Proceedings

Invited Talks

Nils L. Hjort: Counting the Uncounted: How Many Were Killed in Guatemala, 1978-1996?

Abstract: In various application domains, there is a certain `null cell', inside a multinomial setup, where observations are recorded for the other cells, but where one cannot count the number of occurrences for the null cell. I develop inference theory for assessing such unknown numbers, counting the uncounted, in situations where counts are available for the other cells, via parametric modelling. The methods are used to estimate the number of persons killed in Guatemala during the Genocidio guatemalteco years 1978--1995. There are three carefully curated lists of killed people, where the information can be mapped to a Venn diagram with 2³=8 cells. Summing over the seven observed cells, R=47,803 killed individuals can be identified, but how big is N_0,0,0, and hence N=N_0,0,0+R?

Mats J. Stensrud: Optimal decision rules with policy-relevant guarantees

Abstract: Policy makers often wish to use data to decide how treatments, interventions, or policies should be assigned to individuals. For example, the current focus on precision medicine reflects the search for individualized decision rules, adapted to a patient's characteristics. In this presentation, I will study how to define, choose, and estimate effects that inform individualized decisions. A central difficulty, common to most existing approaches, is that as we include more covariates and aim for finer personalization, the required assumptions become stronger. As an alternative, I propose a strategy for detecting and estimating group-level effects, with statistical guarantees that the estimated groups truly differ. I then show that, in realistic settings, group-based decision rules can substantially outperform state-of-the-art optimal-regime methods, even when those methods rely on correctly specified models and are implemented with modern doubly robust machine-learning estimators

Susanne Ditlevsen: Estimating Tipping Points in Climate

Abstract: In recent years there has been an increasing awareness of the risks of collapse or tipping points in a wide variety of complex systems, ranging from human medical conditions, pandemics, ecosystems to climate, finance and society. Even in systems where governing equations are known, such as the atmospheric flow, predictability is limited by the chaotic nature of the system and by the limited resolution in observations and computer simulations. These phenomena are naturally modelled by strongly nonlinear stochastic processes, which permit a statistical description. In this talk I will present methods to analyze data from such complex systems, with application to an important tipping element in the climate, the Atlantic Meridional Overturning Circulation.

Małgorzata Bogdan: Identifying Low-Dimensional Patterns Using Convex Non-Differentiable Regularizers

Abstract: Convex regularization methods are widely used in modern statistics and machine learning, with applications ranging from genomics and signal processing to finance and network analysis. Although they are often introduced as tools for producing sparse models, their role is much broader. Many convex non-differentiable penalties can reveal meaningful structure in high-dimensional parameters, such as groups of variables with similar effects, smooth or piecewise-constant signals, or clusters of related coefficients. In this talk, I will discuss how such structures can be understood through the general notion of a pattern. A pattern describes the low-dimensional structure imposed by a regularizer; it may correspond, for example, to sparsity or to equality relations among model parameters. I will explain how this notion can be characterized geometrically through the subdifferential of the penalty. This perspective provides insight into why regularization can improve estimation: it not only reduces the apparent complexity of the problem, but also shrinks the estimator toward a simpler and more interpretable lower-dimensional structure. I will also discuss recent theoretical results on estimation accuracy and model selection consistency, and explain how pattern recovery differs from the more classical goals of accurate estimation or prediction. The analysis requires tools beyond standard asymptotic theory and leads to conditions closely related to irrepresentability-type phenomena. Examples based on the fused LASSO and SLOPE will illustrate how convex regularization can recover meaningful structure beyond variable selection alone.

Göran Kauermann: 30 years of P(enalized) Splines

Abstract: We give a very personal view on 30 years of penalized regression, which started with the paper by Eilers & Marx (1996). As a contemporary witness, we can report on the early years of P-splines, and like a kaleidoscope, we aim to explain how and why penalized regression became so successful. A few examples illustrate the flexibility of the approach.

Oral Presentations

Thomas Kneib: Bayesian Penalized Transformation Models

Authors: Johannes Brachem, University of Göttingen; Paul F. V. Wiemann, The Ohio State University; Thomas Kneib, University of Göttingen

Abstract: Penalized transformation models (PTMs) are a semiparametric loca\-tion-scale regression family that estimates a response's conditional distribution directly from the data, and models the location and scale through structured additive predictors. The core of the model is a monotonically increasing transformation function that relates the response distribution to a reference distribution. The transformation function is equipped with a smoothness prior that regularizes how much the estimated distribution diverges from the reference. Markov chain Monte Carlo inference for PTMs offers straightforward uncertainty quantification for the conditional distribution as well as for the covariate effects. Applications to the Fourth Dutch Growth Study and the Framingham Heart Study illustrate the usage and practical utility.

Johannes Brachem: Interpretable Transformation Models for Distributional Regression

Authors: Johannes Brachem, University of Göttingen; Thomas Kneib, University of Göttingen

Abstract: Flexible distributional regression models allow covariates to influence higher-order features of a response distribution, but interpretability often deteriorates as flexibility increases. We propose a transformation-based distributional regression model that yields a principled decomposition of covariate effects. The conditional response distribution is represented through a standardized latent variable whose distribution is modeled via a covariate-dependent monotone transformation, while structured additive predictors describe location and log-scale effects. By anchoring the first two moments of the latent distribution, the model separates location, scale, and higher-order shape components. We address the associated identifiability challenges and develop a scalable Bayesian estimation strategy based on variational Gaussian approximations.

Katharina Ammann: Non-Homogeneous Markov-Switching Distributional Regression Models

Authors: Katharina Ammann, Bielefeld University; Timo Adam, Bielefeld University; Jan-Ole Fischer, Bielefeld University

Abstract: We extend Markov-switching generalised additive models for location, scale, and shape (GAMLSS) by allowing covariates to influence not only the distributional parameters but also the transition probabilities via P-splines. In contrast to conventional Markov-switching GAMLSS approaches, which assume time-homogeneous state transitions, our framework models regime dynamics as smooth functions of covariates within a unified penalised likelihood approach with automatic smoothness selection. This added flexibility enables a data-driven characterisation of how external factors shape both the distributional properties of the response and the evolution of latent regimes over time. An application to Spanish energy prices demonstrates how oil prices and exchange rates affect distributional characteristics and regime-switching behaviour.

Jack Moore: Penalised Piecewise Exponential Distributional Regression Model for Time-to-Event Analysis

Authors: Jack Moore, University of Limerick; Andreas Groll, TU Dortmund University; Kevin Burke, University of Limerick; Shirin Moghaddam, University of Limerick

Abstract: Within the field of time-to-event analysis, traditional parametric approaches, such as the Weibull distribution, impose strong assumptions on the hazard shape. In contrast, the piecewise exponential model (PEM) offers a general alternative capable of approximating any survival distribution by partitioning the time axis into intervals within which the hazard is assumed to be constant. Despite this versatility, the PEM has historically been underutilised, partly due to the dominance of the Cox model, though recent years have seen renewed interest. We introduce a distributional regression extension of the PEM that enables flexible modelling of both the baseline hazard and covariate effects, with both interval and variable selection performed automatically via an adaptive lasso penalisation scheme.

Davide Benussi: Focused median bias reduction and post-FIC-selection inference

Authors: Davide Benussi, University of Padova; Ioannis Kosmidis, University of Warwick; Alessandra Salvan, University of Padova; Nicola Sartori, University of Padova

Abstract: We propose a general method for explicit median bias correction of estimators that are functions of maximum likelihood estimators, based on Cornish-Fisher expansions for Wald statistics. Among the numerous applications, the proposed correction can be combined with the hull-based confidence (HulC) interval method to yield post-FIC-selection intervals for a focus parameter with accurate coverage guarantees in finite samples.

Philipp Sterzinger: Maximum softly penalised likelihood for Factor Analysis

Authors: Philipp Sterzinger, London School of Economics and Political Science; Ioannis Kosmidis, University of Warwick; Irini Moustaki, London School of Economics and Political Science

Abstract: Estimation in exploratory factor analysis often yields Heywood cases, which are characterised by non-positive variance estimates and can cause numerical instability, convergence failures, and misleading inferences. We derive conditions that guarantee existence of maximum penalised likelihood estimates, and that the estimators maintains desirable asymptotic properties expected by the maximum likelihood estimator, namely consistency and asymptotic normality. Our maximum softly-penalised likelihood framework further ensures that the resulting estimation and inference procedures are asymptotically optimal. We demonstrate the performance of our MSPL estimators through simulations.

Diego Tovar-Rios: Partitioned Pseudo-Likelihood for Scalable Joint Modelling of Multivariate Longitudinal Data

Authors: Diego Tovar-Rios, KU Leuven, Stavanger University Hospital; Anita Sunde, Stavanger University Hospital, University of Bergen; Dries De Write, KU Leuven; Geert Verbeke, KU Leuven, Hasselt University; Geert Molenberghs, KU Leuven, Hasselt University; Ariel Alonso, KU Leuven, Hasselt University

Abstract: Joint linear mixed models are commonly used to quantify subject-level associations between longitudinal outcomes through correlated random effects. When group-specific covariance structures are introduced to account for heterogeneity, likelihood-based estimation becomes computationally demanding in large samples. We propose a partitioned pseudo-likelihood (PPL) approach within a joint modelling framework to improve scalability while preserving interpretability. The method is illustrated using longitudinal plasma pTau217, plasma NfL, and cognitive data from the population-based HUNT study in the context of Alzheimer’s disease. Increasing the number of partitions reduced runtime and memory requirements without materially affecting correlation estimates. The framework further enables the study of partial and multiple correlations in multivariate longitudinal settings. Overall, PPL provides a computationally efficient and scalable strategy for analysing complex dependence structures in large-scale longitudinal data.

Szurewsky Chloé: Non-parametric Bayesian Approach for the One-Sample Log-Rank Test

Authors: Szurewsky Chloé, Paris-Saclay University; Le Teuff Gwénaël, Paris-Saclay University

Abstract: In oncology, randomized trials may be difficult to conduct. Single-arm trials have emerged as one alternative to compare the survival of an experimental group to that of an external control group S0(t) using the frequentist one-sample log-rank test (OSLRT) or its Bayesian version. In both cases, a parametric estimate of S0(t) is required to calculate the expected number of events, assuming no variability of S0(t). To overcome these two restrictive assumptions, we develop a non-parametric Bayesian OSLRT using a mixture of Dirichlet Process. A Dirichlet process prior is put on the cumulative distribution function with an exponential base distribution and its parameter following a Gamma prior distribution. A simulation study is conducted to estimate the frequentist performance of this approach compared with conventional approaches. The Bayesian OSLRT is less conservative than the other tests, in particular for small sample size. Under the PH scenarios, the Bayesian OSLRT is more powerful than the other tests for small sample sizes and less powerful for higher sample sizes. Further work is to evaluate performance in non-PH settings and to incorporate information from historical data.

Federico Ferraccioli: Resampling the Null Space: Nonparametric Tests for Crossed Random Effects Models

Authors: Federico Ferraccioli (independent), Livio Finos (University of Padova); Angela Andreella (Ca' Foscari University of Venice)

Abstract: Mixed-effects models are powerful tools for analyzing experimental design data involving subjects and items as random effects. Researchers commonly use crossed-random effect models to analyze data from experiments or customer ratings in fields such as psychology, linguistics, neuroscience, and digital markets. In this context, we present a nonparametric approach based on the sign-flip score test. The proposed method does not rely on distributional assumptions for the error components and remains valid even in small samples and unbalanced data. The test’s performance is demonstrated through the experimental results.

Qingzi Huo: Extending the Generalised Covariance Measure for Conditional Independence Testing in Repeated-Measures Data

Authors: Qingzi Huo, Humboldt University of Berlin; Marco Simnacher, Humboldt University of Berlin; Xiangnan Xu, Humboldt University of Berlin; Sonja Greven, Humboldt University of Berlin

Abstract: The Generalised Covariance Measure (GCM) is a regression-based regression-based test of conditional independence that was originally developed for independent and identically distributed observations. However, directly applying the standard GCM to longitudinal data, where repeated measurements within a subject are correlated, can inflate Type I error by ignoring this within-subject dependence. We propose an extension of the GCM that uses mixed-effects models to capture within-subject correlation and thereby recover valid Type I error control in repeated-measures settings.

Arnau Garcia-Fernández: A Strictly Proper Metric for Evaluating Dynamic Predictions for Left-Truncated and Right-Censored Data with Asymmetric Misclassification

Authors: Arnau Garcia-Fernández, Erasmus University Medical Center; Dimitris Rizopoulos, Erasmus University Medical Center; Judith Kooiman, Erasmus University Medical Center; Pedro Miranda-Afonso, Erasmus University Medical Center

Abstract: Dynamic risk prediction models are widely used in precision medicine, but the metrics employed to assess their performance leave an important gap. On one hand, strictly proper metrics such as the dynamic Brier score do not reflect decision asymmetry. On the other hand, clinical utility measures such as net benefit depend on a single threshold and are not strictly proper. We address this by proposing a metric that integrates the cost-weighted misclassification error over the full range of possible thresholds with a user-defined weight function. This yields a strictly proper metric while allowing clinically appropriate asymmetry between false positives and false negatives, without committing to a single decision threshold. We provide a nonparametric inverse-probability-weighting estimator of this metric for left-truncated and right-censored data. The methodology is motivated by the clinical task of dynamically predicting delivery in pregnancies complicated by fetal growth restriction. All methods are implemented in the JMbayes2 R package.

Jonas Moss: Marginal versus typical task success in Bayesian IRT analysis of AI benchmarks

Authors: Jonas Moss, BI Norwegian Business School

Abstract: AI benchmarks that measure how quickly autonomous agents are improving at real-world tasks increasingly guide safety policy and industry forecasts. The most influential such benchmark, maintained by the non-profit METR, assumes that task difficulty is fully determined by how long a task takes a human to complete. We fit a Bayesian item response theory model that lets tasks of the same human completion time vary in difficulty and show that this variation has large consequences. When success is averaged over the distribution of task difficulties (``marginal'' success), it is roughly an order of magnitude lower than success evaluated at the mean difficulty (``typical'' success). Current forecasting methods that ignore this gap may substantially overstate what AI agents can do.

Lars Knieper: Selective Function-on-Surface Regression for Cortical Neuroimaging via Gradient Boosting

Authors: Lars Kneiper, University of Göttingen; Tobias Hepp, Friedrich-Alexander-University Erlangen-Nürnberg; Nadia Müller-Voggel, Friedrich-Alexander-University Erlangen-Nürnberg; Anna von Plessen, University of Göttingen; Elisabeth Bergherr, University of Göttingen;

Abstract: Neuroimaging data observed on the cortical surface are characterized by high spatial dimensionality and strong dependence across locations, complicating statistical inference and effect estimation. Cluster-based permutation tests are widely used to identify significant activation regions but inference is restricted to the cluster level and does not yield spatially resolved effect estimates or a regression framework for incorporating multiple covariates or frequency-specific effects. We propose a function-on-scalar regression model for cortical surface data in which source-space measurements are represented by Wendland radial basis functions defined on the cortical surface. Model-based component-wise gradient boosting is used for estimation and implicit selection of basis functions, yielding spatially localized and interpretable activation estimates within a unified regression framework.

