Young Researchers’ Conference 2014

5 juin 2014

Valérie Chavez Demoulin (Université de Lausanne) Mervat Cluzeau UniGe (coordinatrice)

Laura Florina Turbatu, William Aerberhard, Reto Bürgin, Mark Hannay, Carlos De Porres, Setareh Ranjbar

La YRC est une conférence organisée par des doctorants pour les doctorants en statistique et probabilités appliquées. C'est une excellente opportunité de rencontrer d'autres doctorants, initier des connections entre institutions et présenter/discuter de son travail de recherche dans un environnement décontracté

Des présentations orales, des posters, des discussions sont organisés par les doctorants pour les doctorants.

Fancy non-(non)linearity, Setareh Ranjbar, UNIGE

 Recent developments in the field of econometrics and statistics provide us with a vast variety of analytical tools. This calls for more caution when using or adopting them to applications of real life. In this paper we provide two examples that highlight some of the common mistakes in applying sophisticated methods in empirical studies. In the first example we use a class of ordered discrete response model which is often used in psychology or socio-economics to study the subjective well-being. We show that adopting generalised partial linear model for making inferences on the non-parametric part of the model, will not free one from the correct  specification of the parametric part. In contrary, misspecification in the parametric part will adversely affect the estimation of the non-parametric part. The second example refers to techniques that have been widely used in policy evaluation. It shows why global estimators are not an adequate tool for prediction. Especially in cases where the distributions of the confounders differ seriously over groups, a situation that is typically desired, they easily lead to severely biased estimates.

 Partial differential equation, Matthieu Wilhelm, UNINE

We will present the weak form of a partial differential equation (PDE) and the Galerkin method. Both are the main "ingredients" to numerically solve PDE's. We then will show that solving a problem of penalized regression is equivalent to solving a PDE. We briefly present the finite element method and show how it can be used to solve PDEs. Next, we show how this methodology can be applied in the context of certain statistical problems. For example, it may be suitable to estimate a function defined over a region, using sparse observations. In order to assess the good performance of the proposed method in various contexts, we illustrate it through several examples and simulations.erequisites: basic linear algebra and basic calculus.

A broad introduction to Bayesian statistics, Thomas Lugrin, EPFL

Frequentist and Bayesian inferences are generally presented as two conflicting ways of deriving "causes" (probability distribution) from "effects" (observations). We aim to present the Bayesian paradigm as a complementary approach to frequentist inference, and show how the former can be justified through fundamental principles of statistical theory and decision theory. We also address the delicate question of prior distribution selection and if time allows, we shall give a quick introduction to Bayesian computations and Markov chain Monte Carlo methods."

 Empirical Bayes Unfolding of Elementary Particle Spectra at the Large Hadron Collider, Mikael Kuusela, EPFL

We study the high energy physics unfolding problem where the goal is to estimate the true spectrum of elementary particles given observations corrupted by the limited resolution of a particle detector. This is a challenging statistical inverse problem arising in data analysis at the Large Hadron Collider at CERN. Mathematically the problem reduces to estimating the intensity function of an indirectly observed Poisson point process, but estimating this intensity suffers from two major complications: how to choose the regularization strength of the unfolding algorithm and how to quantify the uncertainty of the solution. We propose solving the first problem in the framework of empirical Bayes estimation. In particular, we employ a Monte Carlo expectation-maximization algorithm to find the marginal maximum likelihood estimate of the hyperparameter controlling the strength of the regularization. The use of empirical Bayes invalidates using Bayesian credible intervals for uncertainty quantification. Instead, we propose using bootstrap resampling for constructing purely frequentist confidence bands for the true intensity. The performance of the proposed methodology is demonstrated using both simulations and real data from the Large Hadron Collider.

Automatic module detection in microarray gene expression data, Alix Leboucq, EPFL 

Independence of genes is a common assumption in microarray data analysis, although it is not true in practice. Indeed, genes are activated in groups called modules: sets of co-regulated genes. These modules are usually defined by biologists, based on the position of the genes on the chromosome or known biological pathways (KEGG, GO for example). Our goal in this work is to be able to define common modules to several studies, in an automatic way. We use an empirical Bayes approach to estimate a sparse correlation matrix common to all studies, and identify modules by clustering. Simulations show that our approach performs better than existing methods in terms of detection of modules across several datasets, and we illustrate our method on a set of 4 serous ovarian cancer studies.

 Robust score and robust residual regression tests in GLM, Mark Hannay, UNIGE

The robust quasi likelihood ratio test is the standard test in the GLM setting for robust tests. While it performs well both in terms of level and power, it is computationally expensive especially in a forward search. The computational cost comes from the fact that the standard test must fit the full model for each of the possible extensions. We propose two robust tests in the GLM setting, which do not need to fit the full model with the robust scores. This makes them computationally much cheaper than the robust quasi likelihood ratio test. In a simulation study we compare the performance of these tests with those of the standard test, both in terms of level and power but also in terms of robustness performance.

UNIVERSITE DE GENEVE

L'inscription pour les doctorants du programme CUSO est gratuite. Le repas, les pauses café ainsi que les grillades sont offerts. Les frais de transport sont remboursés sur la base CFF 1/2 tarif 2ème classe de l'institut d'attache à l'Université de Genève.

Les frais d'inscription des doctorants d'autres institutions se monte à CHF : 60,- et couvrent le repas, les pauses cafés ainsi que les grillades.

La participation est toutefois sous réserve de places suffisantes. 

25