Seminario de Estadística

Ana Justel. Universidad Autónoma de Madrid, España.
Métodos para la Selección de Variables en Análisis Cluster
Sala A-4 Ingeniería - Campus San Joaquín - 15:00 Hrs.
Resumen: En este trabajo se presentan dos procedimientos para seleccionar variables en análisis cluster, que también se pueden extender a problemas de clasificación. El primero de los métodos es especialmente adecuado para detectar variables no informativas, mientras que el segundo también identifica variables con información redundante. Ambos métodos se emplean para seleccionar variables tras la aplicación de un procedimiento cluster que proporciona resultados satisfactorios. Para problemas de alta dimensión la selección de las variables se lleva a cabo mediante un algoritmo tipo forward-backward. Los resultados que se obtienen son muy satisfactorios tanto en simulaciones como en datos reales.
Wesley Johnson. University of California - Irvine
Joint Modelong of Longitudinal and Survival Data Using Mixtures of polya Trees
Sala 1 - Facultad de Matemáticas - 16:30 Hrs.

We discuss a new approach to jointly modeling survival data with time dependent covariates. Historically, time dependent covariates (TDC) were not modeled but rather analysis proceeded with them regarded as fixed. For example, blood pressure is monitored on a regular basis for dialysis patients. It is known that blood pressure that is "too low" can lead to death in these patients. Blood pressure is assumed to be constant in between readings in a standard Cox model analysis of such data. Moreover, blood pressure is also measured with error. If TDC’s are monitored frequently enough, the necessity to model them is diminished. However, to alleviate potential problems associated with these issues in general, methods have been developed whereby the longitudinal (TDC) process is modeled jointly with time to event data. We develop and discuss the problem assuming the need for such modeling and allow baseline distributions to be drawn from broad families in the context of the traditional Cox model, the proportional odds model and the Cox and Oakes model, all with time dependent covariates. We compare results based on these and parametric counterparts using mediterranean fruit fly data. These data were analyzed in the literature using a Cox and Oakes model since it was deemed that the Cox model was inappropriate. Our results indicate that the Cox and Oakes model is also not appropriate for these data since the proportional odds TDC model was preferable due to having a better model fit. This work is joint with Adam Branscum and Tim Hanson.

Georg M. Görg. Facultad de Matemáticas - UC
Lambert W Random Variables - a Family of Generalized Skewed Distributions
Sala 2 (Vìctor Ochsenius) - Facultad de Matemáticas - 16:30 Hrs.
Rosangela Loschi. Universidade Federal de Minas Gerais, Brasil
A new approach for the piecewise exponential model
Sala 2 Víctor Ochsenius - Facultad de Matemáticas - UC

One of the greastes challenge related to the use of piecewise exponential models is to find an adequae grid of time-point needed in its construction. In general, the number of intervals in such grid and the position of their endpoints are ad-hoc choices. We extend previous works by introducing a full Bayesian approach for the piecewise exponential model in which the grid of time-point (and, consequently, the endpoints and the number of intervals) is random. We estimate the ailure rates using the proposed procedure and compare the results with the non-parametric piecewise exponential estimates. Estimates for the survival function using the most probable partition are compared with the Kaplan-Mier estimators. A sensitivity analysis for the proposed model is prov

Alejandro Murua. Départament de Mathématiques Et de Statistique, Université de Montréal
Sobre el Agrupamiento de Datos Con el Modelo de Potts y el Método de K-Centroides Con Núcleo (On Potts Model Clustering and Kernel K-Means)
Sala 2 (Vìctor Ochsenius) - Facultad de Matemáticas - 16:30 Hrs.
Maris Laan. Universidad de Tartu, Estonia
Study of The Genetic Components of Essential Hypertension in Estonia
Sala 2 (Vìctor Ochsenius) - Facultad de Matemáticas - 16:30 Hrs.
Sébastien Van Bellegem. Université Catholique de Louvain
Sala 2 (Vìctor Ochsenius) - Facultad de Matemáticas - 15:00 Hrs.
We consider the problem of identification and inference in nonparametric regression under conditional moment conditions.
In contrast to the parametric case, the nonparametric regression is known to lead to an ill-posed inverse problem and the
goal of the talk is to explain how to find stable (consistent) estimators by regularization methods.

We illustrate the problem by developing inference in partially linear models under endogeneity. We derive a $sqrt{n}$-consistent estimator for the parametric part under mild assumptions and
show that the estimator achieves the semiparametric efficiency bound, even if the model is heteroscedastic.
An application of the model to the analysis of school performance (using the SIMCE database),
Esley O. Johnson. University of California, Irvine
Bayesian semiparametric inferences for disease risk, ROC curves and prevalence
Auditorio Ninoslav Bralic - 15:00 Hrs.
We discuss the application of Mixtures of Polya Trees to the semi-parametric estimation of mixture distributions in the context of assessing disease risk and of assessing the quality of a diagnostic procedure. Sampled individuals are either diseased or not and their status is unknown. Diagnostic procedures result in a score from a corresponding mixture distribution where the mixing parameter is the prevalence of the disease. Bayesian nonparametric and semi-parametric inferences are provided for Receiver Operating Characteristic curves and areas under them, as well as for prevalences. Covariates are modeled for the purpose of classification of individuals with unknown status as either diseased" or "non-diseased" and are also incorporated as factors that might affect the quality of a diagnostic test."
Michel Mouchart. Universidad Católica de Lovaina
Ignorable Missignerss
Sala D104 (Bachillerato)
Emmanuel Lessafre. Universidad Católica de Leuven
Bayesian Methods in The Context of Irt Models
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
Guillermo Marshall. P. Universidad Católica de Chile
Discriminant Models for Unbalanced Logitudinal Data With Multivariante Responses
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
María Gloria Icaza. Universidad de Talca
Metodologías Para el Análisis de la Agrupación Espacio Temporal de la Diabetes Infantil
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
Antonio Sanhueza. Universidad de la Frontera, Temuco
Métodos Robustos en Modelos No Lineales Multivariados
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
Wesley Johnson. Universidad California Davis
Eliciting Prior Distributions in Applied Problems
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
Pilar Iglesias. P. Universidad Católica de Chile
Factores de Bayes Globales
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
Paul de Boeck. Universidad Católica de Leuven
Multidimensionality At The Item and At The Person Side in Irt Models
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
Héctor Allende. Universidad Técnica Federico Santa María
Aprendizaje Robusto en Modelos de Pronóstico Mediant Uso de Redes Neuronales Artificiales
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
Heleno Bolfarine. Universidad de Sao Paulo, Brasil
Measurement Error Models: An Overview
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
Fernando Quintana. P. Universidad Católica de Chile
Un Modelo Bayesiano No Paramétrico Para Datos Ordinales Multivariados
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.
Bernardo Lagos. Universidad de Concepción
Bartlett Correction for The Signed Square Root of Likelihood Ratio Statistic With Uncensored Samples
Sala Víctor Ochsenius, ERC 15:00- 16:00 Hrs.