Statistics and Operation Research Seminars 1999, Department of Economics, Pompeu Fabra University

• October 28, 16:00, Room B30. Gabriele Fiorentini (Universitat d'Alacant):Exact likelihood-based estimation of conditionally heteroskedastic factor models
• November 11, 17:30, Room B25 András Antos (Computer and Automation Research Institute, Budapest): "Lower bounds for the rate of convergence in nonparametric pattern recognition"
• November 22, 17:30, Room B25 M. Koda(U. Tsukuba, Japan): "Stochastic Neural Network Formulation using SDE"
• November 25, 17:30, Room B25 Santiago Velilla (Universidad Carlos III de Madrid):"Variable Deletion, Confidence Regions and Bootstrapping in Linear Regression"
• December 2, Alberto Maydeu (University of Barcelona): "Thurstonian Modeling of Ranking Data via Mean and Covariance Strructure Analysis"
• December 16, 17:30, Room B25 Alfredo Ibáñez (ITAM, Mexico) Monte Carlo valuation of american options through computation of the optimal exercise frontier
• February 4, 17:30, Room 137 Jai Heui Kim (PUSAN NATIONAL UNIVERSITY, Korea) Stochastic integrals of set-valued processes
  Abstract: This will be an expository talk where we will define the stochastic integrals of a set valued process and a fuzzy process with respect to cylindrical Brownian motion on a Hilbert space. We will also discuss some of its properties which are useful for the study of fuzzy stochastic differential equations.
• February 10, Room 101, 17:30 Ricardo Cao (UNIVERSIDADE DA CORUÑA) Estimadores de Kaplan-Meier y Nelson-Aalen presuavizados
  Abstract: En este trabajo se introduce una modificación de los estimadores de Kaplan-Meier y Nelson-Aalen en el modelo de censura aleatoria por la derecha, vía la estimación no paramétrica de la regresión. Se prueba la normalidad asintótica de los estimadores, dejando patente la ganancia de segundo orden en su eficiencia, lo cual también se refleja en un estudio de simulación.
• March 9, 17:30, Room 101, Roger de Llúria Ivailo Partchev (Sofia University, Bulgaria and University of Jena, Germany) Some simple strategies to explore public opinion in countries about which we know little or nothing: a look at the empirical evidence
  Abstract: The question is how to go about doing public opinion research in countries we know little or nothing about. One such country is Albania, a country in which I have conducted political research. I was surprised to see that my attempts to explore the political scene had sometimes done a better job at revealing the identity of local politicians and predicting their behaviou r than other examples of political research. This has given me insight into some possible sources of error in empirical social research. In this talk, I outline the strategy used in the Albanian polls: 1. Identify politically relevant entities -- parties, politicians, institutions, media, and events -- on which simple questions can be asked. 2. Use a combination of multiple- and single-choice questions to find out which of these entities are really important politically. 3. Use multiple correspondence analysis or simpler multivariate techniques to identify groupings among entities of a kind -- politicians, parties, newspapers etc. 4. Use simple and multiple correspondence analysis to link entities of different kinds -- e.g. parties to politicians. 5. If possible, compare results with experience from other countries, watching out for occasions where data analysis techniques seem to behave strangely. And, as a general conclusion: while political concepts and theories do not belong in questionnaires at all, they should better be avoided in interviews with local politicians and analysts as well. Subtle differences in meaning can cause much trouble in subsequent analysis.
• March 16, 13:30, Room 137, Jaume I. Antonio Cabrales, Walter García-Fontes (POMPEU FABRA UNIVERSITY) Estimating learning models from experimental data
  Abstract: We study two procedures that are commonly used to obtain parameters of learning models from experimental data: the quadratic deviation method (or QDM) and maximum likelihood. We show that both methods give asymptotically consistent estimators and provide the asymptotic variance-covariance matrix, which is typically not given in studies that make use of the QDM. Using Monte Carlo analysis we show that both methods give seriously biased estimates in samples which have the typical length of actual experiments.
• April 6, 17:30, Room 107, Jaume I. Marc Saez (UNIVERSITAT DE GIRONA) A modified pseudo-quasi likelihood method to fit generalized linear mixed models
• May 11, 17:30, Room B26, Jaume I. Rosemarie Nagel and Albert Satorra One, two, (three), infinity: lab and newspaper experiments with a beauty contest (Joint work with Antoni Bosch and Jose Garcia-Montalvo.)
• May 16, 12:00 Room 137 Jaume I. Stephen Pollock Frequency-Domain Methods for Detrended Data
• May 23, 12:00 Room 233 Jaume I. Nicholas T. Longford Department of Medical Statistics De Montfort University, England Synthetic estimation and model selection
  Abstract: The general problem of combining several estimators of the same parameter vector is considered. Such a synthetic estimator is in many settings far superior to the most efficient of the constituent estimators or to various estimators derived by hypothesis testing. This will be illustrated on two examples: prediction based on ordinary regression, and the problem of carryover in 2x2 crossover trials. It will be emphasized that instead of relying on a single model from which various estimates are derived, each estimand be considered separately.
• June 22, 11:00, Room 137, Jaume I. Márta Pintér (TECHNICAL UNIVERSITY OF BUDAPEST) Prediction from randomly right censored data
  Abstract: Let $X$ be a random vector taking values in $\R^d$, let $Y$ be a bounded random variable, and let $C$ be a right censoring random variable operating on $Y$. It is assumed that $C$ is independent of $(X,Y)$ and the distribution function of $C$ is continuous. Let $m(x):= \EXP \{ Y | X=x \} $ be the regression function. In many, e.g. medical, studies it is not possible to observe a sample of $(X,Y)$. In case of right censoring a sample $\{ X_i, Z_i, \delta_i \}_{i=1}^n$ is given, where $Z_i=\min\{ Y_i,C_i\}$ and $\delta_i=I_{[Y_i \leq C_i]}$. Based on the data and a vector of covariates $X$, we want to construct an estimate of the regression function. We modify regression estimates that are consistent in the uncensored case (kernel, partitioning, nearest neighbor, least squares and smoothing spline estimates) and show that in the right random censoring model these modified estimates are consistent.