Dienstag, 29.04.2014, 13-14:30 Uhr - Raum: U3-140

Prof. Dr. Harry Haupt
Universität Passau

Nichtparametrische Prognose trendbehafteter, saisonaler Zeitreihen

A simple procedure for simultaneous smoothing in time series regressions with deterministic components, lagged endogeneous and exogeneous components is proposed. Using monthly data on traffic fatalities in Germany the forecasting performance is compared to several commonly used forecasting methods such as ARIMA and innovation state space models. Besides an analysis of the forecasting performance, we compare the proposed methods in terms of identifying and adapting to breaks in the trend.

 

Dr. Joachim Schnurbus
Universität Passau

A fully nonparametric approach to leapfrogging and switching of leading and lagging position in economic models of local growth convergence

Classical growth convergence regressions fail to account for various sources of heterogeneity and nonlinearity. Recent contributions are able to address either the one or the other. For example heterogeneity is addressed by specifiying models of local convergence within clubs of similar regions. This similarity may be defined in a data driven way or by economic explanation. Unfortunately algorithms to empirically identify members of those clubs are sensitive with respect to choices to be made by the experimenter, similar to statistical methods such as model based clustering. An example is that different rules to initiate such an algorithm — e.g. choices on clustering variables — lead to instability of within and between structure of clubs. That is membership in and number of clubs is subject to change. However, such a behavior is not an artefact of the algorithms but can be explained by modern stochastic growth theory, allowing for leapfrogging and switching of leading and lagging positions in economic growth. In the present paper we provide a unified empirical framework to treat (a) heterogeneity by identifying and explaining non-stable club structures exhibiting a local converge behavior, and (b) nonlinearities by using a fully nonparametric regression approach, which allows to estimate club specific trajectories and leapfrogging effects while providing an empirical check for potential club misclassification. The proposed approach is illustrated using the most-current Penn World Table data set.

 

Dienstag, 13.05.2014, 12-13 Uhr - Raum: U3-140

Prof. Dr. Dietmar Bauer
Universität Bielefeld

Asymptotische Eigenschaften von Subspace-Schätzern für lineare dynamische Systeme mit Einheitswurzeln

Subspace estimators are alternatives to quasi maximum likelihood estimators for multivariate, linear dynamical models. They use the state space framework and hence are attractive in particular for high dimensional data sets. In Bauer and Wagner (2002, Journal of Econometrics) the suitability of the CCA  (canonical correlation analysis) subspace estimators proposed in Larimore (1983) for estimating VARMA models for cointegrated processes has been investigated.  The main finding there was that an adaptation of the method using knowledge of the dimension of the cointegrating space leads to consistent estimators of the system. This talk extends the analytical results of Bauer and Wagner (2002) in three respects:
· First it is shown that in fact no adaptation of CCA is required in order to obtain (strongly) consistent estimators of the long- and short-run dynamics. Rates of convergence for the estimated system matrices are provided.
· Second the asymptotic distribution of the estimated system is derived leading to simple pivotal tests on the number of common trends.
· Thirdly also exogenous (possibly integrated) variables are included in the analysis.
It is shown that consistency is robust to whether the exogenous variables are stationary or integrated. The asymptotic distribution of the noise dynamics, however, depends on the characteristics of the exogenous variables. These facts show that CCA provides estimates that can be used as initial guesses for subsequent quasi maximum likelihood estimation or as estimators on their own. In addition the estimates deliver information on structural indices such as the order of the system and the number of cointegrating relations.

