Todagesmøde efterår 2023




November 29, 2023

Chilly weather, comfortable seats, and thrilling statistics welcome 106 participants to DSTS’s 2023 Autumn two-day meeting in Copenhagen.

The second meeting of 2023 was held at the KU Science facilities on November 7th and 8th and was organized by the MATH department at KU. On the first day the voluminous Lundbeck Auditorium made an ideal venue for delving into presentations and discussion of the latest statistical research in much the same way the audience could delve into the softness of their seats. On the next day, the traditional “Lille UP1” Auditorium at the Department of Computer Science lended its historical setting to the second part of the event. The seats may have been tougher on the behinds, but the more intimate and old-school setting made up for it.

The conference was a great success, with an enriched professional and social programme, including a dinner at Food Club right by the lakes of Copenhagen. The conference featured a wide range of topics, from convergence rates of Markov Chain Monte Carlo chains, through assessment of average bioequivalence, to high-dimensional stochastic processes. The conference provided ample opportunity to learn about cutting edge statistical methods, as well as to make new friends and catch up with old ones. Of the 106 participants, 30 were students, suggesting, like the last time, a strong interest in the two-day meetings from the statisticians of tomorrow. This interest may have been helped by the free Young Statisticians Denmark lunch offered before the event, at the courtesy of the Danish Statistical Society.

Jun Yang started off the official programme by taking advantage of the freshness of the listeners, as he jumped right into formulas and complexity results for Markov Chain Monte Carlo algorithms with his talk “Complexity results for MCMC derived from quantitative bounds”. Jun explained how traditional quantitative bounds for the low dimensional setting do not directly generalize to useful bounds in the high-dimensional setting and then showed that one can form a coupling on a high probability well-behaving subset of the domain and use this technique to calculate a quantitative convergence rate bound.

Anne Helby Petersen continued with her talk “Causal discovery for observational life course studies”. Anne first outlined a traditional method for independence test based causal discovery, namely the PC algorithm and then explained how information from an ordering into temporal tiers can be used to enhance the procedure. For this adjusted algorithm, Anne compared the machine’s output with that of two experts. This gave rise to a discussion of what the machine may offer the health researcher, and vice versa. Finally, Anne gave views on what needs to be done to get rid of the assumption of no unmeasured confounders in applications.

Jonas Wallin followed the break for coffee and cake with his talk “Gaussian Whittle–Matern fields on metric graphs”. Jonas discussed the challenge in specifying a valid covariance function on metric graphs and gave an account of his and his collaborators’ work on doing exactly that. Jonas showed us both real life applications where the data generating mechanism can be viewed as a metric graph, such as traffic networks, and simulations indicating the properties of the approach that he had described.

To round off the talks of the day, Susanne Ditlevsen spoke about “Estimation of time to a tipping point”. First, Susanne talked about what a tipping point actually is and underlined a specific feature of a tipping point, namely that the system can not go back to the state it was in before the tipping point. Next, Susanne described how, due to the physics and math behind the problem, early warning signals such as increased variance of the stochastic process considered to generate the data, can be used to predict the actual time of tipping. Susanne rounded off the presentation by talking about the substantial media attention that her and Peter Ditlevsen’s work has attracted, and how they dealt with the intense situation.

Before ending the day with dinner at Food Club, the organizers of the meeting provided drinks, and the UFP (Udvalg for Formidling og Presse / Committee for Dissemination and Press) offered the opportunity to blind-taste ketchup and discover associations between various background variables and ketchup preference.

On the next day, Christian Pipper brought the lecture hall to life with his talk “Properties of a confirmatory two-stage adaptive procedure for assessing average bioequivalence”. Christian took us through the process of developing a method for testing in a particular two-stage adaptive trial that he and his collaborators had worked with. In particular, Christian discussed the optimizing of power by means of focused testing and simultaneous modeling of the two outcomes considered. Lastly, Christian discussed how certain untraditional choices in the analysis method had to be argued for and discussed with regulators.

Erin Gabriel followed with her talk “Propensity score weighting plus an adjusted proportional hazards model does not equal doubly robust away from the null”. In this talk, Erin brought to light that putting together any combination of propensity score and outcome models does not yield doubly robust estimation in general, and in particular for survival outcomes. Erin showed simulations demonstrating this lack of double robustness away from the null and walked through a proof of double robustness under the null. Further, she showed how this misunderstanding has infiltrated scientific literature and, therefore, has become a serious issue.

Munir Hiabu then took us into the realm of interpretable machine learning with his talk “Unifying local and global model explanations by functional decomposition of low dimensional structures”. Here, Munir discussed how to quantify and illustrate the contributions of each feature and its interactions in a low dimensional feature set for predicted values as obtained from a machine learning model that can be decomposed into a sum of main and interaction components of arbitrary order. More specifically, Munir and his collaborators have proposed a new identification constraint that allows for calculation of interventional SHAP values and partial dependence plots, the former of which is a weighted sum of the main component and all the interaction terms that include a feature, and the latter of which corresponds to the main effects term in the model plus the intercept. This work has also brought a new perspective to SHAP values.

Claudia Strauch ended the programme with her talk “On the statistical analysis of high- dimensional stochastic processes”. Here, Claudia discussed estimation of the drift parameter in high dimensional Lévy-driven Ornstein-Uhlenbeck processes under sparsity constraints and remarked that both the Lasso and Slope estimators achieve the minimax optimal rate of convergence in this case. Claudia also touched upon unresolved challenges and future work within the field of high-dimensional stochastic processes as part of her talk.

Before the participants went their separate ways, the Danish Statistical Society served sandwiches outside the lecture hall. Also, the UFP invited all participants to come and discuss initiatives in dissemination and press for the statistical community at an open meeting, which was held after the conference.

In summary, the scientific conference on statistics held in Copenhagen provided a great platform for researchers and professionals to exchange ideas and discuss the latest developments in many different aspects of the field.

Looking forward to seeing new and familiar faces at the next two-day meeting in Aarhus!

This blogpost was written by Christoffer Sejling, member of DSTS UFP (Udvalg for Formidling og Presse / Committee for Dissemination and Press).

Audience applauding. AI-generated image by Christoffer Sejling