Le vendredi 18 mars 2011, 15 h 30 – 17 h, Concordia University, Library Building, 1400 de Maisonneuve O. / LB 921-4, SGW

Combining forward selection and shrinkage techniques for variable selection in regression and classification
Subhashis Ghosal, North Carolina State University

Variable selection is a major statistical issue in contemporary data analysis because modern data typically involve a lot of predictors, many of which are nearly irrelevant. Various sparse regression methods such as the LASSO have been developed in the literature to estimate the regression function and make predictions by setting many regression coefficients to zero using specially devised penalty functions. We propose new variable selection techniques for regression in high dimensional linear models based on a forward selection version of the LASSO (or its variants) to be called Forward Iterative Regression and Shrinkage Technique (FIRST). We exploit the fact that LASSO-type methods have closed form solutions when the predictor is one-dimensional. The explicit formula is then repeatedly used in an iterative fashion until convergence occurs. A simulation study shows that our method works better for extremely sparse high dimensional linear models. We apply the method in a gene expression study.

We also develop similar iterative procedures for classification problems based on a variant of the support vector machines (SVM). In this case, even for a one-dimensional predictor, the classifier does not have a closed form solution. However, a careful study of the objective function reveals an efficient algorithm for locating the optimal solution, which can then be iterated as in the case of linear regression, leading to a new procedure, to be called the CLASsification and Selection using Iterative Cycles (CLASSIC). We consider several variations of CLASSIC and compare their performance with other standard classification algorithms such as L1-SVM, SCAD-SVM, LASSO and penalized logistic regression through simulations. Although CLASSIC generally needs more computational time, our simulations show that the misclassification rate of CLASSIC is significantly smaller than its competitors, and generally it leads to more parsimonious models.

The talk is based on joint work with Wookyeon Hwang and Hao Helen Zhang

Mise-à-jour : Le jeudi de l’ASSQ du 31 mars est annulé. Nous vous aviserons si l’événement est remis à une date ultérieure.

Bonjour chers membres,

Vous êtes tous conviés au prochain jeudi de l’ASSQ qui aura lieu le 31 mars 2011 à l’UQÀM.

Pour l’occasion, nous adopterons une nouvelle formule. Dès midi, vous êtes invités à partager un lunch (inclus dans les frais d’inscription) tout en partageant avec les autres participants. Puis, à 13h15, Monsieur Daniel Morissette, fondateur de StatExpert Inc., présentera une conférence intitulée :

Biostatisticien dans l’industrie pharmaceutique – 20 ans plus tard !

dans laquelle il partagera son expérience de l’industrie pharmaceutique afin d’éclairer les étudiants en statistique ayant un coté entrepreneur face aux possibilités et défis de fonder leur propre entreprise. Il présentera notamment deux projets qui l’ont inspiré depuis le lancement de sa compagnie.

Des détails additionnels se trouvent sur la fiche d’inscription.

Réservez votre place dès maintenant!

Le vendredi 11 mars 2011, 15 h 30, CRM, UdeM, Pav. André-Aisenstadt, 2920, ch. de la Tour, salle 6214

Wavelet-based 2-D Spectra and Applications
Brani Vidakovic, Georgia Institute of Technology and Emory University School of Medicine

Measured bio and neuro responses, geophysical and socio-economic phenomena often possess intrinsic high frequency components and strong persistent serial correlations inhibiting statistical modeling by traditional techniques. In many modeling scenarios the low-frequency-trends are irrelevant and researchers focus on the high-frequency component and its low-dimensional descriptors. The talk overviews several traditional wavelet-based techniques for assessing the scaling in 1-, 2-, and 3-D data and some novel related techniques that are under ongoing research by the speaker and colleagues from Georgia Institute of Technology and Emory University. The focus will be on 2-D wavelet-based spectral tools. The applications include analysis and modeling of spectral responses in 1H NMR spectroscopy describing metabolic fluctuations in human plasma, classification of mammograms, turbulence, and assessment of doppler radar rainfall data.

La 24e édition de la Journée de la statistique, organisée par le CASUL (Comité pour l’avancement de la statistique à l’Université Laval), se tiendra le 24 mars 2011 au Boudoir (local 1550) du Pavillon Desjardins. Un dépiant est disponible ici. Des informations supplémentaires sont présentées sur le site web du CASUL.

Ne manquez pas ce rendez-vous annuel! Inscrivez-vous d’ici le 17 mars, c’est gratuit!

Vendredi 4 mars 2011, 15:35, McGill University, Burnside Hall, 805 Sherbrooke Street West, room 1B24

Limit theorems of functional data analysis with some applications
Lajos Horvath, University of Utah

Functional data analysis is concerned with observations which are viewed as functions defined over some set T. It can be temperature at a given location, stock prices, exchange rates, components of a magnetic field, growth curves and so on. Clearly, due to finite resolution, the values of the curve are available ata finite grid of points but due to the density of the grid and the physical interpretation of the curve, it is assumed that the observation is defined on T. We describe the basic principles of functional data analysis, including projections, the choice and estimation of the directions of the projections. We provide several examples for the applicability of the functional interpretation of the data, including the comparison of the mean curves of two populations, volatility models and functional regression.

