Professur für Raumbezogene Datenanalyse und Statistische Lernverfahren









Drei Studierende, zwei Frauen und ein Mann, sitzen in einem Kreis vor dem Zentralen Lern- und Studiengebäude und unterhalten sich.










RESEARCH




Our research focuses on different types of regression models and their estimation methods, such as Bayesian approaches and model-based gradient boosting.
















OUR CURRENT RESEARCH SPOTLIGHTS












  • General Scope of Research





    General Scope of Research


    In general, we are developing statistical methods in the area of spatial temporal modelling. The methodological background includes classical Bayesian and likelihood based methods, as well as statistical learning methods or neural networks. We further cooperate with several researchers in more data driven areas, such as the Cystic Fibrosis Institute Germany, a research group on tinnitus in the center of Biomagnetism Erlangen or different groups at the university here in Göttingen. Below, we describe a bit of our methodological research. Feel free to contact us, if you want to learn more!












  • Joint Modelling of Longitudinal and Time-to-Event Data





    Joint Modelling of Longitudinal and Time-to-Event Data


    In many research areas, longitudinal and time-to-event data is recorded and then modeled separately, based on overlapping covariates. Joint modelling aims at reducing the bias induced by separately estimation of two different regression models.
    In our group we work in different parts of the development of advances in joint modelling. We for example introduced different types of statistical learning methods to facilitate variable selection and allocation. We further introduced a way of linking covariate dependent quantiles instead of the mean to the time-to-event outcome and work on remedies to prevent from biased estimation in sparse data situations. We also aim at transferring joint models from life sciences, where they are mainly used, to social sciences, where they are currently underused.












  • Time- and Space in Statistical Learning methods





    Time- and Space in Statistical Learning methods


    Random as well as Spatial effects have a long standing history in Bayesian inference as well as likelihood based methods. They were transferred in a straight forward way to statistical learning methods such as gradient boosting. This way of introducing them to an approach, which does not capture information on the whole distribution of the regression parameters leads to severe problems in the variable selection process and gives those effects, which are in many cases meant to be an informative error structure, a weight, which it should not have. By different techniques, like disentanglement or  decorrelation, we seek to avoid those problems while still maintaining the advantageous characteristics of the methods, such as reliable variable selection and prediction properties.











  • Steplength Selection in Gradient Boosting for Generalized Additive Models for Location, Scale and Shape





    Steplength Selection in Gradient Boosting for Generalized Additive Models for Location, Scale and Shape


    In model based gradient boosting, which is an iterative method, updates are made by small steps, usually regularized with a fixed value of e.g. 0.1. When implementing gradient boosting for models, which estimate the influence of covariates on different parameters of the outcome distributions, the likelihood is evaluated on different scales. It is hence not possible to compare update steps for different parameters fairly, when using the same step-length. We develop different ways of choosing fair and adaptive step-lengths for those algorithm. The problem is tackled both numerically and analytically




















OUR RESEARCH OUTPUT





















Publications








External Publications