Research Seminar

Would you like to be notified about these presentations via e-mail? Please subscribe here.

×

Modal title

Modal content

Autumn Semester 2015

Date / Time Speaker Title Location
18 September 2015
15:15-16:00
Robin Evans
University of Oxford
Event Details

Research Seminar in Statistics

Title Causal Models and How to Refute Them
Speaker, Affiliation Robin Evans, University of Oxford
Date, Time 18 September 2015, 15:15-16:00
Location HG G 19.1
Abstract Directed acyclic graph models (DAG models, also called Bayesian networks) are widely used in the context of causal inference, and they can be manipulated to represent the consequences of intervention in a causal system. However, DAGs cannot fully represent causal models with confounding; other classes of graphs, such as ancestral graphs and ADMGs, have been introduced to deal with this using additional kinds of edge, but we show that these are not sufficiently rich to capture the range of possible models. In fact, no mixed graph over the observed variables is rich enough, however many edges are used. Instead we introduce mDAGs, a class of hyper-graphs which is appropriate for representing causal models when some of the variables are unobserved. Results on the Markov equivalence of these models show that when interpreted causally, mDAGs are the minimal class of graphs which can be sensibly used. Understanding such equivalences is critical for the use of automatic causal structure learning methods, a topic in which there is considerable interest. We elucidate the state of the art as well as some open problems.
Causal Models and How to Refute Themread_more
HG G 19.1
25 September 2015
15:15-16:00
Damian Kozbur
Seminar für Statistik
Event Details

Research Seminar in Statistics

Title Testing-Based Forward Model Selection
Speaker, Affiliation Damian Kozbur, Seminar für Statistik
Date, Time 25 September 2015, 15:15-16:00
Location HG G 19.1
Abstract This work develops theory for a Testing-Based Forward Model Selection procedure in linear regression and m-estimation problems. Forward Selection is a greedy procedure which inductively selects covariates which reliably add predictive power into a statistical model. A testing procedure, whose properties resemble traditional statistical hypothesis testing, is assumed as a primitive. It is used as a tool for guiding the forward selection by deciding which variable to include next and when to stop including variables. This offers a flexible and data- driven way to estimate models and construct predictions in high dimensional statistical problems. Probabilistic bounds for prediction error and number of selected covariates are proved for the proposed Testing- Based Forward Selection. The regularity conditions assumed on the set of hypothesis tests are verified for several example problems. The per- formance of Testing-based Forward Selection is compared to Lasso and Post-Lasso in Simulation studies.
Testing-Based Forward Model Selectionread_more
HG G 19.1
28 September 2015
11:00-12:00
Claudia Klüppelberg
Technische Universität München
Event Details

Research Seminar in Statistics

Title Extremes on directed acyclic graphs
Speaker, Affiliation Claudia Klüppelberg, Technische Universität München
Date, Time 28 September 2015, 11:00-12:00
Location HG G 19.1
Abstract We consider a new max-linear structural equation model, where all random variables can be written as a max-linear function of their parents and noise terms. For the corresponding graph we assume that it is a DAG. We present basic probabilistic results of our model; in particular we characterise those max-linear random vectors, which originate in a max-linear structural equation model and are, hence, max-linear distributions on DAGs. We also determine the minimal DAG corresponding to the max-linear structural equation model. As an example we consider the structural equation model with max-stable noise variables, resulting in a max-stable max-linear DAG. This is joint work with Nadine Gissibl.
Extremes on directed acyclic graphsread_more
HG G 19.1
23 October 2015
15:15-16:00
Alessandra Luati
University of Bologna
Event Details

Research Seminar in Statistics

Title Generalised Linear Cepstral Models for the Spectrum of a Time Series
Speaker, Affiliation Alessandra Luati, University of Bologna
Date, Time 23 October 2015, 15:15-16:00
Location HG G 19.1
Abstract The paper introduces the class of generalised linear models with Box-Cox link for the spectrum of a time series. The Box-Cox transformation of the spectral density is represented as a finite Fourier polynomial, with coefficients, that we term generalised cepstral coefficients, providing a complete characterisation of the properties of the random process. The link function depends on a power transformation parameter and encompasses the exponential model (logarithmic link), the autoregressive model (inverse link), and the moving average model (identity link). One of the merits of this model class is the possibility of nesting alternative spectral estimation methods under the same likelihood-based framework, so that the selection of a particular parametric spectrum amounts to estimating the transformation parameter. We also show that the generalised cepstral coefficients are a one to one function of the inverse partial autocorrelations of the process, which can be used to evaluate the mutual information between the past and the future of the process.
Generalised Linear Cepstral Models for the Spectrum of a Time Seriesread_more
HG G 19.1
27 October 2015
15:15-16:15
Po-Ling Loh
University of Pennsylvania
Event Details