Antoni Sieminski: Fast boosted GAM via Adaptive Smoothness Relaxation with scale-invariant tensor product interactions

Authors: Antoni Sieminski, University of Edinburgh, Maxwell Institute Graduate School; Simon Wood, University of Edinburgh, Maxwell Institute Graduate School

Abstract: Component-wise gradient boosting (CWGB) is attractive for generalised additive model (GAM) estimation when large numbers of potential covariates are to be screened, as it integrates variable selection with estimation. But it suffers from two problems: 1. CWGB is very slow to converge if some effects should have high effective degrees of freedom; 2. Even without this, smooth interactions are difficult to include without ad hoc assumptions about relative covariate scaling. By analysing the cause of slow convergence we are able to develop an effective acceleration method, which also allows us to properly incorporate smooth interaction terms without ad hoc scaling assumptions.

Annika Strömer: Optimal categorization of continuous variables in high-dimensional models via boosting and deselection

Authors: Annika Strömer, University of Marburg; Irantzu Barro, University of the Basque Country UPV/EHU; Susana García-Gutiérrez, Galdakao-Usansolo University Hospital; Andreas Mayr, University of Marburg

Abstract: In biomedical research, continuous predictors are often categorized to obtain clinically useful decision rules, but arbitrary cut-points may obscure risk relationships and result in information loss. We propose a two-stage, model-based procedure that derives data-adaptive candidate cut-points and then selects the most suitable partition for each predictor within a multivariable regression framework. The method can be applied in multiple type of models to categorize multiple predictors simultaneously and is suitable for high-dimensional settings.

Qiong Wu: Genetic prediction modelling in large cohort studies via boosting distributional regression

Authors: Qiong Wu, University of Marburg; Hannah Klinkhammer, University of Marburg; Kiran Kunwar, University of Marburg; Christian Staerk, IUF–Leibniz Research Institute for Environmental Medicine, TU Dortmund University; Carlo Maj, University of Marburg; Andreas Mayr, University of Marburg.

Abstract: Polygenic prediction modelling is used to develop scores representing the individual genetic liability for diseases or other health outcomes. Current methods primarily focus on modelling the mean but not other distribution parameters. We introduced snpboostlss, a batch-wise distributional regression boosting algorithm, to model multiple distribution parameters in large scale cohorts where classical methods become unfeasible. Our approach has so far been restricted to Gaussian location and scale models. We propose an adapted version extending it to further data situations such as count data. We illustrate our approach on childhood sunburns in the UK Biobank (n = 263,693, p = 549,278). Our results indicate that modelling the dispersion parameter could identify novel genetic variants that have not been associated with phenotypic mean, and the prediction by snpboostlss could correctly capture the trend of mean and variance in the UK Biobank.

Niccolò Cao: Latent variable models for longitudinal data under missing at random

Authors: Niccolò Cao, University of Bologna; Silvia Cagnone, University of Bologna; Livio Finos, University of Padova;

Abstract: Patient-reported outcome measures (PROMs) are increasingly used to collect self-reported data on patients' health-related latent constructs, e.g., functional mobility. In this context, a standard analytic framework is provided by the longitudinal confirmatory factor analysis (CFA) model, which enables modelling and evaluating a latent construct over time along with multiple methods to handle missing at random (MAR) data. Although full-information maximum likelihood (FIML) guarantees valid inference under MAR, it can be computationally demanding with high-dimensional data. Limited-information methods - as pairwise maximum likelihood (PML) - reduce computational burden, but require additional strategies (e.g., estimator corrections). In this study, we compare FIML with PML estimators through an application based on a dataset provided by the IRCSS Rizzoli Orthopaedics Institute (ROI).

Pavel Hernández Amaro: Variable Domain Multivariate Functional Principal Component Analysis

Authors: Pavel Hernández Amaro, Universidad Carlos III de Madrid; María Durbán, Universidad Carlos III de Madrid; M. Carmen Aguilera-Morillo, Universitat Politècnica de València; José María Quintana, Galdakao-Usansolo University Hospital; Irantzu Barrio, University of the Basque Country; Sonja Greven, Humboldt-Universität zu Berlin.

Abstract: Multivariate functional principal component analysis (MFPCA) assumes all observations share a common fixed domain, an assumption frequently violated when observation periods differ across subjects. We propose a Variable Domain Multivariate Functional Principal Components (VD-MFPC), which extends MFPCA to the variable domain setting by modeling the covariance structure of stacked univariate scores as a smooth function of domain length using penalized splines. Simulation studies confirm consistent improvements over binning strategies of 50--80\% in both reconstruction and eigenfunction recovery. Applied to Peripheral Capillary Oxygen Saturation (SPO$_2$) and body temperature trajectories in COVID-19 patients with varying hospitalization lengths, the extracted scores show a strong association with mortality outcomes.

Matteo D'Alessandro: A semiparametric nonlinear mixed effects model with penalized splines using automatic differentiation

Authors: Matteo D'Alessandro, University of Oslo; Magne Thoresen, University of Oslo; Øystein Sørensen, University of Oslo;

Abstract: We present an estimation procedure for the semiparametric nonlinear mixed-effects model (SNMM) in which the population trajectory is represented by penalized splines and adapted to individuals via subject-specific transformation parameters. By exploiting the mixed model representation of penalized splines, the level of smoothness can be estimated jointly with other variance components. The integration over random effects needed to obtain the marginal likelihood is carried out using Laplace approximation. Exact derivatives for evaluation and optimization of the resulting likelihood are obtained via automatic differentiation. In simulation studies, the method produces improved inferential performance when compared to the existing procedure. The approach is illustrated through simulation and a data example on infant height growth.

Nikolai Spuck: Sparse modeling of category-specific effects in partial proportional odds models using trees

Authors: Nikolai Spuck, Heidelberg University; Moritz Berger, Heidelberg University

Abstract: The proportional odds model is arguably the most widely applied regression approach for ordinal response variables. It assumes that the effects of the covariates are constant across the response categories, which allows a parsimonious model specification but may be rather restrictive. Alternatively, partial proportional odds models, where proportionality is only enforced for a smaller subset of covariates, can be applied. In this paper, we introduce a model fitting algorithm that forms clusters of categories sharing the same effect on the response, with a focus on model parsimony and ease of interpretation. The predictive ability and selection properties of the proposed approach were compared to alternative methods in a simulation study. For illustration, the model was applied to clinical data on the disease course of patients presenting in the emergency department.

Haoxiang Gao: Optimally Weighted Random Forest for Predicting Binary Outcomes

Authors: Haoxiang Gao, UCL; Rumana Omar, UCL; Menelaos Pavlou, UCL

Abstract: We propose an accelerated optimally weighted random forest (OWRF) for probability prediction in binary outcomes. Tree weights are estimated using a generalised Mallows-type criterion and the optimisation is computationally accelerated using a 2-step quadratic programming (QP). The proposed approach improves probability estimation relative to standard random forests, particularly in scenarios with moderate prevalence and large number of predictors.

Dayasri Ravi: Penalized Finite Mixtures of Proportional Hazards Models

Authors: Dayasri Ravi, University of Augsburg; Samyajoy Pal, Technical University of Kaiserslautern-Landau; Christian Heumann, LMU Munich

Abstract: We propose an ℓ1-penalized finite mixture of proportional hazards models for heterogeneous, high-dimensional time-to-event data. Each component uses a flexible piecewise-constant baseline hazard, and parameters are estimated via an EM algorithm with Newton-type updates and soft-thresholding to induce sparsity. A real-world application demonstrates improved predictive performance over Cox and ℓ1-penalized Cox models.

Terry W.S. Chan: Cancer Recurrence Prediction Over Time: Moving beyond proportional hazards with discrete-time neural networks

Authors: Terry W. S. Chan, Netherlands Cancer Institute; Mark Opdam, Netherlands Cancer Institute; Sabine C. Linn, Netherlands Cancer Institute; Renee X. Menezes, Netherlands Cancer Institute

Abstract: Cox proportional hazards models commonly used in clinical risk prediction are often used even when hazards are non-proportional, leading to poor estimation of absolute risks, especially when follow-up period is long or competing risks exist. We present a discrete-time neural network (NN) model that flexibly models cause-specific hazards without the proportionality assumption. Using a Dutch breast cancer cohort (PARADIGM), we show that this approach delivers similar discrimination but markedly improved calibration versus Cox models, while providing patient-specific time-dependent risk profiles. Temporal smoothing and granularity of time discretisation are also explored. Discrete-time NNs are a practical alternative for absolute risk prediction in clinical prognosis under non-proportional hazards.

Jessica Doohan: A statistical mixture neural-additive model

Authors: Jessica Doohan, University of Limerick; Aliaksandr Hubin, NMBU; Lucas Kook, Vienna University of Economics and Business; Kevin Burke, University of Limerick.

Abstract: Neural networks (NNs) are flexible prediction models due to their universal approximation properties, yet their interpretability remains limited, often leading to them being regarded as “black-box” models. In contrast, statistical models prioritise interpretability, which is crucial in high-risk domains such as health applications. We explore hybrid approaches combining the strengths of both model classes. We propose statistical mixture models that combine linear model, generalised additive model (GAM) and NN components, allowing data to determine their relative contributions. This framework enables a test of whether a linear component, a GAM, an NN, or their combination is most appropriate, preserving interpretability when linear/GAM effects dominate while also enabling adaptation to more complex patterns. Flexibility constraints are also considered to balance interpretability and predictive performance.

Cristiano Varin: Ridge estimation in network meta-analysis

Authors: Cristiano Varin, Ca' Foscari University of Venice

Abstract: This paper explores ridge regression for ranking treatments within a network meta-analysis framework. We show that the intrinsic dependency structure of the experimental graph complicates standard shrinkage estimation, rendering classical tuning methods—such as cross-validation—unreliable. To address this, we propose an alternative approach based on minimizing the total mean square error evaluated at pilot estimates. Various extensions and complications of the basic framework will be discussed during the conference presentation.

Marco Grzegorczyk: Tied-Covariance Mixtures of Gaussian Bayesian Networks

Authors: Marco Grzegorczyk, University of Groningen

Abstract: Mixtures of Gaussian Bayesian networks (GBNs) have been developed and applied under full-covariance assumptions, where each mixture component has its own mean vector and covariance matrix. Here we propose a new mixture GBN model with tied-covariance, in which all components share a common covariance matrix and only the mean vectors vary. We evaluate the performance of tied- and full-covariance mixtures of GBNs on benchmark data.

Ilia Negri: Hierarchical Bayesian Estimation of Underreporting with Multiple Data Sources

Authors: Ilia Negri, University of Calabria; Maura Mezzetti, University of Rome ``Tor Vergata"

Abstract: A hierarchical Bayesian framework to estimate under-reported event counts by jointly modeling occurrence and reporting processes and integrating individual-level survey data with aggregated regional counts observed at different spatial scales is proposed. The model separates victimization and reporting mechanisms through distinct covariate structures and links individual and regional components by integrating configuration-specific risks with respect to the regional covariate distribution. Applied to assaults and sexual violence in Italy, preliminary results highlight considerable posterior uncertainty, particularly for sexual violence, reflecting limited observed cases and reinforcing the need for uncertainty-aware inference when combining imperfect data sources.

Yilin Chen: A functional approach for variance decomposition in function-valued traits

Authors: Yilin Chen, King's College London; Davide Pigoli, King's College London; John A.D. Aston, University of Cambridge; Patrick A. Carter, Washington State University

Abstract: Function-valued (FV) traits, such as growth trajectories, are phenotypes of living organisms whose value can be described by a function of a continuous index. These traits can be assessed for continuous genetic variation using a quantitative approach that decomposes phenotypic variation into additive genetic and environmental components, if family relationships are observed. However, FV traits from different individuals often have variation in both amplitude and phase. In this work, we show how removing phase variability can impact the analysis and propose a functional mixed-effects model with functional principal components as the data-driven basis for genetically correlated data. The proposed method is applied to a real-case study of flour beetle growth trajectories.

Hedvig Borgen Reiersrud: Confidence Curve Combination as an Alternative to Wilks in High-Energy Physics

Authors: Hedvig Borgen Reiersrud, University of Oslo; Clara Bertinelli Salucci, University of Oslo; Anders Kvellestad, University of Oslo; Riccardo De Bin, University of Oslo

Abstract: In high-energy physics, inference for parameters of interest is commonly based on Wilks' theorem, using the asymptotic chi-square pivot of the profile likelihood ratio to perform hypothesis testing or construct confidence intervals. However, the validity of this approximation relies on asymptotics and regularity conditions that may fail in practice, leading to miscoverage of the derived confidence intervals in nonregular or small-sample settings. We propose an alternative approach based on confidence curves, which provides coverage-exact inference without relying on asymptotic approximations. Results from a simulation study show that Wilks-based intervals can severely overcover in small-signal scenarios, while our method maintains nominal coverage for all confidence levels.

Klazien de Vries: Bayesian D-optimal designs for GAMLSS in psychological test norming

Authors: Klazien de Vries, University of Groningen; Marieke E. Timmerman, University of Groningen; Casper J. Albers, University of Groningen

Abstract: We develop Bayesian D-optimal designs for GAMLSS models commonly used in continuous norming of psychological tests, namely heteroscedastic normal, beta-binomial, and Box-Cox power exponential distributions. In a proof-of-concept example, we derive exact Bayesian D-optimal designs for the beta-binomial model. For moderately informative priors, this yielded designs similar to local D-optimal designs, while weak priors produced more uniform designs.

Despo Theocharous: Variational Logistic Regression via the Cressie Read Family of Divergences

Authors: Despo Theocharous, RWTH Aachen University; Maria Kateri, RWTH Aachen University

Abstract: Bayesian logistic regression has an intractable marginal likelihood, making variational inference (VI) a tractable tool for posterior approximation. Due to non-conjugacy, standard coordinate ascent variational inference (CAVI) updates are not available; Pólya–Gamma augmentation restores conjugacy and yields closed-form updates based on the reverse Kullback–Leibler (KL) divergence. We instead consider the Cressie–Read (CR) family of divergences, controlled by a parameter $\lambda$. Replacing the reverse KL by a CR divergence breaks conjugacy of the augmented posterior, preventing direct closed-form updates. To overcome this, we employ a tangent transform approximation that recovers tractable variational distributions analogous to the KL-based case. This yields closed-form CR-based CAVI updates (CAVI-CR), which are more robust than their KL counterparts (CAVI-KL). Moreover, CAVI-CR coincides with CAVI-KL in the limit $\lambda \to -1$, and converges asymptotically to the KL-based solution for all $\lambda$. Simulation studies demonstrate improved robustness to outliers for small to moderate sample sizes.