Le jeudi 3 mars 2011, à 13 h 30, en la salle 2500 du pavillon Adrien-Pouliot

On Nonparametric Density Estimation for Length Biased Data
Yogendra Chaubey, Université Concordia

This talk will highlight some recent development in the area of nonparametric functional estimation with emphasis on nonparametric density estimation for length biased data. Such data entail constraints that many traditional nonparametric density estimators may not satisfy. A lemma attributed to Hille, and its generalization [see Lemma 1, Feller (1965) An Introduction to Probability Theory and Applications, xVII.1)] is used to propose estimators in this context. After describing the asymptotic properties of the estimators, we present the results of a simulation study to compare various nonparametric density estimators.

Le jeudi 24 février 2011, à 13 h 30, en la salle 2500 du pavillon Adrien-Pouliot

Contrôle de la divulgation de données : l’approche de la confidentialité différentielle
Anne-Sophie Charest, Université Carnegie Mellon

Toute organisation qui souhaite publier des données ou conclusions statistiques doit faire attention de respecter son engagement de confidentialité envers ses répondants. Avec la multiplication de l’information disponible en ligne sur chaque individu, cette tâche est de plus en plus ardue. Dans cette présentation, nous ferons d’abord un bref survol des différentes méthodes utilisées pour le contrôle de la divulgation, et des critères employés pour évaluer la confidentialité des données publiées. Puis, nous présenterons un critère relativement nouveau dans la littérature de contrôle de la divulgation: la confidentialité différentielle. Dans cette approche, on considère une méthode de contrôle de la divulgation comme une fonction générant un objet (statistique résumée, paramètres de modèles, jeux de données synthétiques…) à partir d’un jeu de données qu’on désire garder confidentielles. La confidentialité différentielle est une propriété de cette fonction. Il s’agit d’un critère très strict qui n’assume rien à propos des ressources de l’adversaire cherchant à révéler les données confidentielles. Nous nous attarderons à la création de données synthétiques satisfaisant le critère de confidentialité différentielle, et en particulier à l’analyse de tels jeux de données synthétiques.

Le vendredi 18 février 2011, 15 h 30 – 16 h 30, Concordia University, Concordia University, Library Building, 1400 de Maisonneuve O., salle LB 921-4

Some remarks on random matrix theory and its applications to multivariate statistics
Noureddine El Karoui (UC, Berkeley)

Over the last 10-15 years, there has been renewed interest in understanding various probabilistic properties of large random matrices. These objects also come up naturally in high-dimensional statistics problems. In this talk, I will describe some of the light random matrix results shed on various classical and modern problems in multivariate statistics, discuss robustness issues (and lack thereof) of random matrix models and present some consequences for problems in optimization, classification, and time permitting, numerical linear algebra.

Le jeudi 17 février 2011, à 13 h 30, en la salle 2500 du pavillon Adrien-Pouliot

Optimal Design for Clinical Trials with an Unknown Delay in Treatment Effect
Juli Atherton, Université McGill

Sometimes clinical trials data consist, for each subject, of a sequence of baseline pre-treatment observations that are followed by a sequence of post-treatment observations. In such trials subjects serve as their own controls. Alternatively, in a randomized controlled trial with say, a control group and one or more treatment groups, a sequence of post-treatment observations is taken on each treatment group from the start of the study. In either situation, if there is an unknown delay before a treatment takes effect (assuming it does), the subject specific times-to-effect can be viewed as the unknown change-points in a multipath change-point setting. Once the data are collected, i) they may be used to estimate the size of each treatment effect and ii) to estimate the proportion of subjects that will experience a treatment effect. Before the data are collected, however, design is the main concern, and optimal timing of the observations will enhance the precision of the eventual inference.

In a broader setting, we consider, optimal design for multipath change-point problems, focusing on problems motivated by i) and ii) above. These optimal design problems are highly non-linear and a Bayesian approach is advantageous. By introducing a design measure we are able to convert a multimodal design criterion function of the design points, into one that is a concave function of the design measures over a simple region. Through this device we reduce the optimal design problem to a feasible numerical search. We carry out simulations whose results have ramifications for the design of clinical trials in which there is an unknown lag between the start of treatment and the occurrence of an effect. In particular, our results suggest that the standard practice-no doubt historically used for convenience-of having a common observation protocol for all subjects, is, in fact, under “objective modeling”, optimal. Further, in many situations, choosing equally spaced observations, which is another common and convenient practice, does not entail a substantial loss in efficiency.

Le CA de l’ASSQ vous convie au prochain « jeudi de l’ASSQ » qui aura lieu le 24 février prochain au pavillon Vachon de l’Université Laval, salle 1240.

La première partie du 5 à 7 sera consacrée à une présentation de M. Ronald Frenette intitulée « Mesure de la satisfaction de la clientèle – résumé de lectures ». Dans la deuxième partie, vous êtes invité à échanger avec l’ensemble des participants en dégustant quelques canapés et bons vins.

Ce document vous fournit toutes les informations pour l’inscription. Pour vous assurez d’une place, réservez d’ici au 18 février 2011.

C’est un événement à ne pas manquer! On vous attend en grand nombre.