 

Dienstag, 27.05.2014, 12-13 Uhr - Raum: U3-140

Vortrag fällt aus

 

Dienstag, 10.06.2014, 12-13 Uhr - Raum: U3-140

Prof. Dr. Göran Kauermann
Institut für Statistik der Ludwig-Maximilians-Universität München

Statistical Analysis of Network Data – A gentle Introduction to Exponential Random Graph Models

Network data are becoming increasingly available and popular and the development of statistical network models is an emerging field of research. The talk introduces to the basic models in this area by emphasizing Exponential Random Graph Models (ERGM) as common tool. ERGM describes the distribution of a network graph with an exponential family distribution, where the statistics are counts of edges (=number of friendships), star constellations (= number of friends) or triangles (= number of joint friends), for instance. Though the model mirrors the welcome properties of exponential families, its behavior is quite unstable and the parameters space resulting in suitable network graphs is peculiarly shaped. In practice this results in either fully connected or fully unconnected networks.  Moreover, fitting of ERGMs is numerical a burden due to a non-feasible normalization constant in the exponential model. We discuss the state of the art of MCMC based sampling strategies for estimation and suggest approximate alternatives. We also include random nodal effects to compensate for heterogeneity in the nodes of a network. The talk ends by discussing the scalability of the available models and methods. How can networks with hundreds or thousands of nodes be modeled and where are current research deficits. This section of the talk has more a task force character.

 

Dienstag, 24.06.2014, 12-13 Uhr - Raum: U3-140

Dr. Odile Sauzet
Universität Bielefeld

Eine Verteilungsmethode für die Dichotomisierung einer stetigen Variable

In randomisierten kontrollierten Studien werden meistens zwei (oder mehr) Gruppen verglichen, und die statistischen Methoden hierfür sind unkompliziert. Aber was kann man tun, wenn die „beste“ Methode nicht hilft, um die Studienergebnisse richtig zu verstehen? Man kann eine andere, neue Methode entwickeln, die genau diese Fragestellung beantwortet. Aber dann kommen die Herausforderungen: die Methode muss die „Wahrheit“ (es gibt einen Unterschied) zeigen, aber muss verständlich (für einen Nicht-Statistiker) und anerkannt (von Gutachtern von Zeitschriften) sein. Anhand des Beispiels einer Dichotomisierung einer normalverteilten Variable werden wir zeigen, wie man eine einfache und (fast) genaue Methode entwickeln kann, um zwei Gruppen zu vergleichen. Somit werden wir die oben genannten Herausforderungen beantwortet.

 

Dienstag, 08.07.2014, 12-13 Uhr - Raum: U3-140

Dr. Christian Schellhase
Universität Bielefeld

Estimation of Non-Simplified Vines Using (Nonparametric) Trivariate Copula Constructions

Recently, vine copulas (or pair-copula constructions) have become an important tool in high-dimensional dependence modeling.  A commonly used assumption is that each bivariate conditional copula in the vine collapses to a bivariate unconditional copula, which is called the simplifying assumption. In this paper we consider ways and means to weaken the simplifying assumption. We show how bivariate conditional copula densities with one conditioning argument can be used to approximate bivariate conditional copula densities with an arbitrary number of conditioning arguments.  We call this a trivariate copula construction since bivariate conditional copula densities with one conditioning argument can be recognized as trivariate copula densities with particular restrictions. Using trivariate copulas we obtain a vine copula that is still feasible in practice and gives a better approximation to the multivariate density. We also present the non-parametric estimation of conditional copulas with a penalized hierarchical B-splines approach. The great advantage of this approach is that we directly estimate a conditional copula with uniform margins by placing simple linear restrictions on the spline coefficients. Thus, there is no need to extract the conditional copula from an unrestricted estimation as would be the case for other nonparametric approaches. As a result, our approach allows for fast evaluation of the conditional copulas which is crucial for applied work.

 

Dipl.-Vw. Christian Heinze
Universität Bielefeld

A constraint least-squares approach to diffuse initialization

The talk identifies the diffuse Kalman filter with a sequential restricted least-squares algorithm. This representation allows to unify the seemingly diverse approaches to diffuse initialization and also sheds light on phenomena such as the transition to the usual Kalman filter. The Kalman filter is a key algorithm in linear time series, e.g. ARIMA models and structural models. However, its application requires the specification of means and (co-)variances for the first element of the series. In case of a stationary process the first moments of the stationary distribution provide natural candidates. Diffuse initialization circumvents the specification of these starting conditions and thereby allows the application of the Kalman recursions to non-stationary processes.