Monika Zimmermann: Scalable Estimation of Large Generalized Additive Models: Extended Fellner-Schall via Hutch++ and Conjugate Gradients

Authors: Monika Zimmermann, University Duisburg-Essen; Simon Wood, University of Edinburgh; Florian Ziel, University of Duisburg-Essen

Abstract: We address the computational bottleneck arising from formation and factorization of the penalized Hessian in large generalized additive models (GAMs) with high-dimensional parameter vectors. We combine the generalized Fellner--Schall smoothing parameter update with stochastic trace estimation (Hutch++) and preconditioned conjugate gradients (PCG), thereby avoiding formation or Cholesky factorization of the GAM penalized Hessian. The resulting procedure relies only on matrix--vector products, enables low memory-bandwidth parallelization and allows exploitation of model term sparsity with minimal infill. The performance of the approach is demonstrated on the NMMAPS respiratory mortality data with over one million observations and more than 20,000 coefficients, fitted in just over half an hour.

Jan-Ole Fischer: Fast and scalable inference in hidden Markov models with Gaussian fields

Authors: Jan-Ole Fischer, Bielefeld University

Abstract: Hidden Markov models (HMMs) and Gaussian processes (GPs) offer powerful frameworks for time series analysis, but their naive combination is computationally prohibitive due to the cubic complexity of GP inference. We resolve this by employing the stochastic partial differential equation (SPDE) approach which approximates Gaussian fields as sparse Gaussian Markov random fields (GMRFs), enabling linear-time computations. The HMM log-likelihood's dense Hessian, however, prevents efficient Laplace approximation in frameworks like Template Model Builder (TMB) when integrating out high-dimensional random effects. To overcome this, we propose a banded forward algorithm that partitions the time series into local blocks, approximating the forward recursion using only recent observations. This induces the required sparsity, enabling efficient high-dimensional inference. We demonstrate our approach through stellar flare detection in TESS photometric data, modelling the light curve as a quasi-periodic GP trend combined with a state-switching flaring process. Our method reduces runtime from hours to minutes, showcasing its potential for real-time analysis in astrophysics and related fields.

Johannes Piller: Capturing Non-Markov Dynamics with Piecewise Exponential Additive Models

Authors: Johannes Piller, LMU Munich; Julian Ostertag, Technical University of Munich School of Medicine; Felipe Gentilini, LMU Munich; Lena Zelinka, LMU Munich; Andreas Bender, LMU Munich

Abstract: Multi-state and recurrent event models are central tools for analyzing complex event histories in biomedical research. However, when future transitions depend on an individual’s event history, standard Markov assumptions are violated and Aalen–Johansen–based transition estimates become inconsistent. We propose Piecewise Exponential Additive Models (PAMs) and their multi-state extensions as a flexible framework for directly modeling non-Markov dynamics through covariates. Using simulations covering Poisson, Markov, and non-Markov recurrent events, we compare PAMs to specialized non-parametric estimators. PAMs show higher robustness to misspecification and lower bias in non-Markov settings. An application to eICU delirium data illustrates their practical utility.

Fiete Sieg: Independent truncation of the exponential distribution

Authors: Fiete Sieg, University of Rostock; Rafael Weißbach, University of Rostock

Abstract: Left-truncated (LT) lifespans are considered as a truncated point process, approximated by a Poisson process. After profiling out the non-parametric distribution of the truncation time, we estimate the parameter of exponentially distributed lifetimes, including right-censoring (RC). Similarities and differences to the usual counting process approach are explored. Assuming consistency, we derive the limiting distribution of the estimator through the theory of least favourable submodels. Simulations and results for an example are reported.

Touqeer Ahmad: A new extension of the generalized Pareto distribution for modeling low threshold exceedances

Authors: Touqeer Ahmad, University of Oslo; Thordis Linda Thorarinsdottir, University of Oslo

Abstract: We propose a new extended generalized Pareto distribution (eGPD) that employs a truncated logistic generator to model extremes with flexible bulk behavior. The model retains the GPD upper-tail properties and introduces a new shape parameter to make the bulk more flexible. This makes sure that the density is positive at the origin and that there is a smooth transition between moderate and extreme values. A simulation study shows that the proposed model outperforms existing eGPD variants in high-quantile estimation, especially at low thresholds. An application to summer temperature records from Montréal–Trudeau International Airport demonstrates that the eGPD remains superior for low-threshold exceedances, a setting where the standard GPD yields poor results.

Samyajoy Pal: Generalized Linear Models with Mixtures of Zero-Inflated Count Distributions

Authors: Samyajoy Pal, TU Kaiserslautern-Landau; Jochen Einbeck, Durham University

Abstract: Zero-inflated count data often exhibit excess zeros and heterogeneous dispersion that classical Poisson or Negative Binomial regression cannot capture. While Zero-inflated Poisson (ZIP) and Zero-inflated Negative Binomial (ZINB) models address these features, their assumptions can be restrictive when multiple zero-generating mechanisms coexist. We propose a generalized linear model whose response distribution is a convex mixture of ZIP and ZINB components, forming a flexible family that nests Poisson, Negative Binomial, ZIP, ZINB, and hybrid submodels. We establish key properties including alternative representations, the induced zero-inflation structure, moment expressions, and identifiability up to label switching. Parameters are estimated by direct likelihood maximization, with score and observed information derived to support likelihood-based inference, including likelihood ratio tests under boundary constraints. Simulations evaluate finite-sample performance and test behaviour under varying zero-inflation and overdispersion, and applications to cytogenetic count datasets demonstrate practical gains and interpretable separation of zero-inflation and dispersion mechanisms.

Ferdinand Buchner: Identifying risky flight landings using multivariate vine copula regression models

Authors: Ferdinand Buchner, Technical University of Munich; Claudia Czado, Technical University of Munich

Abstract: Runway overruns remain a major operational risk in aviation, often caused by long touchdown distances, slow deceleration, or both. Understanding the joint tail behavior of these components is essential for identifying landings with high overrun potential. We propose and compare three multivariate vine copula regression frameworks for modeling the conditional joint distribution of touchdown distance, the distance to reach controllable speed, and average deceleration. The first framework introduces the new class of fork vine copulas for multivariate regression; the second combines separate D-vine regressions with an R-vine copula; and the third uses generalized additive models combined with an R-vine copula. Applied to a data set of 711 cargo landings under identical runway and aircraft conditions, our methods reveal pronounced non-Gaussian dependence between the responses. Our best-performing model identifies flights with substantially increased risk, particularly when manual braking is initiated late or brake duration is short. Our findings demonstrate that explicitly modeling the conditional dependence between the responses improves risk identification.

Claudia Czado: Noise Contrastive Estimation for Calibrating Simplified Vine Copulas

Authors: Michael D. Kraus, Technical University of Munich; David Huk, University of Warwick; Claudia Czado, Technical University of Munich

Abstract: Vine copulas flexibly model multivariate dependence but rely on the simplifying assumption, which prevents conditional pair copulas from varying with conditioning values. We propose an observation-specific calibration for simplified vines using noise contrastive estimation (NCE). Using the simplified vine as the noise model, NCE yields posterior probabilities that adjust each observation’s log-likelihood, providing data-driven flexibility.

Rouven E. Haschka: Beyond Gaussian dependence: Nonparametric Vine copulas to handle endogenous regressors

Authors: Rouven E. Haschka, Technical University of Kaiserslautern-Landau; Julien Hambuckers, University of Liège; Thomas Kneib, Georg-August University of Göttingen

Abstract: Empirical researchers increasingly rely on the instrumental-variable-free Gaussian copula (GP) approach to identify and correct endogeneity in regression models estimated with non-experimental data. This approach models the joint distribution of endogenous regressors and the regression error using a Gaussian copula, allowing endogeneity bias to be mitigated by basing inference on a likelihood based on the joint distribution rather than on the marginal error distribution. We propose a nonparametric generalization based on density-estimated vine copulas. Our framework employs a pseudo-penalized maximum likelihood estimator in which (i) the joint distribution of endogenous regressors and errors is modeled using a flexible nonparametric copula, and (ii) the copula structure and smoothing parameters are selected via cross-validation. Monte Carlo simulations benchmark our approach against the Gaussian copula method and demonstrate that the latter fails to correct endogeneity under a wide range of non-Gaussian dependence structures, whereas the proposed method yields unbiased estimates. The method is later on applied to retail scanner data to estimate demand elasticities.

Maya Natascha Vienken: Jump-augmented stochastic volatility: a discrete-time state-space approach

Authors: Maya Natascha Vienken, Bielefeld University; Jan-Ole Fischer, Bielefeld University; Roland Langrock, Bielefeld University; Sina Mews, Bielefeld University

Abstract: Major events such as the COVID-19 pandemic, the Russia-Ukraine war, and recurring trade conflicts repeatedly cause sudden and severe spikes in financial market uncertainty that basic stochastic volatility models fail to capture. We propose a discrete-time state-space model in which the latent log-volatility follows an autoregressive process augmented by a jump component. The jump term is specified using an exponential distribution. This reflects the empirical asymmetry: volatility increases abruptly but reverts to its mean more gradually. The parameters are estimated by direct numerical approximate maximum likelihood. We consider four nested specifications, differing in whether the observation equation assumes Gaussian or Student-t errors and whether the state equation includes a jump component. Applied to crude oil returns, jump-augmented specifications clearly outperform non-jump counterparts across standard information criteria, with identified jumps aligning closely with known market-disrupting events.

Iain Bell: Multivariate state space models for cross-correlated censored heavy metal concentrations

Authors: Iain Bell, University of Glasgow; Ruth O'Donnell, University of Glasgow; Jafet Belmont Osuna, University of Glasgow; Claire Miller, University of Glasgow; Michael Dunbar, The Environment Agency

Abstract: River water catchments can consist of a complex mixture of chemicals with determinands often exhibiting complex seasonal patterns, autocorrelation, co-occurrence and censoring below a detection limit. To accurately account for these complexities, a novel state space modelling approach is proposed to investigate multivariate relationships, temporal trends and the possibility of prediction of water quality data. This is applied to heavy metal concentrations sourced from the south-west of England.

Eduardo Janotti Cavalcante: Beyond Kaplan–Meier: Modelling Survival Tails

Authors: Eduardo Janotti Cavalcante, Universidade de São Paulo; Antonio Pedroso de Lima, Universidade de São Paulo; Lígia Henriques-Rodrigues, Universidade de Lisboa

Abstract: We propose a semiparametric survival model that embeds extreme-value structure into a latent intensity framework, enabling coherent tail extrapolation under right censoring. Event times are modeled via a Log-Gaussian Cox Process whose mean corresponds to the log-hazard of a Generalized Pareto distribution, allowing the extreme value index to vary over $\mathbb{R}$ while preserving flexibility at moderate times. Inference relies on a tractable discretization combined with Hamiltonian Monte Carlo. Simulations under bounded, exponential, and heavy-tailed regimes show stable estimation of the extreme value index and scale parameter, even under substantial censoring. An application illustrates that the model captures the main empirical survival features with uncertainty comparable to the Kaplan--Meier estimator, while providing smooth extrapolation beyond the observed follow-up. The framework offers a unified alternative to classical extreme-value and nonparametric survival approaches when long-term tail behavior is of primary interest.

Luisa Ferrari: Asymmetric Dirichlet prior for informed clustering of multivariate binary data

Authors: Luisa Ferrari, University of Modena and Reggio Emilia; Maria Franco Villoria, University of Modena and Reggio Emilia; Garritt Page, Brigham Young University

Abstract: We introduce the recent asymmetric Dirichlet prior approach (aFMM) to the context of multivariate Bernoulli mixture models for multivariate binary data. The chosen prior allows users to encode prior knowledge about the expected number of clusters and associated uncertainty. Estimation employs a single-chain MCMC with smart initialization and simulated annealing to mitigate multimodality. Findings show that aFMM accurately recovers latent cluster structure, achieving full posterior inference while remaining computationally efficient, and outperforms competing approaches in both clustering accuracy and runtime.

Tobias Hepp: Model-based clustering for time series and functional data with Mixture Density Networks

Authors: Tobias Hepp, Friedrich-Alexander-Universität Erlangen-Nürnberg; Elisabeth Bergherr, Georg-August-Universität Göttingen

Abstract: In this work, we propose a novel approach utilizing mixture models to identify clusters of similar functions. We present the results of a simulation study in which unsupervised classification performance of the model was investigated under controlled conditions before demonstrating its application to a dataset of daily traffic patterns of bicycles on a major thoroughfare in the city of Göttingen.

Elisa Carolina Gonzalez Santacruz: Interval Wavelets based Multiscale Normalizing Flow for Forcasting Irregularly Sampled Time Series (CDV-FIMTS)

Authors: Elisa Carolina Gonzalez Santacruz, Universidade de São Paulo; Chang Chiann, Universidade de São Paulo, Brazil

Abstract: This work addresses the problem of probabilistic forecasting of irregular multivariate time series (IMTS), a common scenario in domains such as healthcare, climate, and economics, characterized by heterogeneous sampling and missing values. Traditional statistical methods and deep learning techniques, although effective in certain contexts, face limitations when dealing with nonlinearities and dynamic dependencies among variables. Recent studies indicate that graph-based architectures, normalizing flows, and wavelet transforms can help overcome these limitations, yet they have not been unified into a single model. To fill this gap, we propose a probabilistic framework that combines graph-based representations, invertible normalizing flows, and multiresolution decompositions via wavelets, aiming to estimate the joint distribution of IMTS. The method is evaluated on three real-world datasets from the medical and climate domains and compared against benchmark models from the literature. The results demonstrate improvements in predictive accuracy. To the best of our knowledge, this is the first work to integrate graph structures and wavelets within a unified normalizing flow.

Adriano Zanin Zambom: Additive ARCH-m(X) Model via Backfitting Algorithm

Authors: Adriano Zanin Zambom, California State University Northridge; Qing Wang, Wellesley College

Abstract: We propose an additive ARCH-m(X) model that incorporates nonlinear covariate effects in conditional variance while enforcing positivity. Three backfitting strategies are developed: log-scale smoothing, shifted fits, and variance smoothing. Simulations show improved predictive performance over traditional linear models. The application to S&P 500 returns demonstrates substantial out-of-sample gains.

Kristína Sakmárová: Novel Bayesian hierarchical group lasso and its application in pharmaceutical modelling

Authors: Kristína Sakmárová, Charles University; Arnošt Komárek, Charles University; Martin Otava, Janssen-Cilag s.r.o.