Research Seminar in Statistics

Title New perspectives for robust/high-dimensional estimation
Speaker, Affiliation Po-Ling Loh, University of Pennsylvania
Date, Time 27 October 2015, 15:15-16:15
Location HG E 41
Abstract Robust statistics provides a framework for quantifying the behavior of an estimator when data are subject to various imperfections which deviate from standard model assumptions. Such an understanding may aid in designing estimators that are robust to particular forms of contamination. In this talk, we will discuss two high-dimensional estimators suited for robust linear regression and precision matrix estimation. The form of the estimators is quite natural and the analysis follows from standard high-dimensional techniques; what is fascinating is the relationship between the results we derive and their analogs in low dimensions. In particular, we demonstrate that the same guiding principles for selecting estimators in classical robust statistics provide provably good behavior in high dimensions, as well.
New perspectives for robust/high-dimensional estimationread_more
HG E 41
6 November 2015
15:15-16:00
Egil Ferkingstad
NTNU Trondheim, Norway / University of Iceland
Event Details

Research Seminar in Statistics

Title Recent progress in the R-INLA project for Bayesian computation
Speaker, Affiliation Egil Ferkingstad, NTNU Trondheim, Norway / University of Iceland
Date, Time 6 November 2015, 15:15-16:00
Location HG G 19.1
Abstract After briefly introducing the INLA approach for Bayesian computation, I will discuss two topics we're currently working on. Topic 1: A new copula-based correction (http://arxiv.org/abs/1503.07307) for difficult cases, allowing us to push the boundaries of applicability of INLA while adding minimal computational cost. Topic 2: Providing sensible, default, weakly informative priors. I will present the ideas behind Penalized Complexity priors (http://arxiv.org/abs/1403.4630), focusing on work in progress showing how the idea of distributing the overall variance between different model components gives a good strategy for specifying joint priors for hierarchical models.
Recent progress in the R-INLA project for Bayesian computation read_more
HG G 19.1
4 December 2015
15:15-16:00
Benedikt Pötscher
University of Vienna
Event Details

Research Seminar in Statistics

Title Finite-Sample Properties of HAC-Tests
Speaker, Affiliation Benedikt Pötscher, University of Vienna
Date, Time 4 December 2015, 15:15-16:00
Location HG G 19.1
Abstract Testing restrictions on regression coefficients in linear models often requires correcting the conventional F-test for potential heteroskedasticity or autocorrelation amongst the disturbances, leading to so-called heteroskedasticity and autocorrelation robust (or consistent) test procedures (HAR/HAC-tests). These procedures have been developed with the purpose of attenuating size distortions and power deficiencies present for the uncorrected F-test. We develop a general theory to establish positive as well as negative finite-sample results concerning the size and power properties of a large class of heteroskedasticity and autocorrelation robust tests. Using these results we show that nonparametrically as well as parametrically corrected F-type tests in time series regression models with stationary disturbances have either size equal to one or nuisance-infimal power equal to zero under very weak assumptions on the covariance model and under generic conditions on the design matrix. In addition we suggest an adjustment procedure based on artificial regressors. This adjustment resolves the problem in many cases in that the so-adjusted tests do not suffer from size distortions. At the same time their power function is bounded away from zero. As a second application we discuss the case of heteroskedastic disturbances.
Finite-Sample Properties of HAC-Testsread_more
HG G 19.1
15 January 2016
15:15-16:00
Benjamin Frot
Oxford University Statistics
Event Details

Research Seminar in Statistics

Title Latent variable model selection for Gaussian conditional random fields
Speaker, Affiliation Benjamin Frot, Oxford University Statistics
Date, Time 15 January 2016, 15:15-16:00
Location HG G 19.1
Abstract We consider the problem of learning a conditional Gaussian graphical model in the presence of latent variables. Building on recent advances in this field, we suggest a method that decomposes the parameters of a conditional Markov random field into the sum of a sparse and a low-rank matrix. We derive convergence bounds for this estimator and show that it is well-behaved in the high-dimensional regime as well as "sparsistent" (i.e. capable of recovering the graph structure). We then describe a proximal gradient algorithm which is able to fit the model to thousands of variables. Through extensive simulations, we illustrate the conditions required for identifiability and show that there is a wide range of situations in which this model performs significantly better than its counterparts. We also show how this problem is relevant to some of the challenges faced by instrumental variable methods.
Latent variable model selection for Gaussian conditional random fieldsread_more
HG G 19.1

Note: if you want you can subscribe to the iCal/ics Calender.

JavaScript has been disabled in your browser