Seminar overview
×
Modal title
Modal content
Autumn Semester 2010
Date & Time | Speaker | Title | Location |
---|---|---|---|
Thr 30.09.2010 16:15-17:30 |
Thordis Thorarinsdottir University Heidelberg |
Abstract
The theory of point processes offers a realistic class of models for many processes that arise in fields such as epidemiology, ecology, and geology. However, these models are often difficult to analyse and model selection methods have not been adequately investigated. The primary focus of existing model selection methodology aims to detect repulsion or aggregation in the point pattern. While this is an important first step in the modelling of the processes, there is now a need to directly compare different repulsion or aggregation models.
In many applications, this is a vital step in the modelling procedure, as the different models lead to very different interpretation of the underlying physical processes. In this talk, I will discuss this issue for climate data and introduce a general Bayesian framework that allows for such comparisons to be naturally conducted while simultaneously performing parameter inference and out of sample prediction.
Joint work with Peter Guttorp.
ZüKoSt Zürcher Kolloquium über StatistikDoes Bayes beat squinting?read_more |
HG G 19.2 |
Fri 01.10.2010 15:15-16:15 |
Alex Lenkoski University Heidelberg |
Abstract
We introduce efficient Markov chain Monte Carlo methods for inference and model determination in multivariate and matrix-variate Gaussian graphical models. Our framework is based on the G-Wishart prior for the precision matrix associated with graphs that can be decomposable or non-decomposable. We extend our sampling algorithms to a novel class of conditionally autoregressive models for sparse estimation in multivariate lattice data, with a special emphasis on the analysis of spatial data. These models embed a great deal of flexibility in estimating both the correlation structure across outcomes and the spatial orrelation structure, thereby allowing for adaptive smoothing and spatial autocorrelation parameters. Our methods are illustrated using simulated and real-world examples, including an application to cancer mortality surveillance.
Joint work with Adrian Dobra (University of Washington) and Abel Rodriguez (University of California, Santa Cruz).
Research Seminar in StatisticsBayesian inference for general Gaussian graphical models with application to multivariate lattice data read_more |
HG G 19.1 |
Thr 07.10.2010 16:15-17:30 |
Andrew Hector Universität Zürich |
Abstract
Traditionally, biology and biologists have not been heavily mathematical. However, many areas of modern biology, including ecology, are becoming increasingly sophisticated in their use of mathematics and statistics. In this talk I shall try and give an idea of some of the things community ecologists need to use statistics for. In the first part of the talk I shall take one of the hot topics in ecology over the last 10-20 years - the impact of biodiversity loss on ecosystem functioning and stability - and show how the research led us from traditional least squares linear models to more complex approaches including Generalized Linear Models, Mixed-Effects models and Bayesian multilevel models. In the second part of the talk I shall show how the analysis of plant growth has required us to get to grips with non-linear mixed-effects models and spawned collaboration with Microsoft Research on developing more mechanistic process-based models of plant growth implemented using MCMC techniques. The goal of my talk is to promote greater communication and collaboration between ecology and statistics, despite the mathematical barriers that many ecologists face.
ZüKoSt Zürcher Kolloquium über StatistikSome Challenges from the Statistical Analysis of Ecological Dataread_more |
HG G 19.2 |
Fri 08.10.2010 15:15-16:15 |
Peter Hansen Stanford University |
Abstract
We consider the case where a parameter, theta, is estimated by maximizing a criterion function, Q(X;theta ). The estimate, thetaha t, is then used to evaluate the criterion function with the same data, X, as well as with an independent data set, Y. The in-sample fit and out-of-sample fit relative to that of the true, or quasi-true, parameter, theta*, are denoted by eta = Q(X; thetahat) - Q(X; theta*) and etatilde = Q(Y; thetahat) - Q(Y; theta* ), respectively. We derive the joint limit distribution of (eta, etatilde ) for a broad class of criterion functions and the joint distribution reveals that eta and etatilde are strongly negatively related. The implication is that good in-sample fit translates into poor out-of-sample fit, one-to-one.
The result exposes a winner's curse problem when multiple models are compared in terms of their in-sample fit. The winner's curse has important implications for model selection by standard information criteria such as AIC and BIC.
Research Seminar in StatisticsA Winner's Curse for Econometric Models: On the Joint Distribution of In-Sample Fit and Out-of-Sample Fit and its Implications for Model Selectionread_more |
HG G 19.1 |
Thr 04.11.2010 16:15-17:30 |
Stephan Stahlschmidt Humboldt-Universität, Berlin |
Abstract
We present an investigation into the link between the age of an offender and the characteristics of a specific crime, namely sex-related homicides. The work provides insight into how the age of an offender affects the crime and in reverse, if knowledge of events at a crime scene could be exploited to predict the age of an unidentified offender.