Abstract: Throughout the development of the pharmaceutical manufacturing process, settings need to be addressed for series of factors that can have an impact on the quality of the product. Regression models are used to this end. There can be numerous interactions among the factors. At screening stage, the aim is to choose those factors that have an important effect, taking into consideration possible hierarchy of the model terms and grouping of the factors. To be able to choose as few factors which have the effect on the product as possible, penalization of the regression coefficients needs to be applied. Several methods of variable selection are compared from both Bayesian and frequentist perspective and Bayesian hierarchical group lasso is proposed as a novel method surpassing the capabilities of already existing methods.

Razvan-Andrei Morariu: Hierarchical Lasso for multi-factor designs

Authors: Razvan-Andrei Morariu, University of Vienna; Tatyana Krivobokova, University of Vienna; Gianluca Finocchio, University of Vienna

Abstract: High-throughput experimentation (HTE) generates large, structured datasets that require specialized statistical methods for analysis. In particular, in chemical HTE, reaction properties are measured for all combinations of multiple components, forming a factorial design with a single observation per factor-level combination. We propose a hierarchical Lasso estimator specifically designed for such multi-factor datasets, enforcing sparsity and hierarchical constraints on main effects and interactions. The novelty of our approach lies in extending hierarchical estimators to accommodate higher-order (beyond two-way) interactions, which is essential for this type of data. In the case of a linear model with a normal response, our method achieves consistent variable selection and estimation under mild conditions. The extension to generalized linear models is straightforward, although formal proofs of consistency are still in progress.

Aaron Coats: Bayesian Variable Selection for Large Datasets through Bi-Directional Splitting

Authors: Aaron Coats, University of Glasgow; Vinny Davies, University of Glasgow; Mayetri Gupta, University of Glasgow

Abstract: Variable selection in a Bayesian setting often becomes computationally infeasible in large modern datasets with increasing numbers of observations and covariates. We propose a framework that partitions the original dataset by both rows and columns into smaller, more manageable batches processed in parallel, and then uses a two-phase procedure to combine the batches and efficiently perform variable selection in the large n and large p setting. Our experiments show that the approach makes substantial computational gains compared to processing the full dataset, with only a slight decrease in selection accuracy.

Yaoyao Li: Multi-level Variable Selection Using a BART-Enhanced Mixed-Effects Framework

Authors: Yaoyao Li, Rutgers University; Keming Zhang, Brown University; Jungang Zou, Columbia University; Sijian Wang, Rutgers University; Bernadette Fausto, Rutgers University; Liangyuan Hu, Rutgers University

Abstract: We propose a flexible, fully Bayesian unified framework for simultaneous variable selection of both fixed and random effects in hierarchical data. Our method integrates Bayesian Additive Regression Trees (BART) for nonlinear fixed effects with a hierarchical component using covariance decomposition and spike-and-slab priors for random effects. While effective when cluster-level covariates vary, the Unified Model faces stability issues due to near-collinearity when these covariates are constant within clusters. To address this, we introduce a Two-Step Approach that decouples random intercept estimation from cluster-level selection. Simulation studies demonstrate the Unified Model's superiority for varying covariates and the Two-Step Approach's robustness for constant covariates. We apply these methods to identify Alzheimer's disease risk factors in the New Jersey Social Determinants of Health dataset.

Inge Helland: The foundation of quantum mechanics as seen by a statistician

Authors: Inge Helland, University of Oslo

Abstract: It is shown how basic elements of quantum mechanics may be derived from simple assumptions using the notion of theoretical variables, variables that may be accessible or inaccessible. This notion can also be shown to have consequences for statistical modeling, and for quantum decision theory.

Sheng-Tsaing Tseng: A Physics-Guided Approach to Multi-Run Step-Stress Tests for Rechargeable Batteries

Authors: Sheng-Tsaing Tseng, National Tsing-Hua University; Nan-Jung Hsu, National Tsing-Hua University

Abstract: Rechargeable lithium-ion batteries are essential to modern electronics, making accurate lifespan estimation a key research focus. Tseng et al. (2025) used degradation data from multi-run k-level step-stress tests, in which batteries experience varying stress levels across repeated cycles, to predict end-of-performance (EOP) under normal-use conditions. While this advanced method, which combines the cumulative exposure property with a trend-renewal process

Aliaksandr Hubin: Bayesian Fractional Polynomials for Optimal Dosage Estimation

Authors: Aliaksandr Hubin, NMBU and University of Oslo; Åshild Krogdahl, NMBU; Guro Løkka, NMBU; Trond M. Kortner, NMBU

Abstract: The problem of optimal dosage estimation arises in diverse scientific domains, from pharmacology and toxicology to aquaculture and environmental studies. Statistical modeling of nonlinear dose–response relationships is essential to quantify biological effects and determine response-maximizing levels. This paper introduces a flexible Bayesian fractional polynomial (BFP) framework for modeling such relationships, allowing for model uncertainty quantification and robust prediction through Bayesian model averaging. Extensive simulation results demonstrate that the proposed BFP approach yields accurate estimation of optimal dose levels, outperforming standard benchmarks. The approach is further demonstrated on real data from fish nutrient requirement experiments.

Mikis Stasinopoulos: Statistical Modelling Process: Basic Ideas and Common Pitfalls

Authors: Mikis Stasinopoulos, Greenwich University; Robert Rigby, University of Greenwich; Fernanda De Bastiani, Universidade Federal de Pernambuco; Julian Merder, Aarhus University

Abstract: In this article we define basic ideas involved statistical modelling (SM) process. The idea is that if we clarify the basic ideas behind the process we would be able to avoid the pitfalls associated with statistical modelling. We will introduce the idea of focussed statistical modelling (FSM) through a simple example of distributional regression, centile estimation

Sonja Greven: Principal component analysis for sparsely sampled density functions

Authors: Lisa Steyer, Humboldt-Universität zu Berlin; Sonja Greven, Humboldt-Universität zu Berlin

Abstract: We present a novel functional principal component analysis (FPCA) approach for density functions, when these are available only via potentially small samples. We directly model the observed data to account for all sources of uncertainty, instead of relying on a two-step approach with presmoothing of densities, which can be problematic for small or heterogeneous numbers of samples per density. We model densities in Bayes spaces to account for their constrained nature, or equivalently in the subspace L2_0 of L2-functions integrating to zero after centered log-ratio (clr) transformation. We treat underlying densities as latent and employ a Monte Carlo Expectation Maximization (MCEM) algorithm for maximum likelihood estimation. The new method is applied to analyze the dis- tributions of increasing summer month temperatures over the last 70 years in Berlin.

Timo Adam: Flexible modelling of higher-order moments in stochastic volatility models

Authors: Timo Adam, Bielefeld University; Jan-Ole Fischer, Bielefeld University

Abstract: Stochastic volatility models are a key tool for capturing market uncertainty in financial time series. While they describe the evolution of conditional variance well, they typically assume a fixed return distribution and thus ignore variation in higher-order moments such as skewness and kurtosis. Yet these features often vary systematically with volatility, especially during financial stress. We extend the stochastic volatility framework by allowing flexible parametric return distributions whose parameters are smooth functions of a latent state process. These functions are estimated using penalised B-splines, enabling a data-driven yet regularised adaptation to complex distributional dynamics. By capturing the full shape of the conditional return distribution, the model improves estimation of risk measures such as value-at-risk and expected shortfall, leading to more accurate risk assessment in financial applications.

Poster Presentations

Stepan Mazur: Estimation of minimum VaR and minimum CVaR optimal portfolios for small sample and singular covariance matrix

Authors: Stepan Mazur, Örebro University; Muhammad Asif, Örebro University; Svitlana Drin, Örebro University; Nataliya Shchestyuk, National University of Kyiv-Mohyla Academy

Abstract: We study sample estimators of minimum VaR and minimum CVaR optimal portfolios in the setting where both the population and sample covariance matrices of asset returns are singular. We derive stochastic representations that completely characterise the distributions of the portfolios’ expected returns and variances, and we obtain expressions for the corresponding means and variances.

Su-Fen Yang: A Phase II Sign-Based Ratio-of-dispersions Control Chart

Authors: Su-Fen Yang, National Chengchi University; Yu-Cheng Chou, National Chengchi University

Abstract: For a distribution-free bivariate quality variable, we propose a sign-based ratio-of-dispersions control chart to monitor changes in the ratio of two variances. Under in-control conditions, and assuming that the process covariance matrix is known, this ratio can be equivalently expressed as the difference between the two variances. Based on this idea, we derive an unbiased estimator for the difference of the sample variances, along with its mean and variance. Using these results, we construct a sign-based ratio-of-variances control chart. Numerical studies demonstrate the properties and detection performance of the proposed chart under three different bivariate distributions. Keywords: Sign control chart; ratio of variances; bivariate distribution.

Michela Battauz: A joint longitudinal model for students’ performance on exams and academic outcomes

Authors: Michela Battauz, University of Udine; Giuseppe Alfonzetti, University of Udine

Abstract: Students’ performance on exams is determined by grades and completion times. We model these responses as functions of two latent variables, ability and speed, which also influence the probability of observing the academic career outcomes, namely graduation, dropout, or transfer to another program. A joint longitudinal model is proposed to account for informative censoring in exam records, arising when the students do not complete their studies.

Leila Barmoudeh: DeepKriging for Soil pH Prediction: Capturing Spatial Dependencies for Improved Spatial Mapping

Authors: Leila Barmoudeh, University of Bergen; Panagiotis Tziachris, Hellenic Agricultural Organization “DIMITRA”; Bård Støve, University of Bergen

Abstract: Accurate spatial prediction of soil pH is essential for digital soil mapping and precision agriculture. Classical geostatistical methods such as Kriging rely on assumptions of stationarity and Gaussianity and require explicit covariance model specification, which may be restrictive for complex environmental processes. Although deep neural networks (DNNs) offer strong predictive flexibility, they do not inherently account for spatial dependence. This study evaluates the DeepKriging framework, which integrates spatial basis representations within a neural network architecture, for predicting soil pH in northern Greece. We compare DeepKriging with conventional DNN and ordinary Kriging. Model performance is assessed using cross-validated mean squared error (MSE) and mean absolute error (MAE). DeepKriging consistently achieves superior predictive accuracy, demonstrating its ability to capture nonlinear relationships and spatial dependence simultaneously. Variable importance analysis identifies Fe, Mn, and Mg as dominant predictors. These results highlight DeepKriging as a principled and scalable alternative to classical geostatistical models for spatial prediction tasks.

Brunel Darwin Ngoune Domo: Regularised Multi-Omics Regression modeling for Transcriptomic-Proteomic Integration in Liver‑Damaged Mice

Authors: Brunel Darwin Ngoune Domo, TU Dortmund University; Jan G. Hengstler, IfADo-Leibniz Research Centre for Working Environment; Andreas Groll, TU Dortmund University

Abstract: Numerous previous studies have addressed the question of whether protein levels can be predicted by their corresponding mRNAs. However, these studies have shown that mRNA levels explain only a small fraction of the variability of the respective proteins. To address this limitation, we analysed how accurately protein levels can be modelled based on the corresponding mRNA levels, if a set of mRNA-protein data pairs from an independent training data set is available. Building on the pipeline proposed by Heiner et al.~(2025), we introduce an additional random forest preselection step before Lasso regression. This hybrid approach, further developed here, allows us to filter variables more effectively and to focus the modeling on the most informative predictors. We test the extended framework on a new data set on mice with liver damage. The method reduces noise, highlights biologically meaningful gene-protein relationships, and can be used to impute missing proteomic values or detect outliers.

James Booth: Predicting disease prevalence in a clustered population

Authors: James Booth, Cornell University; Brenda Hanley, Cornell University; Krysten Schuler, Cornell University

Abstract: The impact of sampling schemes on the prediction of disease prevalence in animal populations is studied in the presence of intra-cluster correlation

Simon Pauli: Estimating Mating Behavior Through Pairwise Mutation Counts

Authors: Simon Pauli, Johannes Kepler University; Andreas Futschik, Johannes Kepler University;

Abstract: This short paper sums up the proposal of a method to estimate parameters that describe the mating behavior of a species based on the mutations that separate pairs of alleles. Furthermore, the inbreeding models by Campbell (2015) and Severson et al. (2019) are extended to also model non-monogamous behavior. Simulations suggest that the estimates are unbiased and able to reliable capture inbreeding behavior, even some that has not been modelled explicitly.

Tsai-Hung Fan: An EM Algorithm for Log-Normal Generalized Limited Failure Population Models

Authors: Tsai-Hung Fan, National Central University; Chih-Ying Tai, National Central University

Abstract: The generalized limited failure population model (GLFP) provides a meaningful framework for capturing both infant mortality and wearout failures while explicitly accounting for unit defectiveness. This paper considers a log-normal GLFP model and develops an efficient estimation procedure based on the expectation-maximization (EM) algorithm for the failure time data under a Type-I censoring scheme. In addition to using latent variables for the failure mode of a unit and for whether the unit is defective, we apply data augmentation for the unobserved failure times to construct a complete likelihood function. The primary contribution is the derivation of analytical solutions for parameter estimation via the EM algorithm, and the observed Fisher information matrix is accommodated with the missing information principle, which significantly reduces computational cost. The proposed method is applied to a subset of hard-drive failure data from the Backblaze company to further illustrate its practicality and feasibility.

Sinead Moylett: Evaluating imputation strategies for longitudinal cohort studies

Authors: Sinead Moylet, University of Limerick; George Smith-Kolff, University of Canterbury; Blair Robertson, University of Canterbury

Abstract: Missing data is a structural feature of longitudinal cohorts and can bias inference when attrition is systematic. We evaluate imputation strategies for ordinal outcomes in the Irish Longitudinal Study on Ageing (TILDA) across five waves. The cohort includes 8,504 respondents; for evaluation we construct wave-specific ground-truth datasets using stratified samples of 1,000 respondents per wave (5,000 records). Realistic Missing at Random (MAR) patterns are introduced via weighted amputation. We compare a naive mode baseline, three Random Forest (RF) configurations, and Multiple Imputation by Chained Equations (MICE) with proportional odds models. Mode imputation achieves the lowest RMSE (0.994) but introduces bias (0.079) and distributional distortion (KL 0.0293). MICE shows minimal bias (−0.036) and best distributional preservation (KL 0.0009), with RMSE 1.121. Bagged RF is the most stable RF approach (RMSE 1.009), while optimised variants can perform worse and fail under some waves. Results highlight trade-offs between point accuracy and population-level validity

Annette Möller: The zero-degree of freedom chi-squared distribution in a gradient-boosting framework for probabilistic precipitation forecasting

Authors: Annette Möller, Bielefeld University; David Jobst, University of Hildesheim; Jürgen Groß, University of Hildesheim

Abstract: Forecasts from numerical weather prediction (NWP) models typically suffer from systematic bias and dispersion errors. Statistical postprocessing methods such as ensemble model output statistics (EMOS) have been shown to successfully correct the NWP forecasts. Recently, the zero degree of freedom noncentral chi-squared (Chi0) distribution was introduced for postprocessing precipitation amount. This work extends the Chi0 based EMOS model in a gradient-boosting framework. A case study indicates improved performance of Chi0-EMOS-GB over the basic Chi0-EMOS model.

