) [abstract]
http://www.stat.uni-muenchen.de/~tutz/
Abstracts
Good Confidence Intervals for Discrete Statistical Models
Alan Agresti and Euijung Ryu
Department of Statistics, University of Florida, Gainesville
Florida 32611, USA, and
Division of Biostatistics, Mayo Clinic, Rochester,
Minnesota 55905, USA
We survey good methods for constructing confidence intervals
for parameters in discrete statistical models, with emphasis on
categorical data. The method of inverting score tests for parameter
values performs well, usually much better than inverting Wald tests
and often better than inverting likelihood-ratio tests. Exact
small-sample methods are conservative inferentially, but inverting a
test using the mid-P value provides a sensible compromise. For some
models ordinary score inferences are impractical, such as when the
likelihood function is not an explicit function of the model
parameters. For such cases, we propose pseudo-score inference based
on a Pearson-type chi-squared statistic that compares fitted values
for a working model with fitted values of the model when the
parameter of interest takes a fixed value. Finally, we briefly
summarize a different pseudo-score approach that approximates score
intervals for proportions and their differences by adding artificial
observations before forming simple Wald confidence intervals.
Bayesian Item Response Models For Complex Survey Data
Jean-Paul Fox
Twente University, Faculty of Behavioural Sciences, Enschede, The Netherlands
IRT methods have become an important tool in analyzing large-scale survey data. The
application of the common IRT models raises several issues like the implicit assumption of conditionally independent observations, handling collateral information, and dealing with misreporting. It is shown that the Bayesian IRT approach leads to a very flexible modeling framework for analyzing large-scale survey data. The Bayesian IRT models are extended to provide a better fit to the data and to extract richer information from the survey data. A variety of extensions will be discussed.
Fast Sparse Regression and Classification
Jerome H. Friedman
Department of Statistics, Stanford University,
Stanford, CA 94305 (jhf@stanford.edu
Regularized regression and classification methods
fit a linear model to data, based on some loss criterion,
subject to a constraint on the coefficient values. As special
cases, ridge-regression, the lasso, and subset selection all
use squared-error loss with different particular constraint
choices. For large problems the general choice of
loss--constraint combinations is usually limited by the
computation required to obtain the corresponding solution
estimates, especially when non convex constraints are used to
induce very sparse solutions. A fast algorithm is presented
that produces solutions that closely approximate those for any
convex loss and a wide variety of convex and non convex
constraints, permitting application to very large problems. The
benefits of this generality are illustrated by examples.
Random effects meta-analysis in the framework of the general(ized) linear mixed model
Theo Stijnen and Taye H. Hamza
Department of Medical Statistics and Bioinformatics, Leiden
University Medical Center, P.O. Box 9600, 2300 RC Leiden, The Netherlands,
and
Department of Biostatistics, Erasmus University
Medical Center, P.O. Box 2040, 3000 CA Rotterdam, The Netherlands
In this paper we review advanced meta-analysis methods.
We discuss univariate, bivariate and multivariate (regression)
methods, all put into the framework of the (generalized) linear
mixed model. In particular we pay attention to particular cases
where an exact within study likelihood can be used. We show that
the advent of the flexible (generalized) linear mixed model
programs in the widely available statistical packages has made it
feasible to fit advanced meta-analysis models relatively easily in
practice.
Boosting Strategies in Semiparametrically Structured Regression
Gerhard Tutz
Ludwig-Maximilians-Universitä München,
Akademiestraße 1, 80799 München
Early boosting procedures were very successful in
improving classification algorithms by applying reweighted versions
of the input data. With the representation of the procedure as a
functional optimization algorithm boosting has become a tool with
strong potential in high dimensional regression problems. Here a
general likelihood-based boosting procedure is considered that
provides tools for model selection, regularization and feature
selection in semiparametrically structured regression. From the wide
range of applications we will select three: the structuring of the
predictor in high-dimensional regression problems, constrained
regression where the effect function is constrained to be
monotonically decreasing or increasing and signal regression where
predictors have an underlying metric but the number of predictors is
in the hundreds.