Dienstag, 15.04.2025, 12-13 Uhr in W9-109
Prof. Dr. Dietmar Bauer
Universität Bielefeld
Using Subspace Methods for the Estimation of Approximate Dynamic Factor Models
For multivariate time series with a large number of variables classical vector autoregressive (VAR) models are not appropriate because they contain too many parameters. Alternatively in the literature in such situations factor models are used to reduce the dimensionality. Approximate dynamic factor models represent the high-dimensional time series as generated by a common factor part and idiosyncratic terms, where the common factors are latent. Estimating the dynamics of the common factors often is done using a VAR model for the principal components. The estimation of more flexible state space models via maximum likelihood methods is more complicated. Subspace methods are a numerically stable alternative that can be used in this respect. In this talk we show that the subspace methods provide a very robust and computationally simple means to obtain consistent estimators for the latent factor dynamics.
Dienstag, 29.04.2025, 12-13 Uhr in W9-109
Houda Yaqine
Universität Bielefeld
When to Collect Data? A Sobol Index Strategy for Model-Based Experimental Design
Optimizing measurement timing in experimental studies is crucial for reducing animal sacrifices, especially when investigating systems characterized by ordinary differential equations (ODEs). Conventional sampling strategies frequently collect excessive data points without substantially improving parameter estimation accuracy. Our research introduces a new experimental design framework that analyzes the temporal evolution of variance-based sensitivities in ODE-driven processes. By integrating time-dependent Sobol indices with underlying system dynamics, this approach identifies the most informative sampling points, particularly during phases where parameter interactions are the primary contributors to system variance. Both theoretical analyses and computational experiments confirm that our strategically selected measurement times yield better parameter estimation outcomes compared to traditional protocols.
Dienstag, 13.05.2025, 12-13 Uhr in W9-109
Prof. Dr. Christoph Kayser
Universität Bielefeld
Insights from cognitive neuroscience – statistical models to understand behavior
In this talk I will present studies on how we (and our brain) combine multisensory signals to guide behavior. The talk will focus on examples such as spatial ventriloquism and multisensory perceptual decision making, but also speak more broadly about how well we can explain behavior in typical laboratory tasks. The goal of the presentation is to illustrate the different types of statistical models employed in these studies, including Bayesian approaches and drift-diffusion (random walk) models and how they can be used to understand behavior, but also how we can link them to neural processes in the brain.
Dienstag, 27.05.2025, 12-13 Uhr in W9-109
Lars Knieper
Georg-August-Universität Göttingen
Advances in gradient boosting approaches for geoadditive models
One of the key features of model-based component-wise gradient boosting is a data-driven variable selection mechanism which takes place while estimating the effects simultaneously. When spatial effects, of areal or point-reference data, are included as a potential model-component a drastic increase of chosen fixed effects can be observed. This is accompanied by a high selection frequency of the spatial component without achieving an impactful reduction of the loss function. To address this ineffective variable selection, we propose to eliminate the competition between fixed and spatial effects by separating the spatial part from the component-wise mechanism and ensuring complete estimation in each iteration. Additionally, we suggest using spatial cross-validation, which accounts for the auto-correlated structure of spatial data when constructing folds. Our approach is applied to yield deviation estimates of coffee farmers in Colombia, resulting in a sparser model with improved predictive performance. In addition, there will be an extended outlook on spatial confounding—an estimation bias arising from collinearity between fixed and spatial effects—which tends to be more pronounced in gradient boosting than in maximum likelihood approaches. A straightforward remedy, known as Spatial+, proves effective in mitigating this issue.
Dienstag, 10.06.2025, 12-13 Uhr in W9-109
Prof. Dr. Göran Kauermann
Institut für Statistik der Ludwig-Maximilians-Universität München
More on Uncertainty in Machine Learning
The quantification of uncertainty is continuously gaining interest in the machine learning community. Tackling this question can be done with statistical tools, getting back to the foundation of statistics, namely how to cope with uncertainty and ignorance. The talk extends our previous work on this topic and presents some new results and insights. We discuss different sources of uncertainty and advocate an increased use of statistics and statistical methods in the machine learning world. To showcase our view we look at ambiguity in computer linguistics, we sketch the use of statistics in network weight reconstruction and focus on labelling ambiguity. While these examples are heterogeneous, the data centric view and hence its statistical foundation serves as common thread.
Dienstag, 24.06.2025, 12-13 Uhr in W9-109
Paul Heimerl
Department of Economics and Business Economics der Aarhus University
Estimation of Latent Group Structures in Time-Varying Panel Data Models
We introduce a panel data model where coefficients vary both over time and the cross-section. Slope coefficients change smoothly over time and follow a latent group structure, being homogeneous within but heterogeneous across groups. The group structure is identified using a pairwise adaptive group fused-Lasso penalty. The trajectories of time-varying coefficients are estimated via polynomial spline functions. We derive the asymptotic distributions of the penalized and post-selection estimators and show their oracle efficiency. A simulation study demonstrates excellent finite sample properties. An application to the pace of microclimate change highlights the relevance of addressing cross-sectional heterogeneity and time-variance in empirical settings.
Dienstag, 08.07.2025, 12-13 Uhr in W9-109