Agnes Mallinger: Approximate Bayesian computation for stochastic hybrid systems with ergodic behaviour

Authors: Agnes Mallinger, Johannes Kepler University; Sascha Desmettre, Johannes Kepler University; Amira Meddah, Johannes Kepler University; Irene Tubikanec, Johannes Kepler University

Abstract: Piecewise diffusion Markov processes (PDifMPs) form a versatile class of stochastic hybrid systems that combine continuous diffusion processes with discrete event-driven dynamics, enabling flexible modelling of complex real-world hybrid phenomena. In this work, we present a novel framework for parameter inference in PDifMPs based on approximate Bayesian computation (ABC). We extend existing ABC summary statistics for diffusion processes to account for the hybrid nature of PDifMPs, showing particular effectiveness for ergodic systems. Our results show that the proposed ABC method reliably recovers model parameters for example PDifMPs that empirically exhibit ergodic behaviour. These findings highlight the potential of ABC as a practical tool for inference in various complex stochastic hybrid systems.

Rouven Michels: The Best of Both Worlds: Predicting Coverage Schemes in American Football with Supervised and Unsupervised Learning

Authors: Rouven Michels, TU Dortmund University; Robert Bajons, WU Vienna; Jan-Ole Fischer, Bielefeld University

Abstract: This study introduces a statistical framework that integrates supervised and unsupervised learning to predict defensive coverage schemes in American football. Using high-resolution player tracking data, we aim to classify plays into man- or zone coverage. For this, two supervised models are employed: an elastic-net logistic regression, providing an interpretable framework; and an XGBoost model, allowing for flexible nonlinear patterns. Our feature set consists of i) pre-motion information, ii) naive post-motion characteristics and iii) summary statistics and a play-specific random effect, derived from results of a hidden Markov model that captures latent guarding assignments during motion phases. We show that the latter, integrated approach improves predictive performance compared to a feature space solely consisting of pre-motion and naive post-motion information.

Izzati Izyani Japar: Sample Size Determination for Developing Risk Prediction Models with Clustered Binary Data

Authors: Izzati Izyani Japar, University College London; Gareth Ambler, University College London; Menelaos Pavlou, University College London; M. Shafiqur Rahman, University of Dhaka; Rumana Z Omar, University College London

Abstract: Risk prediction models for health outcomes are often developed using multi-centre data, where within-cluster correlation must be accounted for. Adequate sample size for model development is essential to avoid overfitting and ensure good predictive performance. Existing sample size approaches for clustered data are not explicitly designed to target the predictive performance and recently proposed methods for independent data may be unsuitable for clustered data. We used simulations to evaluate whether models developed using existing sample size approaches achieved pre-specified targets for the mean absolute prediction error (MAPE). We found that existing methods often failed to achieve the target MAPE. We developed a new meta-model-based sample size formula which consistently achieved the target MAPE across a wide range of scenarios.

Zachary James: A Framework for Nonstationary Gaussian Processes with Neural Network Parameters

Authors: Zachary James, Cornell University; Joseph Guinness, Washington University in St. Louis

Abstract: We present a nonstationary Gaussian process model where nonstationary parameters are expressed with a neural network. We fit the model to a geospatial dataset and show that it outperforms existing models in log-score and computation time and is competitive in mean square error.

Nirian Martin: Selecting the most efficient robust log-linear regression for Rayleigh distribution

Author: Nirian Martin, Complutense University of Madrid

Abstract: Sintetic Aperture Radar (SAR) Imagery techniques are considered to be an efficient way to monitor the Earth's surface. The Rayleigh regression model was presented in Palm et al. (2019) for ground type detection in SAR Imagery. Inspired in the log tranformed lifetime modeling with regression, considered in Kalbfleisch and Prentice (2002), this paper explores three models, theoretically equivalent to the one considered by Palm et al. (2019) using log-linear regression models, when maximum likelihood estimation (MLE) is being used. It is very important to highligh that such an equivalence is based on the invariance property of the MLEs for estimating parameters when lifetime random variables are tranformed. The presence of outliers in SAR data is very common and they can give inaccurate inferences. This papers aims to present for the first time density power divergence based estimators (Basu et al., 1998) and test statistics based on them, applicable in SAR Imagery. The asymptotic distribution of the estimators of the three models behind the initial unique model are separately studied and the most convenient one is selected taking into account efficiency and robustness criteria.

Ralitza Gueorguieva: Longitudinal Multivariate Mixed Models for Partially Ordered Data

Authors: Ralitza Gueorguieva, Yale School of Public Health; Eugenia Buta, Yale School of Public Health; Maria Iannario, University of Naples Federico II

Abstract: Longitudinal multivariate mixed models (LMMMs) are a powerful tool for analyzing repeated measurements of multiple correlated outcomes over time. This paper presents an extended LMMM framework for ordinal outcomes that accounts for the presence of ‘don’t know’ (DK) response options, incorporating them into the model formulation within a partially ordered response structure. Maximum-likelihood estimation using numerical approximations and iterative maximization and model comparison is performed. The proposed approach aims to provide unbiased and efficient statistical inference in cases where missing or uncertain responses are prevalent. An application from a longitudinal survey evaluating perceptions of the addictiveness of various tobacco products is presented. Simulations are used to evaluate the model-fitting algorithm and to study the effects of ignoring DK responses.

Chiu-Hsieh Hsu: A multiple imputation-based sensitivity analysis approach for a MNAR binary variable

Authors: Chiu-Hsieh Hsu, University of Arizona; Nathan Cunicelli, University of Arizona

Abstract: Under the selection modeling framework, we propose a multiple imputation sensitivity analysis approach for a MNAR binary variable, where the sensitivity analysis parameter is the correlation coefficient between event probabilities and selection probabilities under a bivariate probit model. The sensitivity analysis involves multiple imputations of missing values, yet the sensitivity parameter is only used to determine the event probabilities for imputing missing observations. Hence, the proposed approach is expected to be more robust against mis-specifications of the sensitivity parameter. For illustration, the proposed approach is applied to incomplete measurements of HPV status for patients who had head and neck cancers. A simulation study is conducted to evaluate the performance of the proposed approach.

Ming-Che Lu: A Bayesian EWMA Control Chart for Monitoring Multinomial Overdispersion Processes

Authors: Ming-Che Lu, Chaoyang University of Technology; Su-Fen Yang, National Chengchi University

Abstract: Multivariate count data are often overdispersed due to additional sampling variability, such as clumped sampling (Efron, 1986). This study explores the statistical process control of multinomial distribution for overdispersed count data. Bayesian methods are used to construct multinomial control charts due to computational simplicity, ease of interpretation, and evidence in favor of in-control. Using the Dirichlet distribution as a prior, Bayesian exponentially weighted moving average (EWMA) control charts are proposed for monitoring the multinomial overdispersion process. The average run lengths (ARL) of the proposed control charts are calculated by Monte Carlo simulations to evaluate their detection performances. Moreover, the comparison results show that the proposed control charts present fewer false alarms than existing multinomial control charts which do not consider overdispersion.

Simon Hirsch: Towards Online Distributional Regression for Time Series

Authors: Simon Hirsch, University of Duisburg-Essen, Statkraft; Sven Pappert, TU Dortmund University; Jonathan Berrisch, University of Duisburg-Essen; Florian Ziel, University of Duisburg-Essen

Abstract: In a wide array of domains, distributional regression models are often used for time series forecasting. However, the focus is often on (non-linear) effects of the lagged target variable, while lagged effects of latent variables, such as the estimated distribution parameters and residuals, are often ignored. In this work, we discuss the stochastic properties of this general class of time series processes, the relationship with well-known processes such as the GARCH and INGARCH, and the regularized and incremental estimation of distributional regression mod- els for time series. We provide a case study on electricity price data.

M Cristina Miranda: Estimating the extremal index in a robust way

Authors: M Cristina Miranda, University of Aveiro; Manuela Souto de Miranda, University of Aveiro; M. Ivette Gomes, University of Lisbon

Abstract: Modelling extreme events is crucial for assessing and preventing large-scale impacts in many real-world applications. Extreme Value Theory (EVT) provides a statistical framework for modelling distributional tails and quantifying rare events. However, in many applications, extremes occur in clusters due to temporal dependence. In such cases, the Extremal Index (EI) is a parameter that characterises the limiting mean cluster size. This paper discusses the robustness properties of the EI P-estimator and illustrates its performance using real data sets.

Francesco Bartolucci: On a class of Bayesian hidden Markov models for spatio-temporal data

Authors: Francesco Bartolucci, University of Perugia; Daniele Tancini, University of Perugia; Riccardo Rastelli, University College Dublin

Abstract: A class of spatio-temporal hidden Markov models is proposed, where the latent process follows an extension of the autologistic model. These models are extremely difficult to estimate because the latent joint distribution is explicitly available only in trivial cases. We show how Bayesian inference can be carried out using block pseudo-distributions and an approximate exchange algorithm, as an alternative to the more common use of a standard pseudo-distribution.

Annamaria Guolo: Inference in meta-analysis of accuracy of diagnostic tests under different link functions

Author: Annamaria Guolo, University of Padova

Abstract: Random-effects meta-analysis is a widely accepted approach for the evaluation of the accuracy of a new diagnostic test, compared to a gold standard. This work investigates the performance of a likelihood approach based on a multinomial processing tree model and that of a simulation-based alternative (SIMEX), when different link functions are used to model (transformations of) sensitivity and specificity of the diagnostic test at the between-study level.

Davi Oliveira Chaves: Determining the number of factors in locally stationary dynamic factor models

Authors: Davi Oliveira Chaves, Universidade de São Paulo; Chang Chiann, Universidade de São Paulo; Esther Ruiz, Universidad Carlos III de Madrid; Pedro Alberto Morettin, Universidade de São Paulo.

Abstract: This paper introduces a novel method for determining the number of factors in locally stationary dynamic factor models based on ratios of eigenvalues computed from local kernel-based estimates of the time-varying covariance matrix of the process. Simulation studies highlight its effectiveness.

Matteo Ventura: A Pairwise Likelihood Approach for High-dimensional Models for Ordinal Data

Authors: Matteo Ventura, University of Calabria; Sabrina Giordano, University of Calabria; Dimitris Karlis, Athens University of Economics and Business; Paola Zuccolotto, University of Brescia

Abstract: This paper introduces a pairwise likelihood approach for modeling multivariate ordinal data with CUB margins using copula functions. This method is particularly useful in high-dimensional settings where standard likelihood methods become intractable. Application to survey data confirms the model’s ability to effectively capture complex dependence structures across multiple items.

Pere Puig: Statistical Modelling of the 1820 Plague in Mallorca

Authors: Pere Puig, Universitat Autònoma de Barcelona, Centre de Recerca Matemàtica; Joana M. Pujadas-Mora, Universitat Oberta de Catalunya, Centre d'Estudis Demogràfics

Abstract: This paper models the 1820 plague outbreak in Mallorca, which caused over 2,400 deaths, using a system of differential equations to describe the dynamics of susceptible, infected, living, and recovered populations. Key epidemiological parameters are estimated in a Bayesian framework from historical daily records. The results suggest an important role for pneumonic or septicemic transmission beyond a purely bubonic pattern.

Amaia Iparragirre: Design-based inference for AUC-based model evaluation

Authors: Amaia Iparragirre, University of the Basque Country UPV/EHU; Thomas Lumley, University of Auckland; Irantzu Barrio, University of the Basque Country UPV/EHU

Abstract: Inference on the AUC remains challenging for data arising from complex sampling designs. In this work, we propose a design-based framework for AUC inference using replicate weights methods, including Jackknife Repeated Replication and the Rescaling Bootstrap. Simulation results suggest that replicate weights methods outperform the traditional bootstrap for constructing confidence intervals and hypothesis tests.

Irantzu Barrio Beraza: Modeling Bout Distributions in COPD with GAMLSS: From Activity Patterns to Survival

Authors: Irantzu Barrio Beraza, University of the Basque Country UPV/EHU; Cristóbal Esteban, University Hospital of Galdakao, Spain; Thomas Kneib, University of Göttingen, Germany;

Abstract: Using GAMLSS, we characterize the full distribution of bout duration in COPD patients, modeling location and scale as functions of clinical covariates. The derived distributional scores have shown to be strong predictors of survival.

Ina-Marie Berendes: Modelling of break point angles in the Nordic hamstring exercise using exercise variation and muscle activity data

Authors: Ina-Marie Berendes, TU Dortmund University; Jana Bergmann, TU Dortmund University; Andreas Groll, TU Dortmund University; Thomas Jaitner, TU Dortmund University

Abstract: The Nordic hamstring exercise serves as training for the eccentric strength of the hamstring muscles with the goal of injury prevention. We examine how the exercise variation as well as EMG, force, and kinematic covariates influence the break point angle achieved in the exercise execution due to the angle being an indicator for eccentric hamstring strength. Data from 44 participants executing three variants of the exercise are analyzed. A Lasso-penalized linear mixed model (LMM) and a conditional random forest (cRF) are employed to identify crucial covariates and investigate their influence on the break point angle. A nested cross-validation procedure is used to assess the prediction performance of the two modelling approaches via RMSE and MAE. The random forest shows total prediction errors 36-38 % smaller than the mixed model and uses the exercise variation and integrated EMG covariates most heavily, while other covariates are still important for prediction of the break point angle.

Anabela Rocha: Robust strategies in panel data: the case of hidden outliers

Authors: Anabela Rocha, University of Aveiro; M. Cristina Miranda, University of Aveiro; Manuela Souto de Miranda, University of Aveiro

Abstract: The analysis of panel data or longitudinal data is fundamental to modern research, since its multi-dimensional structure allows for the capture of individual-specific effects over time. However, the inherent complexity of these datasets often makes them susceptible to contamination, where outliers can bias traditional estimation techniques severely and lead to misleading conclusions. Consequently, developing and implementing robust estimators is a critical necessity to ensure the reliability of statistical inferences. This research presents a robust estimation technique for random effects models in panel data framing, by incorporating robust methodologies into the Feasible Generalized Least Squares (FGLS) framework. Also provides a method for identifying both cellwise and casewise outliers. The performance of the robust proposed estimator is evaluated using a simulation study and a real data set.

Eirik Høyheim: "What-if" contributions in latent binary Bayesian neural networks

Author: Eirik Høyheim, Norwegian Defence Research Establishment

Abstract: This work exploits a local linearity property that neural networks exhibit around a given input, making it feasible to view them as generalized linear models and hence extract covariate contributions. Latent binary Bayesian neural networks are used, which allows us to learn sparser models and quantify the uncertainty of the provided contributions. Furthermore, we compare explanations between different models trained on the same problem as the input varies, showing that simpler architectures can achieve comparable predictive performance while offering more interpretable structures.

