Zentrum für Statistik
Uni von A-Z
Universität Bielefeld > Zentrum für Statistik

4. Nachwuchsworkshop des ZeSt

Am 27. Oktober 2017 findet an der Universität Bielefeld der 4. Nachwuchsworkshop des Zentrums für Statistik statt.

Im Rahmen des Nachwuchsworkshops stellen sich Doktorandinnen und Doktoranden der am Zentrum für Statistik beteiligen Bereiche gegenseitig ihre Forschungsfelder vor und diskutieren darüber. Der Nachwuchsworkshop findet von 14:00-17:30 Uhr in W9-109 statt.

Bei Interesse melden Sie sich bitte per E-Mail bei Nina Westerheide.

Programm des Nachwuchsworkshops


Vorträge der teilnehmenden Doktorandinnen und Doktoranden (in alphabetischer Reihenfolge)

Timo Adam
Universität Bielefeld, Fakultät für Wirtschaftswissenschaften

Gradient boosting in Markov-switching generalized additive models for location, scale and shape

We propose a novel class of flexible latent-state time series regression models which we call Markov-switching generalized additive models for location, scale and shape. In contrast to conventional Markov-switching regression models, the presented methodology allows to model different state-dependent parameters of the response distribution - not only the mean but also variance, skewness and kurtosis parameters - as potentially smooth functions of a given set of explanatory variables. In addition, the set of possible distributions that can be specified for the response is not limited to the exponential family but additionally includes, for instance, a variety of Box-Cox-transformed, zero-inflated and mixture distributions. We propose an estimation approach based on the EM algorithm and demonstrate how statistical boosting can be exploited to prevent overfitting while simultaneously performing variable selection. The feasibility of the suggested approach is illustrated in a real-data setting, where we model the conditional distribution of the daily average price of energy in Spain over time.


Manuel Batram
Universität Bielefeld, Fakultät für Wirtschaftswissenschaften

Reconsidering hybrid models for joint modelling of time-trade-off and discrete-choice valuation experiments

Health-related quality of life (HRQoL) instruments play a crucial role in economic evaluation of health technologies and resource allocation in health care systems. In order to ensure those capabilities of generic instruments, the construction of a tariff or value set is of paramount importance. Currently, time-trade-off (TTO) and discrete choice experiments (DCE) are among the most widely used and accepted elicitation methods in HRQoL research and efforts have been made to combine their datasets in the estimation of the EQ-5D-5L value sets (e.g. in UK and Germany). However, the first installments of hybrid models ignore issues arising from within-person correlation and the current assumptions on the distributions of the error terms make it hard to derive an utility function encompassing both ?parts? of the hybrid model. Starting from the above observations we develop an econometric specification of a hybrid model with random effects, which we estimate using the maximum likelihood method. We use simulated data, mimicking the data from current EQ-5D-5L valuation studies, to compare this new estimator to standard methods as well as hybrid estimators without random effects. We find that the hybrid model with random effects has the lowest error with regard to parameter recovery. Our preliminary results suggest that hybrid models in general and our more flexible RE-hybrid model in particular have the potential to (I) improve value set estimation and (II) help understand the influence of different error sources.


Anke Erdmann
Universität Bielefeld, Fakultät für Soziologie

The victim-offender-overlap from a life course perspective

Contrary to the traditional assumption of victims and offenders being two distinct groups, criminological research of the past five decades suggests otherwise: Both groups feature similar characteristics and the risk of becoming a victim or an offender is influenced by the same factors. Consequently victims and offenders are often one and the same person. This so-called victim-offender-overlap is currently under discussion in criminology. First, longitudinal empirical research on victimization is rather neglected in comparison to delinquency. Second, the overlap between victimization and offending over the life course requires further investigation. My dissertation project focuses on the development of the relationship between victimization and delinquency. In this presentation, I will give an overview on the work that I have done so far and a prospect on the work in progress. First, I will briefly discuss the different theories for explaining the overlap between victims and offenders. Afterwards, I will present analyses done on the relationship between violent victimization and violent delinquency. Finally, I will discuss current/planned analyses that use Group Based Trajectory Modeling resp. Dual Trajectory Modeling to analyze the over- lapping of victimization and delinquency more detailed.


Jan Klostermann und Daniel Böger
Universität Bielefeld, Fakultät für Wirtschaftswissenschaften

Deriving consumers' perceptions from user-generated content on Instagram

The increasing use of social media services has led to an enormous amount of content that is shared every day. To benefit from this development, marketers can extract brand perceptions from user-generated content to listen to what consumers disclose about brands. In this regard, brand-related usergenerated content was shown to reveal insights into what contributors think and feel about brands in prior research. One major issue of user-generated content is the missing information about characteristics of the anonymous content creators. In the context of market segmentation, a common task is to differentiate according to consumers and non-consumers, as these subgroups require different marketing efforts. Against this background, this talk presents a first idea to use image data from Instagram to classify actual brand consumers and non-consumers.


Marius Ötting
Universität Bielefeld, Fakultät für Wirtschaftswissenschaften

Evaluating fraud detection systems in sports betting using bivariate distributional regression

In this talk, the main strategy used by fraud detection systems to detect fixed matches in soccer is presented. Therefore, betting odds stated by bookmakers are compared to those estimated by a statistical model. In order to estimate betting odds, we use a generalized additive models for location, scale and shape (GAMLSS) with bivariate Poisson distribution to model the number of goals scored by both teams as well as the correlation between these two, and subsequently derive the corresponding odds. Thus, we account for possible smooth functions of a given set of explanatory variables. Furthermore, we estimate our model via a boosting algorithm to prevent overfitting and perform automatically variable selection.We apply this approach to data from the Italian Serie B, since for this league there are several matches which have been proven to be fixed.


Jennifer Pohle
Universität Bielefeld, Fakultät für Wirtschaftswissenschaften

Dominant processes in multivariate hidden Markov models for electronic health record data

Electronic health record data contain important information about the health state of a hospitalized patient. They include the patient?s clinical findings, the sequences of measured vital signs and lab test results. Analyzing these data aims to gain a better understanding of the patient?s course of disease. Furthermore, it can help physicians to identify critical developments in time and to take appropriate countermeasures. Our dataset contains hourly heart rate, respiratory rate and blood pressure sequences of 842 patients in the intensive care unit. Hidden Markov models (HMMs) seem to be a natural choice for analyzing these data. They are flexible time series models for observations which are driven by an underlying, serially correlated hidden state sequence. At each time point, the current state determines the distribution of the observations. In the clinical context, the hidden states can be seen as proxies for the clinical condition of the patient. This seems very appealing. However, the joint modeling of the vital signs leads to an undesirable behavior: For all but one variable, the fitted state-dependent distributions do not differ across states. Hence, in the joint model, a single variable dominates the whole state switching process. In my talk, I would like to introduce this problem and to discuss first ideas to overcome it.



Vortrag im ZeSt

Am 29.05.2018 spricht Prof. Dr. Göran Kauermann vom Institut für Statistik der Ludwig-Maximilians-Universität München im Rahmen des Kolloquiums des ZeSt. Der Vortrag findet zwischen 12:00 und 13:00 Uhr in Raum W9-109 statt. Weiter....

5. Nachwuchsworkshop des ZeSt vom 25. bis 27. Juni 2018

Weitere Informationen zum Nachwuchsworkshop finden Sie hier.

Bewerbungszeitraum Masterstudiengang Statistische Wissenschaften

Die Bewerbungsfrist für das Wintersemester 2018/19 beginnt am 1.6.2018 und endet am 15.7.2018. Weiter...