The interdisciplinary research area FAIR organizes a two-day workshop on Using Resampling and Simulation to Tackle Heterogeneity in Social Science Research "in cooperation with the Research Center Trustworthy Data Science and Security".

This workshop will introduce participants to a range of topics, including simulation techniques, resampling procedures, multivariate data analysis, regression analysis, handling missing values, and imputation techniques.

Our invited speakers are:

• Moritz Berger (University of Bonn)
• Dennis Dobler (Vrije Universiteit Amsterdam)
• Florian Dumpert (Federal Statistical Office of Germany)
• Sarah Friedrich (University of Augsburg)
• Frank Konietschke (Charité – Universitätsmedizin Berlin)
• Łukasz Smaga (Adam Mickiewicz University)

When and where

Date and Time: September 22 and 25, 2023. 9:00-14:30 CET
Location: TU Dortmund University, room: Otto-Hahn-Strasse 12, E.003
Virtual: Online in Zoom (a link will be sent only to registered participants by email) 

How to register

Registration is necessary both for on-site and online participation. In order to register for this workshop, please fill out the following online form: https://umfragen.tu-dortmund.de/index.php/386944?lang=en 

The extended deadline for registration is 19th September 2023.

Invited Speakers

Parametric and Nonparametric Modeling of Dispersion Effects in Ordinal Regression

Abstract: The proportional odds model is probably the most widely applied ordinal regression model. It postulates that the effects of the explanatory variables on the outcome are the same across all categories. This facilitates the interpretation of model parameters in terms of cumulative odds, which makes it very attractive for practitioners. In many applications, however, the assumptions of the proportional odds model are too restrictive. In particular, if the outcome shows differing variability in subgroups of the population (also referred to as dispersion) the proportional odds model will typically show poor goodness-of-fit. An extended model that accounts for heterogeneity of variances is the location-scale model already considered by McCullagh (1980) also known as heteroscedastic logit model. More recently, we proposed the so-called location-shift model as an alternative to account for varying dispersion. The location-shift model contains the familiar location term (as in the proportional odds model) complemented by a linear term that represents variability of the outcome. Parameters of the location-shift model can be interpreted straightforward in terms of log odd ratios. Moreover, it can be embedded into the framework of multivariate generalized linear models.

In this talk, by several real-world applications (e.g. data of the general social survey of social science; ALLBUS) we demonstrate that the location-shift model frequently shows satisfactory goodness-of-fit while being comparably sparse in parameters. In addition, we introduce the class of additive location-shift models allowing for smooth effects, as well as a tree-structured extension of the location-scale model accounting for interactions between the explanatory variables.

Randomization-based Inferece in Nonparametric Repeated Measure Models with Missing Data

Abstract: Relative effects enjoy great popularity across various field of research. Also in statistical methodolgy research, extensions of this method have been developed in many different directions. In this talk, we will focus on repeated measures designs with randomly missing data. Here, relative effects can be used to find a time or other effects on the outcomes. In a previous SMMR-paper by Rubart, Pauly, and Konietschke (2022), relevant theory has been developed for tests based on quadratic forms in this context. They developed Wald- and ANOVA-type tests that are based on approximations using estimated chi-squared and F-distributions. In this talk, we re-visit that testing problem by means of a randomization procedure which will give rise to asymptotically exact inference procedures. Simulations demonstrate the small sample performance and a real data analysis illustrates several aspects of our method.

Classify and code, review and validate, and edit and impute in German official statistics

Abstract: Official statistics face a variety of challenges. Driven by the increased possibilities of obtaining information and the progress in information technology, the demand for information from politics, the economy and society on the most diverse subject areas of official statistics is increasing. In order to meet this demand adequately, the production of statistics must be further developed. This is not only a matter of making new data sources usable, but also of making processes more efficient. This applies in particular to steps in the area of data processing, such as classify and code, review and validate, and edit and impute. Solutions for (partial) automation of the processing steps are therefore being tested and used by Destatis.

On the role of benchmarking data sets and simulations in method comparison studies

Abstract: Method comparisons are essential to provide recommendations and guidance for applied researchers, who often have to choose from a plethora of available approaches. While many comparisons exist in the literature, these are often not neutral but favor a novel method. Apart from the choice of design and a proper reporting of the findings, there are different approaches concerning the underlying data for such method comparison studies. Most manuscripts on statistical methodology rely on simulation studies and provide a single real-world data set as an example to motivate and illustrate the methodology investigated. In the context of supervised learning, in contrast, methods are often evaluated using so-called benchmarking data sets, that is, real-world data that serve as gold standard in the community. Simulation studies, on the other hand, are much less common in this context. In this talk, we will investigate differences and similarities between these approaches, discuss their advantages and disadvantages, and ultimately develop new approaches picking the best of both worlds.

Max – t Test in high dimensional Repeated Measures Designs

Abstract: In many experiments and especially in translational and preclinical research, sample sizes are (very) small. In addition, data designs are often high dimensional, i.e. more dependent than independent replications of the trial are observed. In this talk, we discuss the use of max t-tests (also known as multiple contrast tests) in high dimensional repeated measures and multivariate designs. We hereby relax the usual but rather strict assumptions of multivariate normality and/or equal covariance matrices across the different (independent) samples. We derive the limiting distribution of the test statistics and propose a wild bootstrap approach as approximate solution for small sample sizes. Extensive simulation studies indicate that the test controls the type-1 error very well even when sample sizes are as small as 10. A real data set illustrates the application of the methods.

References:

Konietschke, F., Schwab, K, Pauly, M. (2020). Small sample sizes: A big data problem in high-dimensional data analysis. Statistical Methods in Medical Research DOI: 10.1177/0962280220970228

Lin, Z., Lopes, M. E., & Müller, H. G. (2023). High-dimensional MANOVA via bootstrapping and its application to functional and sparse count data. Journal of the American Statistical Association, 118(541), 177-191.

Introduction to Functional Data Analysis: Modeling, Inference, and Real-Data Applications

Abstract: Functional data analysis (FDA) is a branch of statistics that focuses on modeling and making inferences from observations treated as functions, curves, or surfaces. It provides a powerful approach for analyzing one or more variables measured over time or space, which is a common scenario encountered in various fields. For example, we can consider the measurement of temperature every minute at multiple weather stations over a specific period in a country. Each weather station's temperature values recorded at time points can be modeled as a function, representing a single functional observation. Utilizing a functional representation of time series data helps overcome challenges associated with classical multivariate statistical techniques, such as the curse of dimensionality and missing data. Over the past two decades, a wide range of data analysis methods have been developed specifically for functional data. These methods encompass classification, clustering, dimension reduction, outlier detection, regression, and statistical hypothesis testing. During the presentation, we will explore different aspects and strategies employed in functional data analysis techniques. In particular, we consider the resampling methods for hypothesis testing problems and regression. The presentation will feature real data examples prepared using the R programming language to illustrate the problems and methodologies discussed.

Our Partners

This year's Workshop is offered in cooperation with the Research Center Trustworthy Data Science and Security.