Statistical Colloquium/TRR Research Seminar

PhD Student Colloquium

Title: Designing experiments on networks

Vasiliki Koutra, King's College London

The design of experiments provides a formal framework for the collection of data to aid decision making. When such experiments are performed on connected units linked through a network, the resulting design and analysis are more complex; e.g. is the observed response from a given unit due to the direct effect of the treatment applied to that unit or the result of a network, or viral, effect arising from treatments applied to connected units? In this talk, I propose a methodology for constructing efficient designs which control for variation among the experimental units arising from network interference, so that the direct treatment effects can be precisely estimated. Performance gains over conventional designs will be demonstrated via different example experiments.

29.05.2024 at 14:15, TU Dortmund University, M/ E 21

Title: Backtesting: two applied tales about prediction

Prof. Dr. Ivan Mizera, University of Alberta

Two stories about prediction under uncertainty in the applied setting are told. The first concerns the task of prediction of commercial breaks in targeted TV advertising via stochastic means, a research of industrial character realized in the collaboration with the Edmonton based company INVIDI. The second addresses the important risk-assessment task in non-life insurance, the prediction of the financial reserves guaranteeing the payment of existing and potential future insurance claims via so-called run-off triangles. What both stories have in common is their highly nonparametric, model-free character, with virtually no existing approaches to build on; and their empirical vindication accomplished via so-called backtesting on historical data.

04.06.2024 at 10:15 a.m., TU Dortmund University, M/ E 25 (!!)

Title: Modeling Dynamic Interaction Networks Using Counting Processes

Alexander Kreiß, Junior Professor for Statistics, Leipzig

In an interaction network one observes a network of vertices and edges. Two vertices can interact when they are connected by an edge, where an interaction is understood as an instantaneous event. A typical example would be users of a social media platform (the vertices) who interact by sending messages to each other when they are in a friendship relation (edges). In a dynamic interaction network we allow that the edges may appear and disappear over time. We suppose to observe the edges and the interactions in a given time period. In addition, we observe covariate processes which describe the network, individual vertices and the relations between pairs of vertices. In the talk we will review parametric and nonparametric models for this type of data based on counting process theory. We will show how these models can be used to incorporate global covariates, perform parameter estimation and goodness of fit tests. Mathematically, a challenge in these models is the dependence between pairs of vertices. We will discuss assumptions under which this dependence can be handled. Finally, we will also look at a real-world dataset on rental bikes to illustrate the models.

05.06.2024 at 14:15, University of Duisburg-Essen, room R11 T08 C01 in building R11 (!!)

Title: Inference in Regression Discontinuity Designs with High-Dimensional Covariates

Alexander Kreiß, Junior Professor of Statistics, Leipzig

We study regression discontinuity designs in which many predetermined covariates, possibly much more than the number of observations, can be used to increase the precision of treatment effect estimates. We consider a two-step estimator which first selects a small number of 'important' covariates through a localized lasso-type procedure, and then, in a second step, estimates the treatment effect by including the selected covariates linearly into the usual local linear estimator. We provide an in-depth analysis of the algorithm's theoretical properties, showing that, under an approximate sparsity condition, the resulting estimator is asymptotically normal, with asymptotic bias and variance that are conceptually similar to those obtained in low-dimensional settings. Bandwidth selection and inference can be carried out using standard methods. We also provide simulations and an empirical application. This is joint work with Christoph Rothe.

no seminar

Title: Testing and monitoring speculative bubbles

Jörg Breitung, Institute for Econometrics and Statistics, University of Cologne

We propose a heteroskedasticity-robust LBI statistic to test the hypothesis of a unit root against the alternative of an explosive root related to speculative bubbles. Compared to existing alternatives like Dickey-Fuller type tests, the proposed LBI test has a standard limiting distribution and higher power, especially in the empirically relevant case of a moderate explosive root. Further refinements, such as the point-optimal linear test, come remarkably close to the power envelope. To detect bubbles with an unknown starting date, sequential schemes based on forward and backward expanding windows are considered, with the stacked backward CUSUM procedure proposed by Otto and Breitung (2023) standing out as the most powerful sequential scheme in the homoskedastic case. For the case of time-varying volatility, a heteroskedasticity-robust MOSUM detector is proposed. Finally, we consider simple statistics for consistently estimating the starting date of the bubble.

no seminar

tba

no seminar

no seminar