November 20th, 2019
Dr. Nawid Siassi,, öffnet eine externe URL in einem neuen Fenster TU Wien.
From Dual to Unified Employment Protection: Transition and Steady State.

Abstract: We develop a computationally tractable model to study the allocational and distributional consequences of replacing a highly dual employment protection legislation (EPL) system with a unified EPL scheme. To illustrate our approach, we specialise the discussion to Spain – a country considered as an epitome of a labour market with dual EPL. First, we show that introducing a unified EPL scheme reduces unemployment and worker turnover at short job tenures. However, these changes are quantitatively limited as most of the adjustments occur through a change in bargained wages. Second, as a consequence, the policy reform has very heterogeneous effects on the lifetime value and on the volatility of labour income. The results support the view that replacing dual EPL with a unified scheme creates winners and losers among workers who are employed when the reform is implemented.

November 13th, 2019
Prof. Vladimir Gaitsgory,, öffnet eine externe URL in einem neuen Fenster Macquarie University, Sydney.
Linear Programming Approach to Long Run Average Optimal Control: The Non-Ergodic Case.

Abstract: We will discuss an infinite dimensional linear programming (IDLP) problem, which along with its dual allow one to characterize the limit optimal values of the infinite time horizon optimal control (OC) problem with time discounting and time averaging criteria. One of the results that we will concentrate on is that establishing that the Abel and Cesaro limits of the optimal value of the OC problem are bounded from above by the optimal value of the IDLP problem and from below by the optimal value of its dual, this implying, in particular, that the Abel and Cesaro limits exist and are equal if there is no duality gap. We will also discuss IDLP based sufficient and necessary optimality conditions for the long-run-average optimal control problem applicable when there is no duality gap. The novelty of our consideration is that it is focused on the general case, when the limit optimal values may depend on initial conditions of the system. The talk is based on results obtained in collaboration with V. Borkar and I. Shvartsman.

June 12th, 2019
Prof. Elena A. Erosheva, University of Washington and Université Côte d’Azur, Nice.
Analyzing JSTOR Corpus of Publications: Is There Gender Homophily in Scientific Collaborations?

Abstract: We develop methodology to analyze  gender-based homophily  in scholarly collaborations – the tendency for researchers to co-author with  individuals of the same gender. In contrast with previous efforts, which were limited to single academic disciplines, we investigate gender collaboration patterns across the scientific landscape using the corpus of JSTOR articles.  The methods we propose allow for a nuanced analysis of homophily necessitated by the fact that the data comprises heterogeneous subdisciplines and that authorships within these subdisciplines are not necessarily exchangeable. We distinguish between three components of gender homophily in collaborations:  a structural component that is due to the demographics and non-gendered authorship norms of a scientific community, a compositional component which is driven by varying gender representation across sub-disciplines, and a behavioral component which we define as the remainder of observed homophily after its structural and compositional components have been taken into account. Using minimal modeling assumptions, we measure and test for behavioral homophily. We find that significant behavioral homophily can be detected across wide swaths of the JSTOR corpus and show that this finding is robust to missing gender indicators in our data.

(Joint work with Y. Samuel Wang, Carole J. Lee, Jevin D. West, and Carl T. Bergstrom)

May 8th, 2019
Dr. Karl Oskar Ekvall,, öffnet eine externe URL in einem neuen Fenster TU Wien.
Convergence Analysis of a Collapsed Gibbs Sampler for Bayesian Vector Autoregressions.

Abstract: The use of Markov chain Monte Carlo (MCMC) to explore posterior distributions is widespread in Bayesian statistics. In order to assess or ensure the reliability of an analysis using MCMC it is essential to understand some convergence properties of the chain in use. Here, I discuss a collapsed Gibbs sampler for Bayesian vector autoregressions with predictors, or exogenous variables. The emphasis is on how the algorithm’s convergence rate is affected as the length of the sample path from the underlying vector autoregression increases. The main result, which is among the first of its kind for practically relevant MCMC algorithms, establishes an asymptotic upper bound on the convergence rate.

April 24th, 2019
Prof. Hannu Oja, University of Turku.
Information and Structures in Multivariate Data Sets: PCA vs. ICA.

Abstract: In this talk, the principal component analysis (PCA) and independent component analysis (ICA) are compared. These two tools to analyse multivariate data sets find interesting or informative directions and subspaces in the multivariate data. Large amount of information is thought to mean for example large variation, surprises, structures or regularity in the data and sometimes a good ability to compress the data. Classical statistical measures of information such as variance, entropy and Fisher information and their use in the estimation of interesting directions in PCA and ICA are discussed. The Gaussian distribution plays an important role in these analyses.

April 3rd, 2019
Dr. Emanuel Gasteiger,, öffnet eine externe URL in einem neuen Fenster TU Wien.
Endogenously (Non-)Ricardian Beliefs. 