M. Mar Fenoy: Flexible gamma frailty model with long-term survivors

Authors: M. Mar Fenoy, Complutense University of Madrid; N. Balakrishnan, McMacter University; M. Carmen Pardo, Complutense University of Madrid

Abstract: Abstract: A frailty model is a random effects model for time variables, where the random effect (the frailty) has a multiplicative effect on the hazard with the aim of modeling the unobserved heterogeneity in the population. In addition, it is common a proportion of non-susceptible subjects, even if, accompanied by a large time, which is so-called cured. In this work, we extend the generalized time- dependent logistic frailty model proposed by Milani et al. (2015) to a flexible family of frailty models which allows for non-proportional hazards as well as survival data with long-term survivors. Inferential methods are then developed for the models. Finally, the applicability of the proposed models was illustrated with a real bladder cancer data set.

Luca Cibinel: Enhancing Quantum Graph Representations via Heuristics and Model-Based Clustering

Authors: Luca Cibinel, University of Oslo; Claudio Meggio, University of Oslo; Lorenzo Zuccato, University of Oslo

Abstract: We propose a novel class of representations for graph-structured data, with the objective of improving feature prediction performance. We focus on graphs with clearly identifiable and interpretable subgraphs, such as transition metal complexes. Our approach is based on the identification of subgraph populations whose citizens exhibit a similar statistical behaviour. Knowledge of these populations is injected in the original graphs in the form of virtual nodes to make global information available at the local level.

Kunal Rehani: Mixed-Effects Models for High-Frequency Longitudinal Health Data

Authors: Kunal Rehani, Western Sydney University; Kenan M. Matawie, Western Sydney University

Abstract: Digital tinnitus interventions generate intensive longitudinal data through ecological momentary assessment (EMA), characterised by repeated observations and strong within-person dependence. We analysed daily EMA data from a smartphone-based intervention using linear mixed-effects models (LMMs) and a generalised linear mixed-effects model (GLMM). Distress and loudness were recorded on 0–100 visual analogue scales across baseline and intervention phases and clinical change was assessed using the Tinnitus Handicap Inventory (THI). An interaction model tested whether distress trajectories differed by phase, followed by models evaluating level effects and covariate adjustment. Distress was strongly associated with contemporaneous loudness, while intervention effects attenuated after adjustment and did not modify time slopes. A GLMM for THI improvement ı.e. (≥ 7-point reduction in THI score) indicated a strong phase association. These results highlight the importance of mixed-effects modelling for analysing symptom trajectories and clinically meaningful change in intensive longitudinal health data.

Sophie Potts: Comparing Information Criteria for Prediction-based Variable Selection in Model-based Gradient Boosting

Authors: Sophie Potts, University of Goettingen; Martin Refisch, Johann Heinrich von Thünen Institute; Elisabeth Bergherr, University of Goettingen

Abstract: Model-based gradient boosting is a statistical learning routine, allowing for data-driven variable selection in possibly high-dimensional data settings. A substantial corpus of research has been dedicated to the classical algorithm, with the objective of enhancing the quality of variable selection in terms of a low False Positive Rate (FPR) and a high True Positive Rate (TPR) by keeping the good predictive performance. The approach of prediction-based variable selection (Potts et al., 2024) has shown favorable properties in cases of Gaussian distributed outcomes but struggled with other types of Generalized Linear Models. We expand the idea towards other model selection criteria, namely the corrected AIC as well as the BIC and test their performance for Poisson and Binomial distributed outcome variables. The simulation study suggests that BIC-boost is preferable in non-normally distributed cases. The application on data of right-wing populist election wins showcases this superiority in practice.

Andrew McInerney: Selection and fusion of categorical predictors using a smooth information criterion

Authors: Andrew McInerney, University of Limerick; Lea Kaufmann, RWTH Aachen University; Maria Kateri, RWTH Aachen University; Kevin Burke, University of Limerick

Abstract: Categorical predictors often contain levels with indistinguishable effects, motivating the fusion of category coefficients to obtain parsimonious and interpretable regression models. We propose a smooth information-criterion-based method that enables simultaneous estimation, factor selection, and fusion of categorical effects within generalized linear models. The approach replaces the discrete model complexity term underlying the Bayesian information criterion (BIC) with a continuously differentiable surrogate, allowing direct optimisation without requiring a tuning parameter.

Colin Griesbach: Markov-Switching Regression via Gradient Boosting Techniques

Authors: Colin Griesbach, University of Göttingen; Sina Mews, University of Bielefeld

Abstract: In this work we adapt recent findings from statistical boosting in order to construct an estimation approach for Markov-switching regression models based on gradient boosting. The algorithm relies on established concepts for boosting with latent variables and is evaluated via synthetic data.

Jacob Grytzka: A GAMLSS approach for interpretable emulation of climate-vegetation relationships

Authors: Jacob Grytzka, TU Dortmund University; Andreas Groll, TU Dortmund University; Anne Dallmeyer, Max-Planck-Institute for Meteorology, University of Cologne; Nils Weitzel, TU Dortmund University, University Alliance Ruhr

Abstract: Future projections of the Earth system rely on complex numerical models, which are validated by assessing their ability to reproduce past climate changes. Here, we develop a statistical model to emulate the climate – carbon dioxide – vegetation relationship in an Earth system model (ESM), with the goal of identifying drivers of environmental change and quantifying uncertainties. To facilitate interpretability of the emulated relationships, we employ semi-parametric distributional regression, specifically Generalized Additive Models for Location, Scale, and Shape (GAMLSS), trained with spatio-temporal fields from an ESM. We focus on the ability to select predictors from a set of highly-correlated variables, for example through random forest pre-selection and penalization. The ESM implementation facilitates conditional independence between training data such that residual spatio-temporal structure can be used as a criterion for missing covariates. Preliminary results show a high accuracy for spatio-temporal vegetation patterns and an ecologically meaningful hierarchy of predictor importance.

Edoardo Redivo: A Generative Model for Microclustering in DNA Data Storage

Authors: Edoardo Redivo, University of Bologna; Silvia Dallari, University of Bologna; Cinzia Viroli, University of Bologna

Abstract: Synthetic DNA enables digital data storage by encoding information into short nucleotide sequences called reference strands. During data retrieval, sequencing produces multiple noisy copies, called reads, of each reference strand, containing insertions, deletions, and substitutions. This gives rise to the inferential problem of clustering reads according to their unobserved reference strands. This setting falls into the microclustering regime, where the number of clusters grows proportionally with the sample size while cluster sizes remain small. We introduce a generative latent editing mixture model in which each read is obtained from a latent template through independent per-site editing operations. The framework naturally generates variable-length reads and supports likelihood-based inference.

Silvia Dallari: Clustering Microbiome Data via Bray-Curtis Dissimilarity-Based Mixture Models

Authors: Silvia Dallari, University of Bologna; Laura Anderlucci, University of Bologna; Angela Montanari, University of Bologna;

Abstract: Research on the gut microbiome is crucial for understanding human health and the role of microbial communities in biological systems and disorders. A central objective is to identify groups of individuals sharing similar microbiome configurations, as this may support a range of clinical and health applications, including the prevention of dysbiosis and the development of group-specific therapeutic strategies. Characteristic of biological communities is the biodiversity, which can be described through different diversity indices. Among them, there is the class of β-diversity measures (e.g., Bray–Curtis dissimilarity), which capture differences between samples. In this work, the Bray-Curtis dissimilarity is used to build a distance-based density that is then incorporated into a mixture model. In this way, our proposal combines the flexibility of a model-based framework with the use of biological dissimilarities, thereby bridging the gap between model-free clustering methods based on β-diversities and fully probabilistic mixture models. The potential of the proposal is illustrated through a real microbiome application.

Vegard Kvernelv: Depth-to-Bedrock Mapping Using Kriging with Heterogeneous Data Sources

Authors: Vegard Kvernelv, Norwegian Computing Center; Ingrid Aarnes, Norwegian Computing Center; Petter Abrahamsen, Norwegian Computing Center

Abstract: We evaluate a kriging-based approach for mapping bedrock depth, applied to the Stubberud area in Oslo. Accurate bedrock mapping is crucial for urban planning but often relies on heterogeneous data sources with varying reliability. We utilize a data set of approximately 15,000 observations, ranging from high-precision core penetration tests to less precise torque tests. In our approach, we model bedrock depth as a spatial random field with deterministic trend, spatially correlated residuals, and explicit measurement error to handle data quality differences. For the local Stubberud model, we employ an empirical trend and an anisotropic nested variogram. Results indicate a bedrock depth up to 90 meters, and on average 20 meters deeper, than suggested by an uninformed baseline model. This study demonstrates how geostatistical methods can transparently integrate diverse data to produce more geologically plausible models for decision-making.

Maxi Schulz: Supervised Machine Learning Models for Longitudinal and Clustered Data - A Scoping Review

Authors: Maxi Schulz, University Medical Center Göttingen; Olesia Evgenievna Korchevaia, University Medical Center Göttingen; Tim Friede, University Medical Center Göttingen; Thomas Kneib, University of Göttingen; Maike Hohberg, University Medical Center Göttingen

Abstract: Longitudinal or clustered data often arise in clinical research, potentially violating the independent and identically distributed (i.i.d) assumption. Machine learning (ML) based algorithms offer great flexibility, however, usually assume i.i.d observations, at least implicitly. To address this limitation, extensions to ML algorithms for hierarchical data have been proposed. However, a comprehensive review that systematically identifies and compares supervised ML algorithms that can handle longitudinal and clustered data is lacking. This works aims to fill this gap by conducting a scoping review, which involves a systematic examination of the methodological literature. This ongoing work presents the methodology of the scoping review, as well as preliminary findings from the literature search. We anticipate sharing additional findings of the scoping review in the future.

Rui Martins: Sequential Bayesian Updating in Dynamic Hierarchical Models Without Recompilation Using NIMBLE

Authors: Rui Martins, University of Coimbra; Daniel Turek, Lafayette College

Abstract: Sequential Bayesian inference requires posterior updating based only on information available up to each time point. In compiled probabilistic programming frameworks, this is challenging because the model graph is fixed after compilation and future observations remain structurally embedded. We propose a strategy for sequential updating without recompilation in NIMBLE, combining likelihood gating, restriction of MCMC sampling to upstream stochastic nodes, adaptation resetting, and a decreasing iteration schedule. The full study structure is encoded at compilation, while information flow is dynamically controlled to prevent downstream likelihood leakage. The approach is illustrated with a dynamic hierarchical model for football match outcomes, enabling genuine round-by-round prequential evaluation.

Emily Davison: Scalable Bayesian Phylogenetic Models For Capturing Within-Species Trait Variation

Authors: Emily Davison, University of Glasgow; Christina Cobbold, University of Glasgow; Richard Reeve, University of Glasgow; Claire Harris, University of Glasgow; Vinny Davies, University of Glasgow;

Abstract: We present a Bayesian hierarchical phylogenetic model for continuous traits that incorporates within-species variation and explicitly accounts for heterogeneous sampling effort. By exploiting the three-point structure of Brownian motion covariance matrices on trees, our approach achieves substantially improved scalability compared with existing approaches, enabling analyses of phylogenies with hundreds of thousands of species. The key innovation lies in combining this computational efficiency with a hierarchical treatment of uneven sampling across species within a coherent Bayesian model. Simulation studies demonstrate near-linear computational scaling and robust parameter recovery, particularly when within-species variances are assigned hierarchical priors. The method provides a practical route to global-scale comparative analyses and improved prediction of trait values for poorly sampled species.

Ana Elizabeth García Sipols: Nonlinear Cointegration Testing Based on Order Statistics: A Robust Extension of the Engle and Granger Approach

Authors: Ana E. Sipols, Rey Juan Carlos University; M. Teresa Santos-Martín, University of Salamanca; Clara Simón de Blas, Rey Juan Carlos University; Lynda Atill, Mouloud Mammeri University of Tizi Ouzou; Hocine Fellag, Mouloud Mammeri University of Tizi Ouzou

Abstract: Traditional cointegration tests often fail to detect nonlinearities in cointegrated time series, leading to size distortions and power losses. This work proposes a new cointegration testing procedure that extends the classical Engle and Granger (1987) two-step approach by incorporating the nonparametric Range Unit-Root (RUR) and Forward–Backward RUR (FB-RUR) tests of Aparicio et al. (2006). The method applies order-statistic-based unit root tests to residuals from a short-run regression under the null hypothesis of no cointegration, avoiding prior estimation of the cointegration parameter. This design provides robustness to nonlinear adjustment, parameter instability, and structural change. Monte Carlo simulations under nonlinear error-correction models show that the proposed RURC and FB-RURC tests deliver accurate size and superior power to standard procedures, including Engle-Granger, Johansen, and error correction-based tests. An application to real exchange rates illustrates their ability to detect nonlinear cointegration missed by conventional linear methods.

Ivo Sousa-Ferreira: The defective Beta-Gompertz rate model with a time-dependent cure proportion

Authors: Ivo Sousa-Ferreira, Universidade da Madeira, Universidade de Lisboa; Ana Maria Abreu, Universidade da Madeira, Centro de Investigação em Matemática e Aplicações; Cristina Rocha, Universidade de Lisboa

Abstract: A new parametric rate model for gap times between recurrent events is introduced based on the defective Beta-Gompertz (BG) distribution. The model is formulated in terms of the conditional distribution of a gap time given the previous recurrence time, which naturally yields a time-dependent cure proportion through the use of an improper distribution. Several sub-models are available, some defective and others non-defective, enabling the presence of cured individuals in the population to be assessed using the likelihood ratio (LR) test. Inference is performed via maximum likelihood (ML) for right-censored observations. The practical relevance of the proposed model is illustrated by an application to hospital readmission data.

Aluísio Pinheiro: Temporal Patterns of Vehicle Theft in São Paulo: Wavelet–Copula Variational Bayes Framework for Bayesian Conditional Transformation Models

Authors: Giovanni Piccirilli, University of Campinas; Aluísio Pinheiro, University of Campinas

Abstract: This work proposes a Variational Bayes framework for Bayesian Conditional Transformation Models, with application to temporal patterns of vehicle theft. The proposed approach provides a computationally efficient alternative to MCMC-based inference while retaining flexibility to approximate the full posterior distribution of the model parameters. The methodology is applied to vehicle theft data from São Paulo city in Brazil, modeling the conditional distribution of theft counts as a function of time of day, day of the week, month, and year. The results reveal pronounced temporal patterns, with higher expected theft rates during nighttime and systematic differences between weekdays and weekends. Overall, the proposed framework offers a scalable and effective tool for distributional regression in large datasets with complex temporal dynamics.

