Aufgrund der aktuellen Situation greifen wir in diesem Semester auch auf aufgezeichnete Videovorträge zurück. Einige Vorträge sind passwortgeschützt. Bei Interesse am Vortrag, melden Sie sich bitte bei Dr. Nina Westerheide (nwesterheide@uni-bielefeld.de) oder direkt bei der bzw. dem Vortragenden, um das Passwort bzw. den zoom-Link in Erfahrung zu bringen.
Videovorträge
Prof. Dr. Roland Langrock
Universität Bielefeld, Vortrag am 03.07.2017 an der University of Groningen (RUG)
Nonparametric inference in hidden Markov and related models
Hidden Markov models (HMMs) are versatile models for univariate and multivariate time series. In their basic setup, HMMs comprise two stochastic processes, only one of which is directly observed. The observed (state-dependent) process is in some way driven by the unobserved, serially correlated (state) process. In these modelling classes even parametric inference can be computationally challenging, since the likelihood involves the consideration of all possible state sequences that might have given rise to the observations. Nonparametric techniques for flexible modelling, which are nowadays routinely and very successfully used for example in regression or in density estimation, are only beginning to be applied also in HMM-type modelling classes. In this talk, after giving a general introduction to statistical inference for HMMs, I will demonstrate how the powerful HMM machinery can be combined with the general advantages of penalized splines to allow for flexible nonparametric inference in general-purpose HMM-type classes of models, including Markov-switching regression models. The focus of the presentation will lie on practical aspects of nonparametric modelling in these frameworks, with the methods being illustrated in ecological and economic real data examples.
The talk will be held in English.
weitere geplante Vortragstermine (über zoom):
Dienstag, 10.11.2020, 12-13 Uhr - Vortrag über zoom
David Winkelmann, M.Sc.
Universität Bielefeld
Can we beat the market? - Long-term betting strategies in European football
Research on sports betting often attempts to identify biased evaluation by bookmakers, opening opportunities for profitable strategies to bettors. Previous studies have provided evidence for the existence of such inefficiencies. Since most studies cover only a few seasons, the question of whether market inefficiencies persist over time remains unanswered. We analyse the big five leagues in European association football for 14 seasons to detect the occurrence and duration of market inefficiencies. While our results suggest that most biases do not persist for a long time, we still uncover profitable betting strategies throughout the full observation period.
The talk will be held in English.
Dienstag, 24.11.2020, 12-13 Uhr - Vortrag über zoom
Lennart Oelschläger, M.Sc.
Universität Bielefeld
Approximating mixing distributions in probit models via a Bayesian approach
The multinomial probit model is one of the most widely-used statistical models to explain the choices that individuals make among a discrete set of alternatives, which is of central interest in many scientific areas, for example in transportation economics and marketing. In many such choice scenarios it is reasonable to assume, that the preferences of the decision makers are non-homogeneous, which can be modelled by imposing mixing distributions on the coefficients. Currently, the literature does not provide much guidance for the specification of the mixing distributions apart from trial and error procedures using several standard parametric distributions. To fill this gap, this talk presents a new approach that combines Bayes estimation and semi-parametric methods: A Bayesian framework is used for estimating a latent class mixed multinomial probit model where the number of latent classes is updated within the algorithm on a weight-based strategy. Based on simulation results, we show that the approach is capable of approximating the underlying mixing distributions, reproducing the latent classes and thereby guiding the specification of mixing distributions in empirical applications.
The talk will be held in English.
Dienstag, 08.12.2020, 12-13 Uhr - Vortrag über zoom
Dr. Turid Frahnow und Dr. Peter Pütz
Universität Bielefeld
KoCo19: A Corona antibody study in Munich
Starting in the end of 2019, a recently discovered coronavirus (SARS-CoV-2) spread all over the world causing a severe respiratory syndrome with sometimes fatal consequences. Due to its impact on health, economy and daily life, validated information about the virus, the disease (COVID-19) and predictions regarding infection rates are of great interest. In Munich, a population-based serological study was conducted aiming for a better understanding of the pandemic. In this talk, we (on behalf of the whole study team) want to share first results with the audience about SARS-CoV-2 seroprevalence and epidemiological insights into the pandemic.
The talk will be held in English.
Dienstag, 12.01.2021, 12-13 Uhr - Vortrag über zoom
Anja Rappl, M.Sc.
Friedrich-Alexander-Universität Erlangen-Nürnberg
Joint Models in Mean Regression and Beyond
Joint Models for longitudinal and time-to-event data have gained a lot of attention in the last few years, as they are a helpful technique to approach a common data structure in biostatistics where longitudinal outcomes are recorded alongside event times. Those two processes are often linked and the two outcomes should thus be modeled jointly in order to prevent a potential bias introduced. Commonly, joint models are estimated in likelihood based expectation maximization or Bayesian approaches and are focused on the mean trajectory of the longitudinal outcome. Those frameworks lack of straight forward variable selection and do not immediately work for high-dimensional data situations. Gradient boosting is a method from the field of statistical learning which leads to automated variable selection and shrinkage. This talk introduces the boosting framework for joint models, which for the first time allows to estimate joint models in high-dimensional data situations. In addition to the commonly known variable selection feature, an extension which allows for allocation of the variables to the correct part of the model will be presented. Furthermore a Bayesian approach for extensions towards modelling different features beyond the mean of the longitudinal part of the model will be discussed.
The talk will be held in English.
Dienstag, 26.01.2021, 12-13 Uhr - Vortrag über zoom
Sina Mews, M.Sc.
Universität Bielefeld
Continuous- time state-space modelling of delinquent behaviour in adolescence and young adulthood
Using data from a longitudinal study on delinquent behaviour of adolescents in Germany, we investigate the persistence of an individual’s delinquency level over time. We assume the latter to be a latent trait underlying the observed trajectories of adolescents' delinquency, thus using a state-space model (SSM) to analyse the data. As the observations are irregularly spaced in time, we formulate the SSM in continuous time and specify the state process as an Ornstein-Uhlenbeck process. We further include the adolescents’ gender and age as covariates in the observation process. Statistical inference is carried out by maximum approximate likelihood estimation, where multiple numerical integration within the likelihood evaluation is performed via a fine discretisation of the state process. The corresponding reframing of the SSM as a continuous-time hidden Markov model enables us to apply the associated efficient algorithms for parameter estimation and state decoding. The results reveal temporal persistence in the deviation of an individual's delinquency level from the population mean.
The talk will be held in English.
Dienstag, 09.02.2021, 12-13 Uhr - Vortrag über zoom
Prof. Dr. Axel Mayer
Universität Bielefeld
Analyzing Average and Conditional Effects with Non-Linear Structural Equation Modeling
We present an approach for the analysis of the effects of interventions based on nonlinear structural equation mixture modeling (NSEMM). It extends the traditional moderated regression approach to include latent continous and discrete (mixture) variables as well as their higher order interactions, quadratic or more general nonlinear relationships. We combine the recently proposed EffectLiteR approach for analyzing average and conditional effects and NSEMM to provide applied researchers with new possibilities to evaluate the effectiveness of an intervention or a treatment. In the proposed synthesis of the two approaches, we extend the EffectLiteR approach to models with conditional nonlinearities and non-normalities and show how the NSEMM approach can be used to compute various kinds of effects of interest. The approach is illustrated by an example from the educational sciences. We use the R package blavaan to fit the Bayesian NSEMM model and compute the effects of interest based on the fitted model.
The talk will be held in English.