As general sociological and psychological theory on this specific type of crime (and the precise influence of the offender's age) is lacking and therefore no hypothesis could have been tested up until now, the technique applied is an explorative analysis of 252 cases of sex-related homicides, which have been transferred to one of the biggest existing data bases on this specific type of crime. In detail, graphical modelling was applied to learn the structure of the data and a Bayesian network was generated. This Bayesian network constitutes a data driven model on sex-related homicides and highlights influences of the offender's age on the crime.
ZüKoSt Zürcher Kolloquium über StatistikGraphical Models and Sex-related Homicidesread_more |
HG G 19.2 |
Thr 11.11.2010 16:15-17:30 |
Federico Ambrogi Università degli Studi di Milano, Milan, Italy |
Abstract
In clinical studies where multiple events during patients follow-up are of interest, the analysis of the crude cumulative incidence (CCI) is used to support clinical decisions while the analysis of the cause specific hazard (CSH) provides information on the disease dynamics for biological hypotheses generation and follow-up planning. Treatment failure, as the event firstly occurring, may be due to causes having different clinical implications in planning therapeutic strategies. The interest is generally focused on some specific causes of failure. Since only one of them can be actually observed on each patient, the competing risks methodology is appropriate. In this context, the sub-distribution hazard model is applied to infer on the difference among crude cumulative incidences. However, inference on sub-distribution hazards are not directly interpretable from a clinical perspective. To assess treatment or covariate effects, measures of clinical impact based on crude cumulative incidence should be considered. In particular relative risks, excess of risks, relative risk reduction and number of patients needed to be treated are known to be useful to clinical practitioners. The aim of this work is to provide a straightforward approach to obtain point and interval estimates of the above measures, by resorting to the general framework of transformation models, through suitable link functions in presence of competing risks. In particular, the proposal of Klein and Andersen, based on pseudo-values, was considered as starting point. The baseline cumulative risk was estimated resorting to regression spline functions on time. Time-varying effects of covariates were tested through interaction with time functions. A literature data set on a controlled clinical trial on prostate cancer, using causes of death as end-points, was used for illustration. The critical aspects of competing risks analysis will be illustrated using a study of the impact of micrometastases on patients with unilateral breast cancer, classified as node negative at diagnosis, and who had undergone surgery with axillary lymph node dissection. In this situation the endpoint of interest is the subsequent development of distant metastases.
ZüKoSt Zürcher Kolloquium über StatistikClinical useful measures for the study of competing risks in survival analysisread_more |
HG G 19.2 |
Thr 18.11.2010 16:15-17:30 |
Heiko Bailer Clariden Leu AG, Zürich |
Abstract
Nowadays, traders and financial research and sales groups increasingly crunch numbers. However, more often than not, their models do not fit the underlying data. This results in money loosing investments and/or in models that are viewed not as useful by experienced practitioners. This talk shows what can go wrong right from the start and how the choice of simple statistical methods - that are in line with the desired outcome - can deliver very useful results. I will discuss fat-tail properties of financial time series and its implication on asset allocation. In particular, I will show through an example how a robust version of a statistical measure (Mahalanobis) helps to improve the risk adjusted return of a tactical asset allocation model.
ZüKoSt Zürcher Kolloquium über StatistikA simple but effective application of statistics in financeread_more |
HG G 19.1 |
Fri 19.11.2010 15:15-16:15 |
Fadoua Balabdaoui Université Paris Dauphine |
Abstract
We provide a representation of the maximal difference between a standard Brownian bridge and its concave majorant on the unit interval, from which we deduce expressions for the distribution and density functions and moments of this difference. This maximal difference has an application in nonparametric statistics where it arises in testing monotonicity of a density or regression curve.
Research Seminar in StatisticsDistribution of the maximal difference between a Brownian bridge and its concave majorantread_more |
HG G 19.1 |
Thr 25.11.2010 16:15-17:30 |
Simon Wood University of Bath |
Abstract
Generalized additive models (GAM) are generalized linear models where the linear predictor depends on smooth functions of covariates, and the smooth functions are the targets of inference. Generalized additive mixed models (GAMM) are the equivalent generalization of GLMMs. This talk will briefly outline the theoretical framework that allows reliable and efficient estimation and inference with such models in the R package mgcv, making the link between penalized likelihood, Bayesian and mixed model approaches. The types of smooth function that can be used as model components will then be illustrated, followed by examples illustrating the diverse range of models that fall within the scope of penalized GLMs.