Baraah Abu Abbadi: Investigating Generalised Additive and Neural-Network Models: An Empirical Study

Authors: Baraah Abu Abbadi, University of Limerick; Kevin Burke, University of Limerick; Andrew McInerney, University of Limerick

Abstract: Generalised Additive Models (GAMs) and Neural Networks (NNs) are widely used for modelling tabular data, but are often treated as competing approaches because they offer different strengths in interpretability and flexibility. A recent systematic review by Doohan et al. (2026) highlighted substantial inconsistency in methodological reporting across published GAM–NN comparisons. Published results are difficult to interpret because modelling procedures are not consistently reported. This paper presents an independent re-analysis of datasets identified in that literature under a common modelling framework. Preliminary results suggest that reported conclusions depend on modelling choice.

Christian Pfeifer: Power calculations/simulations in case of linar mixed models if we use an accelarated longitudinal design

Authors: Christian Pfeifer, Medical University of Innsbruck; Christopher Schaar, Medical University of Innsbruck; Beatrice Frajo-Apor, Medical University of Innsbruck; Alex Hofer, Medical University of Innsbruck

Abstract: ---

Fernanda De Bastiani: Detection Of Influential Curves in Functional Geostatistics

Authors: Fernanda De Bastiani, Universidade Federal de Pernambuco, Georg-August-Universität Göttingen; Jerfson B. N. Honório, Universidade Federal de Pernambuco; Manuel Galea, Pontificia Universidad Catolica de Chile; Miguel A. Uribe-Opazo, Universidade Estadual do Oeste do Paraná

Abstract: This work presents a method for detecting and visualizing influential curves in functional geostatistical data. The curves are represented using a cubic spline basis, and local influence is applied to the trace-variogram estimator, resulting in a closed-form diagnostic function defined over the temporal domain. The proposed approach is illustrated using daily temperature curves from Canadian meteorological stations. A functional boxplot initially identified three curves as potential outliers; further analysis confirmed that these curves are genuinely influential. These findings demonstrate that the proposed method goes beyond outlier detection by quantifying the impact of individual curves on the spatial dependence structure.

Ruggero Bellio: Integrated likelihood inference for risk difference in meta-analysis with sparse data

Authors: Ruggero Bellio, University of Udine; Annamaria Guolo, University of Padova; Nicola Sartori, University of Padova

Abstract: We propose the use of the integrated likelihood based on the zero-score expectation nuisance parameter for inference on a common risk difference for stratified binomial settings, typically arising in meta-analyses of comparative studies with binary outcomes. The proposal is built upon the Mantel-Haenszel estimator. It produces accurate inferences and, at the same time, resolves the numerical issues arising in sparse-data settings, without requiring any continuity correction or other changes to the original data. The same methodology extends to other risk measures, such as relative risk or log-odds ratio, and can be viewed as a general, unified solution for inference on stratified 2 × 2 tables.

Juan M. Rodriguez-Diaz: Pharmacokinetic Modeling and Personalized Design for Chemotherapy Clearance

Authors: Juan M. Rodriguez-Diaz, University of Salamanca; M. Teresa Santos-Martin, University of Salamanca; M. Isabel Asensio, University of Salamanca; Irene Marinas-Collado, University of Oviedo

Abstract: Therapeutic Drug Monitoring (TDM) is critical in high-dose chemotherapy, where toxicity is governed by the temporal evolution of plasma concentrations. Current protocols rely on fixed sampling schedules that fail to account for substantial inter-patient pharmacokinetic variability, particularly in High- Dose Methotrexate (HD-MTX) therapy. This rigidity leads either to unnecessary blood draws or delayed detection of toxic events. To address these limitations, this study proposes a hybrid Gamma–bi-exponential model to accurately characterize both the initial absorption phase and the multi-phase elimination kinetics. A Bayesian sequential optimal design can be then implemented to adaptively personalize sampling times based on patient-specific data. Simulation studies demonstrate that this methodology provides highly precise estimations of the safety threshold crossing time. Consequently, the proposed approach significantly reduces the clinical burden of frequent blood sampling while optimizing patient safety and care.

Eva Ryan: A Formal Causal Framework for Estimating the Effect of Chronic Pain on Mortality in Older Adults

Authors: Eva Ryan, University of Limerick; Hanna Grol-Prokoczyk, University at Buffalo; Christopher R. Dennison, University at Buffalo; Anna Zajacova, University of Western Ontario; Zachary Zimmer, Mount Saint Vincent University

Abstract: Chronic pain is a serious condition known to negatively impact many aspects of life, such as functional ability and mental wellbeing. However, uncertainty remains regarding the strength and nature of the pain-mortality relationship. To date, almost all research on this topic has assessed association, with few studies taking an explicitly causal approach. In this paper we propose a formal causal framework to estimate the effect of chronic pain on mortality in a large cohort of American older adults (n = 19,971) under the assumption of conditional exchangeability. Our assumed causal structure is illustrated using a directed acyclic graph. We compare multiple propensity score matching, inverse probability weighting, and doubly robust approaches to adjust for measured confounding. We estimate that chronic pain causes a modest increase in mortality hazard. Some results are also compatible with no causal effect.

Kushagri Tandon: Distributional properties of non-Gaussian spatial fields

Authors: Kushagri Tandon, University of Oslo; Thordis Linda Thorarinsdottir, University of Oslo; Geir Olve Storvik, University of Oslo;

Abstract: Environmental spatial fields such as precipitation or wind speed are often strongly non-Gaussian, yet dependence is commonly modelled via Gaussian random fields and Gaussian anamorphosis. The Gaussian-anamorphosis correlation is usually imposed through a Gaussian copula and tuned, rather than derived from an explicit non-Gaussian field. We investigate the distributional properties of two non-Gaussian spatial models---Gamma convolution (random particle) fields and non-Gaussian Matérn fields defined via SPDEs (stochastic partial differential equations) with type-G Lévy noise---and compare these against the Gaussian-anamorphosis correlations implied by their joint laws. Using bivariate Laplace transforms and low-order mixed moments, we construct two approximation routes: a transform-based route using Fejér-smoothed Fourier inversion, and an operator-theoretic route using truncated Lancaster expansions, and compare these to a Hermite--Mehler Gaussian-copula construction in terms of kriging performance.

Carlo Giovanni Camarda: Life Table Functions and Their Uncertainty: A Unified Matrix Algebra Framework

Author: Carlo Giovanni Camarda, Institut national d'études démographiques

Abstract: This paper develops a unified statistical framework for constructing confidence intervals for life expectancy and all major life table functions. It addresses a longstanding gap in demographic methodology, particularly in the analysis of small populations and subnational areas. Building on the Poisson nature of death counts, we derive a coherent approach to quantifying uncertainty in age-specific mortality and propagating it systematically across the life table. The framework is expressed through a matrix algebra formulation that unifies the computation of key life table quantities, including survivorship, deaths, person-years, total years remaining, and life expectancy. All analytical results are derived explicitly, providing both a practical toolkit for uncertainty estimation and a rigorous reinterpretation of the life table as an interconnected stochastic system. The approach is illustrated using 2020 female mortality data from the 20 administrative districts (arrondissements) of Paris, demonstrating how uncertainty varies with population size and mortality levels.

Germaine Uwimpuhwe: Dentifying core components affecting effect sizes in educational trials

Authors: Germaine Uwimpuhwe, Durham University; Sungkyung Kang, Durham University; Mohammad Sayari, Durham University; Akansha Singh, Durham University; Nasima Akhter, Teesside University; Jochen Einbeck, Durham University

Abstract: Educational interventions often consist of multiple interacting components, making it difficult to identify the drivers of effectiveness. Using data from Education Endowment Foundation (EEF)-funded trials, we propose a weighted component-wise meta-analysis incorporating component-specific weights into a random-effects framework. This approach estimates component-level effects while accounting for between-study variability and multiple outcomes per trial. To capture joint effects, we also consider regularised regression and tree-based methods. The approaches consistently identify a small number of influential components.

Arnošt Komárek: Fusion of MCMC and Variational Bayes for posterior approximation in hierarchical models

Authors: Jan Vávra, Charles University; Arnošt Komárek, Charles University

Abstract: In Bayesian statistics, Markov chain Monte Carlo (MCMC) became a gold standard for posterior inference. Nevertheless, it may require a painful amount of computational resources. A newer approach, Variational Inference (VI) is less precise, nevertheless, almost for free with respect to computational power. The purpose of the contribution is to discuss possibilities on how to combine the two approaches in order to improve the posterior inference in particular in context of statistical hierarchical models.

Øystein Evandt: A procedure and a calculator to help determine equal group sizes for confidence intervals for the effect size related to Student’s two-sample t-test

Authors: Øystein Evandt, Evandt Indstrial Statistics

Abstract: Several weaknesses of p-values from Student’s two-sample t-test related to the effect measure difference in means are pointed out. Advantages of confidence intervals over p-values in this connection are discussed. A procedure to help determine equal sample sizes for computing confidence intervals for the difference in means in a rational way is proposed. A calculator in EXCEL for applying this procedure is developed. An effect measure that can supplement the effect measure difference in means in some applications is proposed.

Chiara Rita Cunsolo: Applying a response-time-based hierarchical latent response model to detect engagement responding in a digital competence test

Authors: Chiara Rita Cunsolo, University of Bologna; Mariagiulia Matteucci, University of Bologna; Stefania Mignani, University of Bologna

Abstract: In educational assessment contexts, disengaged responding in both cognitive and non-cognitive tests may distort latent trait estimates and compromise the validity of inferences. The aim of this study is to apply a response-time-based hierarchical latent response model to investigate examinee engagement in INVALSI DigComp data. The model captures item-by-person variation in engagement by jointly modeling item responses and response times. We employ a simplified version of the model proposed by Ulitzsch et al. (2020), excluding the omission process. Model parameters are estimated within a Bayesian framework using Hamiltonian Monte Carlo implemented in Stan. Results indicate that disengaged responses are relatively rare but associated with a lower proportion of correct responses and shorter response times, consistent with rapid guessing behaviour.

Ryan Marinelli: Bayesian Benchmarking: Applications of Statistics for Efficient Model Ranking

Author: Ryan Marinelli, University of Oslo

Abstract: This work explores more efficient benchmarking techniques to com- pare the performance between models. By leveraging Bayesian techniques, it is possible to confidently rank models by their performance without excessive compute.

Ingrid Mauerer: Regularization for Differential Item Functioning in Gender Equality Attitudes

Authors: Ingrid Mauerer, University of Malaga; Maria Iannario, University of Naples Federico II; Gerhard Tutz, LMU Munich

Abstract: The contribution presents a regularization-based approach for detecting differential item functioning within the Generalized Partial Credit Model for the analysis of gender equality attitudes. It enables the simultaneous inclusion of multiple covariates and yields sparse item-specific effects. Using data from the European Values Study, we find substantive yet heterogeneous DIF effects across countries, with gender as the primary driver. The findings underscore the importance of accounting for DIF in cross-group comparisons in the study of gender equality attitudes.

Sabine K. Schnabel: Modelling high throughput phenotyping data over time and space - an example of plant height in wild potato

Authors: Sabine K. Schnabel, Wageningen University & Research; Luigi Fiorentino, Wageningen University & Research; Rick van de Zedde, Wageningen University & Research; Vivianne Vleeshouwers, Wageningen University & Research

Abstract: High-throughput plant phenotyping facilities produce large amounts of data. In addition to classic measurement involving handling the plants often frequent imaging is used to determine a number of traits of interest. Here, data from a study on wild potato subject to inoculation with the fungal plant pathogen Verticillium dahliae is analyzed as an example. Plant height of more than 500 genotypes is modelled using generalized additive (mixed) models as well as within the SpATS framework allowing for both incorporation of trends over space and time.

Iuliia Trifonova: Federated Learning for Ancestry-Aware Cancer Subtype Classification

Authors: Iuliia Trifonova, University of Salzburg; Daniel Katzlberger, University of Salzburg; Natalia Nunes, University of Salzburg; Nikolaus Fortelny, University of Salzburg; Georg Zimmermann, University of Salzburg

Abstract: Large-scale genomic datasets are biased towards individuals of European (EUR) ancestry, causing predictive models to generalise poorly across populations. We apply a federated learning (FL) framework to The Cancer Genome Atlas (TCGA) RNA-seq data for cancer subtype classification, treating two cohorts with different ancestry compositions as federated clients. Ancestry heterogeneity induces non-independent and non-identically distributed (non-IID) structure analogous to cross-silo FL, motivating federated optimisation as a principled solution. We compare five models — centralised ElasticNet, Federated Averaging (FedAvg), Federated Proximal (FedProx), ancestry-reweighted centralised, and their combination — evaluating overall predictive performance and ancestry-specific fairness across African (AFR), East Asian (EAS), South Asian (SAS), and admixed American (AMR) populations.

Philip Eigen: A Novel Benford-Based Approach for Detecting Anomalies in Election Data

Authors: Philip Eigen, Advancing American Freedom; Steven J. Miller, Williams College; Kevin Dayaratna, Advancing American Freedom

Abstract: We introduce L2-BL 10, an extension of Benford’s Law that analyzes the joint distribution of final digit pairs rather than leading digits. We derive closed-form matched-pair probabilities and establish their convergence to 1/10. Using discrete Weibull modeling and Monte Carlo chi-squared testing on precinct-level vote counts from nine U.S. states in the 2016 and 2020 presidential elections, the method flags 50 county–candidate combinations whose digit patterns deviate from expected behavior.

Maria Xose Rodriguez Alvarez: A general framework for estimating ROC curves with and without covariates

Authors: Maria Xose Rodriguez Alvarez, CITMAga, Universidade de Vigo; Philippe Lambert, Université de Liège; Vanda Inácio, University of Edinburgh

Abstract: We propose a unified framework for estimating pooled, covariate-specific, and covariate-adjusted ROC curves. The key insight is that these three functionals can all be expressed as cumulative distribution functions of suitable diseased placement values. This allows estimation to be reduced to modelling either the marginal or the conditional distribution of transformed placement values. To this end, we use the nonparametric double additive location-scale model (DALSM) in two stages: first, to estimate the relevant reference distribution in the nondiseased population, and second, to model the transformed placement values in the diseased population. The three ROC curves are then obtained within the same framework by simply including or excluding covariates in each of these two stages.

Pawan Lamsal: Multistate Models– Comparison of Bayesian and Frequentist Approaches

Authors: Pawan Lamsal, Johannes Kepler University Linz; Jutta Gampe, Max Planck Institute for Demographic Research, Johannes Kepler University Linz; Helga Wagner, Johannes Kepler University Linz

Abstract: Multistate models offer rich and flexible tools to describe processes where individual observations change states over time. Key parameters of a multistate model are the transition rates, often modulated by covariates in a proportional hazards setting. As in many longitudinal settings, intermittent observations, leading to interval-censored data, can complicate the analysis. Estimation of transition rates and regression parameters predominantly is performed in a frequentist framework and Bayesian analyses of multistate models are rare. Here we perform a thorough comparison of frequentist and Bayesian approaches to devise (dis)advantages of the two methods in view of model complexity, sample sizes, observation schemes and also in view of further model extensions by frailties.