Abstract: This paper develops a theory of endogenously (non-)Ricardian beliefs. That is, whether Ricardian Equivalence holds in an equilibrium depends on endogenous private sector beliefs. The novelty here is a restricted perceptions viewpoint: in complex forecasting environments, agents forecast aggregate variables with (potentially) misspecified models that are optimal within the restricted class, i.e., a restricted perceptions equilibrium (RPE). A misspecification equilibrium is a refinement of an RPE where the choice of restricted models is endogenous. Our formalization considers two predictors: in one rule Ricardian beliefs emerge as a self-confirming equilibrium, while the other features an equilibrium with non-Ricardian beliefs. We show that (1.) there can exist misspecification equilibria where beliefs are endogenously (non-)Ricardian, (2.) multiple equilibria exist where the economy can coordinate on Ricardian or non-Ricardian equilibria. The theory suggests a novel interpretation of post-war U.S. inflation data as being generated by endogenous belief-driven regime change and a nuanced trade-off for monetary policy rules.

March 20th, 2019
Dr. Jonas Krampe,, öffnet eine externe URL in einem neuen Fenster University of Mannheim.
Bootstrap Based Inference for Sparse High-Dimensional Time Series Models.

Abstract: Fitting sparse models to high dimensional time series is an important area of statistical inference.
In this paper we consider sparse vector autoregressive models and develop appropriate bootstrap methods to infer properties of such processes, like the construction of confidence intervals and of tests for individual or for groups of model parameters. Our bootstrap methodology generates pseudo time series using a model-based bootstrap procedure which involves an estimated, sparsified version of the underlying vector autoregressive model.
Inference is performed using so-called de-sparsified or de-biased estimators of the autoregressive model parameters. We derive the asymptotic distribution of such estimators in the time series context and establish asymptotic validity of the bootstrap procedure proposed for estimation and, appropriately modified, for testing purposes. In particular we focus on testing that groups of autoregressive coefficients equal zero. Our theoretical results are complemented by simulations which investigate the finite sample.

March 13th, 2019
Dr. Patrick Tardivel, University of Wrocław.
On the Sign Recovery Given by LASSO, Thresholded LASSO and Thresholded Basis Pursuit Denoising.


On the sign recovery given by LASSO, thresholded LASSO and
thresholded basis pursuit denoising
Patrick Tardivel, Wrocaaw university
We consider the high-dimensional regression model $Y = X\beta_0 + \epsilon$, where X is a n × p such that rank(X) = n .
In this presentation, we provide an upper bound for LASSO sign recovery which is reached when non-null
components of $\beta_0$ are infinitly large. This tight upper bound is smaller than 1/2 when the irrepresentable condition does not hold and thus generalizes the famous Theorem 2 of Wainwright [2]. In addition, when the irrepresentable condition holds, this upper-bound provides a guide to select the LASSO's tuning parameter. On the other hand, LASSO can consistently estimate $\beta_0$ under weaker assumptions than the irrepresentable condition. This implies that appropriately thresholded LASSO can recover the sign of $\beta^*$ under such weaker assumptions (see e.g. [1]). In this presentation we also revisit properties of thresholded LASSO and provide new theoretical results in the asymptotic setup under which X is fixed and non-null components of $\beta_0$ become infinitly large. Apart from LASSO, our results cover also Basis Pursuit DeNoising (BPDN). Compared to the classical asymptotic, where X is a n×p matrix and both n and p tends to +∞ our approach allows for reduction of the technical burden. In the result our main theorem takes a simple form: When non-null components of $\beta_0$ are sufficiently large, appropriately thresholded LASSO or thresholded BPDN can recover the sign of $\beta^*$ if and only if $\beta_0$ is identifiable with respect to the $l_1$ norm, i.e. 

If $X\gamma = X\beta_0$ and $\gamma ≠ \beta_0$ then $||\gamma||_1 > ||\beta_0||_1$.

For a fixed design matrix X , we define irrepresentability and identifiability curves which provide the proportion of k sparse vectors $\beta^*$ for which the irrepresentability and identifiability conditions hold. These curves illustrate that the irrepresentability condition is much stronger than the identifiability condition since the identifiability curve is highly above the irrepresentability curve.
Finally, we illustrate how the knockoff methodology can be used to select an appropriate threshold and that thresholded BPDN and LASSO can recover the sign of $\beta^*$ with a larger probability than adaptive LASSO.


[1] Nicolai Meinshausen and Bin Yu. Lasso-type recovery of sparse representations for high-dimensional data.
The Annals of Statistics, 37(1):246-270, 2009.

[2] Martin J Wainwright. Sharp thresholds for high-dimensional and noisy sparsity recovery using constrained
quadratic programming (lasso). IEEE transactions on information theory, 55(5):2183-2202, 2009.