Tuesday, September 22, 3.15 pm (OBS! CHANGED TIME), 2026. Seminar in Statistics.

Dependent Censoring: An Ignored but Not Ignorable Problem in Survival Analysis
Ewa Wycinka, Department of Statistics, Faculty of Management, University of Gdańsk
Abstract: Survival analysis is the primary statistical framework for studying time-to-event data and has applications across medicine, epidemiology, engineering, economics, and computer science. A defining feature of such data is censoring, whereby the event time is only partially observed. Standard methods, including the Kaplan-Meier estimator and Cox proportional hazards model, rely on the assumption that the censoring mechanism is independent of the event process. Although routinely adopted, this assumption is often difficult to justify in practice and may be violated in many real-world settings.
The presentation begins with an overview of the fundamental concepts of survival analysis and its principal fields of application. It then examines censoring from four complementary perspectives, providing a structured framework for understanding the different mechanisms leading to incomplete observations. Particular attention is devoted to dependent censoring, in which the censoring process is associated with the underlying event time. Common sources of such dependence and their consequences for survival estimation, regression modeling, prediction, and statistical inference are discussed.
A central challenge is that dependent censoring is generally not testable from the observed data alone. Nevertheless, under certain conditions its presence may be identified or its plausibility assessed using study-specific knowledge, auxiliary information, external data sources, or explicit modelling assumptions. The opportunities and limitations of these approaches are examined.
The final part of the presentation reviews statistical methods designed to address dependent censoring, including inverse probability of censoring weighting, joint modelling, copula-based approaches, and sensitivity-analysis techniques. Emphasis is placed on the assumptions underlying these methods, their practical applicability, and the extent to which they can mitigate bias arising from informative censoring.

Location: Alan Turing .

Tuesday, October 6, 3.15 pm, 2026. Seminar in Mathematical Statistics.

Recent Advances in High-Dimensional Data Modeling: From Variable Selection and Machine Learning to Post-Selection Shrinkage and Statistical Inference
Reza Arabi Belaghi, Department of Energy and Technology, Swedish University of Agricultural Sciences
Abstract: The rapid growth of high-dimensional data in genomics, biomedical research, environmental sciences, and other scientific disciplines has created statistical problems in which the number of potential predictors can be comparable to, or substantially larger than, the sample size. Although regularization methods such as LASSO and Elastic Net have transformed the analysis of such data, important challenges remain. In particular, variable selection can become unstable in the presence of highly correlated predictors, and procedures designed to produce sparse models may overlook predictors with weak but non-negligible signals. Consequently, selecting a sparse model is not necessarily the end of the statistical problem: how we estimate and make use of the information remaining after selection is equally important.
In this talk, I will discuss recent developments in high-dimensional statistical modeling from the complementary perspectives of variable selection, machine learning, and post-selection estimation. I will first examine the behaviour of classical penalized approaches such as LASSO and Elastic Net and contrast them with machine-learning-based screening methods such as Boruta. Particular attention will be given to the effects of dimensionality, predictor correlation, sample size, and signal strength on the ability of these methods to identify relevant variables.
I will then move from the question of which variables should be selected? to the equally important question of what should be done after selection? A central theme will be the distinction between strong, weak-to-moderate, and null signals. Rather than automatically discarding variables that fail to survive an initial sparse selection procedure, post-selection shrinkage methods can retain and appropriately down-weight information carried by weaker predictors. I will discuss recent extensions of this idea to high-dimensional generalized linear models and Cox proportional hazards models, including weighted-ridge and Stein-type shrinkage strategies designed to improve estimation and prediction in the presence of weak signals and multicollinearity.
Finally, I will discuss a broader "screen-then-shrink" perspective that combines machine-learning screening with statistical shrinkage. Using high-dimensional beta regression as an example, I will illustrate how Boruta-based screening can first reduce an ultra-high-dimensional feature space, after which ridge and Stein-type shrinkage can be used to stabilize estimation among correlated and potentially weak predictors. This provides a bridge between modern machine-learning feature discovery and classical statistical estimation and inference.
Throughout the talk, theoretical ideas will be connected with simulation studies and applications to biomedical and genomic data. The overall message is that high-dimensional modeling should not be viewed simply as a search for the smallest predictive model. Effective analysis requires careful consideration of correlation, weak signals, selection uncertainty, prediction, and post-selection estimation. I will conclude by discussing open methodological questions and potential directions for integrating machine-learning screening, regularization, shrinkage, and interpretable statistical inference in modern high-dimensional data analysis.
Location: BL34

Tuesday, October 20, 3.15 pm (OLD TIME), 2026. Seminar in Statistics.

Adaptive approaches to modeling random functions
Valentin Patilea, CREST ENSAI
Abstract: In many applications, data can be viewed as sample paths of a random function (stochastic process) defined on some bounded domain. A large body of work, commonly known as functional data analysis, has developed nonparametric methods for this type of data, where the observation unit is a curve or surface that is discretely observed, possibly with noise. However, very few contributions have addressed the problem of adaptivity to the regularity and structure of the data - an important practical concern, as data-driven methods free the analyst from the burden of tuning parameters by hand. In this talk, several recent advances in letting functional data speak for themselves will be presented. The methods covered apply to mean and covariance function estimation and inference, functional PCA, and time series forecasting. Throughout, the focus will be on random design settings.
Location: Alan Turing .

Tuesday, November 3, 1.30 pm (NEW TIME), 2026. Seminar in Statistics.

Latent-Variable Learning of SPDEs via Wiener Chaos
Sebastian Zeng, Department of Mathematics and Physics, Linnæus University
Abstract:
We study the problem of learning the law of linear stochastic partial differential equations (SPDEs) with additive Gaussian forcing from spatiotemporal observations. Most existing approaches either assume access to the driving noise or initial condition, rely on deterministic surrogates that fail to capture intrinsic stochasticity, or employ generative models (e.g., diffusion models) that capture local randomness but may accumulate distributional errors when used in an autoregressive rollout setting. We propose a structured latent-variable formulation that requires only observations of solution realizations and learns the underlying randomly forced dynamics. Our approach combines a spectral Galerkin projection with a truncated Wiener chaos expansion, yielding a principled separation between deterministic evolution and stochastic forcing. This formulation reduces the infinite-dimensional SPDE to a finite system of parametrized ordinary differential equations governing latent dynamics. The latent dynamics and stochastic forcing are jointly inferred through variational learning, enabling recovery of stochastic structure without explicit observation or simulation of noise during training. Empirical evaluation on synthetic data demonstrates state-of-the-art performance under comparable modeling assumptions on bounded and unbounded one-dimensional spatial domains. Location: Alan Turing .

Tuesday, December 1, 1.30 pm (NEW TIME), 2026. Seminar in Statistics.

TBA
Maria Karlsson, Umeå School of Business, Economics and Statistics, Umeå University
Abstract:
Location: Alan Turing .