ZüKoSt Zürcher Kolloquium über StatistikGAMs, GAMMs and other penalized GLMs with mgcv in Rread_more |
HG G 19.1 |
Fri 26.11.2010 15:15-16:15 |
Simon Wood University of Bath |
Abstract
Many flexible regression models can be written as penalized Generalized Linear Models, either because the linear predictor depends on random effects or unknown smooth functions, or both. In this talk I discuss selection of the strength of penalization in such penalized GLMs. The two main alternatives are Marginal Likelihood (or REML) methods and prediction error methods (such as GCV or AIC). These are compared in the light of recent work by Reiss and Ogdon (JRSSB, 2009) suggesting that REML may be preferable to GCV in practice. Computational approaches to REML based smoothness selection are considered, and some numerical problems in computing the Marginal or restricted likelihood are identified and illustrated. A computationally robust strategy for dealing with these issues is then outlined, which is implemented in the R package `mgcv'. The methods are illustrated with examples of generalized additive models and signal regression models.
Research Seminar in StatisticsREML estimation of penalized GLMs read_more |
HG G 19.1 |
Fri 03.12.2010 15:15-16:15 |
Mohamed Hebiri Seminar für Statistik |
Abstract
In this talk, I will present some works I have realized during my postdoc in the ETH around the variable selection topic. The main focus is Lasso-type methods. In a first part, we deal with the Lasso with an additional quadratic penalty. Then we present some results obtained on a transductive version of the Lasso. Finally, I will discuss some current works on the graphical model.
Research Seminar in StatisticsGoodbye Talk! read_more |
HG G 19.1 |
Thr 09.12.2010 16:15-17:30 |
Zaid Harchaoui Researcher in the INRIA of Grenoble |
Abstract
We propose a new approach for dealing with the estimation of the location of change-points in one-dimensional piecewise constant signals observed in white noise. Our approach consists in reframing this task in a variable selection context. We use a penalized least-square criterion with a ℓ1-type penalty for this purpose. We explain how to implement this method in practice by using the LARS/LASSO algorithm. We then prove that, in an appropriate asymptotic framework, this method provides consistent estimators of the change-points with an almost optimal rate. We nally provide an improved practical version of this method by combining it with a reduced version of the dynamic programming algorithm and we successfully compare it with classical methods.
ZüKoSt Zürcher Kolloquium über StatistikMultiple Change-point Estimation with a Total Variation Penaltyread_more |
HG G 19.1 |
Fri 10.12.2010 15:15-16:15 |
Jan Beran Universität Konstanz |
Abstract
Olfactory coding is an interesting branch of neurobiology with challenging statistical problems. In this talk, calcium imaging data measuring reactions to olfactory stimuli in the antennal lobe of honeybees is considered. Specific biological questions lead to a combination of regression splines, fractional processes and errors-in-variables regression. Asymptotic results for spline estimators are based on fractional calculus and integration with respect to fractional Brownian motion or Hermite processes. Test statistics for examining the effect of neurotransmitters are derived taking into account errors-in-variables and previously derived limit theorems.
This is joint work with Arno Weiershäuser (Department of Mathematics and Statistics, University of Konstanz) and several colleagues from neurobiology (Giovanni Galizia, Julia Rein, Martin Strauch, Brian H. Smith).
Research Seminar in StatisticsOn bees, splines and fractional integration - an approach to the analysis of calcium imaging dataread_more |
HG G 19.1 |
Fri 17.12.2010 14:15-15:00 |
Caroline Uhler University of California Berkeley |
Abstract
We study maximum likelihood estimation in Gaussian graphical models from the perspective of convex algebraic geometry. It is well-known that the maximum likelihood estimator (MLE) exists with probability one if the number of observations is at least as large as the number of variables. We examine the situation with fewer samples for bipartite graphs and grids. Taking an algebraic approach we find the first example of a graph for which the MLE exists with probability one even when the number of observations equals the treewidth of the underlying graph.
Research Seminar in StatisticsGeometry of maximum likelihood estimation in Gaussian graphical modelsread_more |
HG G 19.1 |
Fri 17.12.2010 15:15-16:15 |
Victor Panaretos EPFL Lausanne |
Abstract
Given two samples of continuous zero-mean iid Gaussian random functions on [0,1], we consider the problem of testing whether they share the same covariance operator. Our study is motivated by the problem of determining whether the mechanical properties of short strands of DNA are significantly affected by their base-pair sequence; though expected to be true, this effect had so far not been observed in three-dimensional electron microscopy data. The testing problem is seen to involve aspects of ill-posed inverse problems and a test based on a Karhunen–Loève approximation of the Hilbert–Schmidt distance of the empirical covariance operators is proposed and investigated. When applied to a dataset of DNA minicircles obtained through the electron microscope, our test seems to suggest potential sequence effects on DNA shape.
Research Seminar in StatisticsSecond-Order Comparison of Random Functions and DNA Geometryread_more |
HG G 19.1 |