Roman Pfeiler: Analyzing retirement data from Austria with a flexible Bayesian panel data model

Authors: Roman Pfeiler, JKU Linz; Helga Wagner, JKU Linz

Abstract: We consider a Bayesian panel data model for analyzing the effect of retirement on health expenditures using administrative data from Austria. We model selection into retirement with a Probit and health expenditures with a Hurdle model. In both models we use time-varying regression effects together with random walk shrinkage priors to prevent the model from overfitting. We find evidence that individuals with higher income as well as blue-collar workers have an increased probability of entering retirement at a given point in time.

Edoardo Pandolfo: Modeling university outcomes: a hidden Markov model approach

Authors: Edoardo Pandolfo, Sapienza University of Rome; Nina Deliu, Sapienza University of Rome; Brunero Liseo, Sapienza University of Rome; Andrea Tancredi, Sapienza University of Rome

Abstract: In this paper, we propose a hidden Markov model to investigate late graduation and dropout outcomes for university students by acounting for latent activity and inactivity periods in their careers. Data are based on the whole cohort of University La Sapienza students at the faculty of Economics enrolled in the academic year 2017/18. Model estimation is performed via a Bayesian approach. A Gibbs sampler algorithm is used to produce posterior estimates.

Btissame El Mahtout: Fast Training of Mixture-of-Experts for Electricity Price Forecasting

Authors: Btissame El Mahtout, Duisburg Essen University; Florian Ziel, Duisburg Essen University

Abstract: This paper proposes a novel Mixture-of-Experts (MoE) framework designed for day-ahead electricity price forecasting in European power markets. The method incorporates expert-specific loss information directly into the training process to enhance expert specialization, improve regime differentiation, and increase utilization of diverse market signals. We introduce an online learning strategy that incrementally updates both the gating mechanism and expert parameters, thereby substantially reducing computational effort by avoiding repeated full model retraining. By combining expert-level loss integration with different calibration window sizes and efficient online optimization, the proposed approach improves learning efficiency while preserving strong predictive accuracy. Empirical results on major European day-ahead markets demonstrate faster convergence, more stable expert allocation across price regimes, and improved forecast performance compared to the state-of-the-art benchmark

Sina Mews: Modelling momentum in tennis: A latent-state approach to point outcomes and rally lengths

Authors: Sina Mews, Georg-August-University of Göttingen; Jan-Ole Fischer, Bielefeld University; Rouven Michels, TU Dortmund University; Christian Deutscher, Bielefeld University

Abstract: We investigate momentum in tennis by jointly modelling point outcomes and rally lengths using a Markov-modulated marked Poisson process (MMMPP), treating the underlying momentum as a latent state. By analyzing point-by-point data from ATP Grand Slam tournaments, we identify momentum shifts within matches, accounting for player- and match-specific effects.

Sigurd Myhre Vittersø: BGET: Bayesian Genetic Ensemble of Trees

Authors: Sigurd Myhre Vittersø, Norwegian University of Life Sciences; Aliaksandr Hubin, Norwegian University of Life Sciences

Abstract: We embed classification and regression trees in the Bayesian generalized nonlinear model framework, defining a Bayesian genetic ensemble of trees with complexity-penalizing priors on tree structures. Inference is performed via genetically modified mode jumping MCMC, which explores large spaces of tree ensembles using posterior inclusion probabilities to guide genetic operators. The approach is evaluated on the Abalone and Boston housing data and performs on par with relevant related baselines

David Winkelmann: Statistical Modelling of Live‑Betting Behaviour to Identify Match‑Fixing

Authors: David Winkelmann, University of Vienna; Maya Vienken, Bielefeld University; Christian Deutscher, Bielefeld Universtiy; Roland Langrock, Bielefeld University

Abstract: Match-fixing undermines the integrity of sport by eroding public trust and threatening the financial sustainability of clubs and leagues. The global expansion of sports betting markets has created new incentives and opportunities for manipulation, calling for rigorous, data-driven monitoring tools. Football, which accounts for the largest share of global betting turnover, remains particularly exposed: integrity reports continue to flag several suspicious matches, with past scandals in Italy and Turkey underlining the problem's persistence. This study uses high-frequency live-betting data from the Italian Serie B (2018/19-2020/21) to explore statistical approaches for detecting abnormal betting behaviour. A state-space modelling framework is employed to describe standard betting market dynamics and to predict expected betting volumes conditional on match events. Deviations from these expectations are then analysed using outlier detection techniques to identify potentially suspicious periods. The results demonstrate how statistical modelling can contribute to the early identification of irregular betting patterns and thereby support integrity assurance in live sports betting markets.

María José Ginzo Villamayor: Surnames and Population Movements in Galicia (Spain): Analysis by Cohorts and Major Cities

Authors: María J. Ginzo-Villamayor, University of Santiago de Compostela, Galician Centre for Mathematical Research and Technology (CITMAga)

Abstract: Surnames are used to analyze population dynamics and migration patterns in Galicia across three periods (pre-1945, 1945–1965, post-1965). Migration flows are normalized using a hyperbolic tangent transformation with a regularization parameter to control the influence of municipality size and ensure robust comparisons. The analysis is conducted at three levels: global, major cities, and specific surnames. Data come from the 2011 Galicia Census (IGE), disaggregated by municipality and age cohorts.

Jungbok Lee: Model Derived Estimate based Cutpoints Determination of Survival Data with Non-Linear Increase Hazard

Author: Jungbok Lee, Asan Medical Center & University of Ulsan

Abstract: In this study, we introduce a method for determining the cut-off using a model-derived estimate rather than a method using test statistics. We present a procedure for finding the cut-off using a β-derivative hazard that represents the non-linear change in hazard. Using the PELT algorithm, we propose a procedure to find only significant change points by pruning change points with similar β-derivative hazard. Additionally, using skewed data as an example, we will examine the differences from methods using test statistics.

Kenan M. Matawie: Statistical Analysis of Query Transformation and Indexing Configurations in Information Retrieval

Authors: Kenan M. Matawie, Western Sydney University; Sargon Hasso, Illinois Institute of Technology

Abstract: Traditional information retrieval metrics like MAP and NDCG quantify effectiveness but often fail to reveal the behavioral differences between algorithms. This study investigates both retrieval performance and document overlap across 225 queries, eleven indexing configurations, and multiple query transformation strategies. Using a factorial ANOVA, we identify significant performance variations driven by query formulation and indexing methods. Furthermore, Jaccard similarity analysis and visual diagrams highlight how different strategies retrieve distinct document sets. By combining statistical performance evaluation with document-level overlap analysis, this framework provides deeper insights into retrieval diversity, supporting the development of more effective ensemble retrieval systems.

Mostafa Raeisi Sarkandiz: Joint Estimation of the Lag Parameter of the Threshold Variable in a STAR Model

Authors: Mostafa Raeisi Sarkandiz, University of Palermo; Sara Ghayekhloo, University of Palermo; Vito M. R. Muggeo, University of Palermo

Abstract: A central challenge in nonlinear time series models such as the Smooth Transition Autoregressive (STAR) model is the selection of the lag parameter for the threshold variable. Existing approaches typically rely on computationally demanding two-stage procedures based on brute-force grid searches over candidate lags. This paper proposes a computationally efficient framework for the joint estimation of the discrete lag parameter and the continuous parameters of a two-regime STAR model. The key innovation is an indicator-function-based reparameterization that renders the lag index an explicit estimable parameter. Monte Carlo simulations demonstrate the effectiveness of the proposed nested sigmoid approach, which achieves accurate identification of the true lag parameter for almost all sample sizes.

Stefania Lando: Sample Size Determination for Non-Linear Effects in Cox Models Using Restricted Cubic Splines

Authors: Stefania Lando, University of Padova, Pediatric Research Institute “Città della Speranza”; Ester Rosa, University of Padova, Pediatric Research Institute “Città della Speranza”; Gloria Brigiari, University of Padova, Pediatric Research Institute “Città della Speranza”; Dario Gregori, University of Padova, Pediatric Research Institute “Città della Speranza”

Abstract: Flexible functions such as restricted cubic splines (RCS) are used to model non-linear covariate effects in Cox regression, yet sample size calculations still typically assume a 1-degree of freedom (df) linear effect or rely on ad hoc events per variable rules. We propose Non-Linear Power (NlP), a design-by-shape framework that powers the planned multi-df hypothesis test of an RCS term by mapping a clinically meaningful hazard ratio curve over the covariate range to spline coefficients, yielding a non-central chi-square approximation with a non-centrality parameter. To link confirmatory inference and prediction modelling, we quantify a “parameter tax” based on Riley-style shrinkage targets to quantify the added events required when moving from linear effects to RCS. In an illustration with anticipated R^2 = 0.15, target shrinkage 0.9, and event fraction 0.10, moving from a linear effect to an RCS with 3 and 4 knots increases the stability-driven required events from around 12 to 24 and 36, respectively. These calculations make the df–stability trade-off explicit and help align sample size planning with the intended non-linear analysis.

Ester Rosa: Kolmogorov–Arnold Networks as Structured Penalized Superposition Models

Authors: Ester Rosa, University of Padova; Stefania Lando, University of Padova; Gloria Brigiari, University of Padova; Dario Gregori, University of Padova

Abstract: Kolmogorov–Arnold Networks (KANs) replace scalar weights with learnable univariate functions along network edges, yielding a structured superposition architecture inspired by Kolmogorov’s representation theorem. When parameterized with B-spline bases and difference penalties, KANs admit a finite-dimensional formulation as penalized likelihood estimators within a generalized regression framework. We develop a spline-based penalized log-likelihood formulation and characterize the resulting block-structured regularization matrix. Under an iteratively reweighted least squares (IRLS) linearization, we derive an approximate effective degrees-of-freedom (DoF) expression given by trace of a generalized influence operator. This provides a continuous measure of functional complexity linking spline basis size, network width, and smoothing strength. The estimator can be viewed as a structured extension of additive spline models, interpolating between generalized linear models and connected neural networks. An illustrative nonlinearity-graded experiment shows how the expanded superposition class reduces approximation bias in nonlinear regimes while preserving interpretable univariate components

Forough Pazhuheian: Personalised Longitudinal Clinical Monitoring Using Multivariate Adaptive Reference Regions

Authors: Forough Pazhuheian, University of Galway; John Newell, University of Galway; Davood Roshan, University of Galway

Abstract: Clinical reference ranges are essential for interpreting laboratory results, but traditional methods are static and population based, limiting their value for individualized longitudinal monitoring. Adaptive reference ranges overcome this limitation by updating subject-specific limits as new observations are collected, enabling learning of personal biomarker trajectories over time. Most existing adaptive methods focus on single biomarkers, whereas many clinical applications measure multiple correlated outcomes simultaneously. Monitoring biomarkers independently may fail to identify abnormal joint patterns. We extend adaptive reference ranges to a multivariate framework by jointly modelling multiple longitudinal biomarkers using a multivariate mixed-effects model. The proposed method produces recursively updated reference regions, ellipsoids in the multivariate space, that account for within-subject variability and between-biomarker correlations. This approach improves detection of atypical joint trajectories, as illustrated with longitudinal athlete biomarker data.

Paolo Girardi: Semi-Parametric Methods for Long-Term Trend Detection in Air Quality: the North-Eastern Italian Alps Case

Authors: Paolo Girardi, Ca' Foscari University of Venice; Claudia Collarin, University of Edinburgh; Mauro Masiol, Ca' Foscari University of Venice; Ilaria Prosdocimi, Ca' Foscari University of Venice

Abstract: Assessing long-term air quality trends is essential for evaluating environmental policies and public health impacts. However, air pollution data are often characterised by skewness, temporal dependence, and censoring due to detection limits. We propose a semi-parametric framework based on Generalised Additive Models (GAMs) integrating a data-driven Box--Cox transformation and a Tobit likelihood. The approach flexibly captures non-linear dynamics while properly accounting for left-censored observations. The methodology is applied to daily concentrations of NO, NO2, O3, PM10, and PM2.5 from 40 monitoring stations in the North-Eastern Italian Alps (2009-2023). Results indicate decreasing trends for NO2 and particulate matter, while ozone remains stable or shows a slight increase, with strong spatial heterogeneity.

Lena Schemet: Structural Limitations of PoSI in Cox Regression

Authors: Lena Schemet, University of Augsburg; Sarah Friedrich-Welz, University of Augsburg

Abstract: Uniform post-selection inference (PoSI) guarantees simultaneous coverage over a prespecified collection of submodels. While such procedures are well understood in Gaussian linear models, their transfer to Cox proportional hazards regression has not been critically examined. We study two asymptotically valid Cox-PoSI constructions, a standardized max-$|Z|$ approach and a score-based supremum approach. We show that in both cases the resulting constants are structurally larger than full-model simultaneous corrections. This inflation is intrinsic to maximization over submodels and persists under ideal Gaussian approximation. Simulation results illustrate that even moderate model spaces yield practically infeasible intervals. A real data illustration using the TCGA-BRCA cohort demonstrates the resulting interval inflation in practice.

Cristina Cametti: The performance of non-adaptive and adaptive spatial sampling designs in low-resource settings

Authors: Cristina Cametti, KU Leuven; Emanuele Giorgi, University of Birmingham; Daisy Lou L. Polestico, Mindanao State University; Thomas Neyens, KU Leuven, Universiteit Hasselt

Abstract: In low-resource settings, spatial data are often collected sequentially in small batches, raising questions about efficient sampling design. We compare non-adaptive designs with prediction variance driven adaptive geostatistical designs under small initial sample sizes. Through a simulation study, we evaluate deterministic and randomized prediction variance based strategies, with and without spatial inhibition, using continuous and binary mean squared error. Results show that non-adaptive designs perform better at very small sample sizes, while prediction variance driven adaptive designs become competitive as more data are collected.

Dae-Jin Lee: Performance aging curves via population functional envelopes

Author: Dae-Jin Lee, IE University

Abstract: Many longitudinal performance metrics are governed by functional limits rather than average behavior, making mean-centered models poorly suited when the scientific question concerns performance relative to an upper limit. We present PACE (Performance Aging Curves via Envelopes), which represents age-related trajectories relative to a population-level functional envelope representing the upper attainable range of performance. The envelope is estimated by smooth conditional expectile regression with shape constraints and treated as a fixed population reference. Individual trajectories are modeled via STAR (Shape Translation And Rotation), an interpretable parameterization capturing differences in level, timing, and tempo through geometric transformations of the envelope. Identifiability depends on envelope geometry: near-linear envelopes (sprint speed) reduce to level and tempo, while non-linear envelopes (bolt rate) identify all three parameters. Using MLB Statcast data, sprint speed shows a strong negative level–tempo correlation (r = -0.89), a performance–longevity tradeoff, whereas bolt rate shows a positive correlation (r = 0.31), indicating distinct aging dynamics for continuous versus threshold-based metrics.