"Am I a statistician, data miner, data scientist or analyst?"

Abstract:

This talk is given in the context of the 25 year anniversary of the statistical consulting course at UCLouvain.

Eric Lecoutre attended this course in the academic year 1999/2000 and has since then explored many aspects of life as a professional in the data sector. He started his career as statistical and computing consultant at the Institut de Statistique at UCLouvain, contributed important R packages including R2HTML, went to work for the company Business and Decision, contributed to the Data Science Hub in Brussels, and founded his own company WeLoveDataScience.
In his talk, he will retrace important steps in his professional journey and point out what may be important for statisticians and data scientists who are entering the market of jobs and opportunities of the data world around Brussels and Louvain-la-Neuve. This will include an analysis of the bewildering variety of profiles published in job adds, a discussion of roles of R and Python as programming languages in the data field, and anecdotes of the joys and frustrations in his career.
Note: Since this talk may draw a larger audience, it will probably take place in the lecture hall Barb 91 but please check the location a few days before the seminar. There will be (limited) remote access via Teams.

Teams link for people who would like to join remotely:
TEAMS LINK
 

Multiple Speakers

There will be several short presentations and an exchange of ideas related to teaching statistical consultancy.

•    16:00: Christian Ritter (UCLouvain) : Greetings and a brief report on a survey among past participants of the statistical consulting course

•    16:15-17h:00: Several short presentations by past participants related to their career and to how a statistical consulting course can help

Teams link for people who would like to join remotely:
TEAMS LINK

Note:

• 17:30 : The session will be followed by a reception for alumni of this course and new participants
  - location : Hall Auditoire BARB 91
   
• 19:00 : The alumni dinner will follow 
  - location : TBA - Reservation required

Abstract:

Statisticians at W.L. Gore & Associates partner with engineers and scientists across the organization to ensure data driven decisions when it comes to quality and reliability in all aspects of the production process. In this seminar we will present how we have applied an OMARS design to optimize a chemical batch process spanning over three process steps. We will briefly explain the application, a polymerization process, followed by an introduction to the concept of OMARS designs as developed by José Núñez Ares & Peter Goos[1]. This new class of designs allowed us to combine screening and optimization phase of the product development cycle. In total, we studied 8 factors of the three-step manufacturing process. During the first part of the seminar, we invite you to accompany our journey of DoE planning and analysis. In the second part of the seminar, we will give the opportunity to do a hands-on analysis of the data yourself.

[1] José Núñez Ares & Peter Goos (2020) Enumeration and Multicriteria Selection of Orthogonal Minimally Aliased Response Surface Designs, Technometrics, 62:1, 21-36, DOI: 10.1080/00401706.2018.1549103

Teams link for people who would like to join remotely:
TEAMS LINK

Abstract:
The growing complexity of experiments in biology requires appropriate data analysis methods for establishing how firmly certain biological functions can be identified from the data. One such biological function relates to how the human brain functions and is arguably one of the most challenging problems in science today. Recent advances in neuroscience have made the landmark discovery of certain types of neurons in the brain, termed grid cells, and have highlighted their role as the most plausible mechanism for how the brain performs a range of tasks including spatial navigation. We argue that statistical modelling of the firing activity of grid cells has not been dealt with in-depth and that commonly employed approaches may potentially mask the effect of key features such as the direction of navigation of the animal as well as residual temporal variation. To our knowledge, there is no formal statistical framework developed so far that can facilitate their inclusion of such features. As a result, we are led to develop likelihood-based procedures for modelling and estimating the firing activity of grid cells conditionally on biologically relevant covariates. Our approach rests on modelling the firing activity of cells using Poisson point process on trajectories of animals with latent Gaussian effects defined in the environment that the animal explores. The latent prior Gaussian effects accommodate for overdispersion and are carefully chosen so that they mimic closely the behaviour of the firing activity from grid cells whilst accounting for unexplained variation. Inference is performed in a fully Bayesian manner which allows us to quantify uncertainty and provide evidence that supports the hypothesis of the presence of effects that are typically missed out from most of the existing analyses.

"Sparse Models for Sparse Networks"

Abstract:: Network data are frequently collected in modern society and science.  Stylized features of a typical network include network sparsity, degree heterogeneity and homophily among many others. This talk introduces a framework with a class of sparse models that utilize parameters to explicitly account for these network features. In particular, degree heterogeneity is handled by node-specific parameters while homophily is captured by the use of covariates. To avoid over-parametrization due to the former, we differentially assign parameters to nodes that are important in certain sense. We start by discussing the sparse \beta model when no covariates are present, and proceed to discuss a generalized model to include covariates. Interestingly for the former we can use \ell_0 penalization to identify and estimate the heterogeneity parameters, while for the latter we resort to penalized logistic regression with an \ell_1 penalty, thus immediately connecting our methodology to the lasso literature. Along the way, we demonstrate the fallacy of what we call data-selective inference, a common practice in the literature to discard less well-connected nodes in order to fit a model, which can be of independent interest.

Abstract:
As a major dimension reduction tool, the idiosyncratic components of a tensor time series factor model can exhibit serial correlations, especially in financial and economic applications. This rules out a lot of state-of-the-art methods that assume white idiosyncratic components, or even independent/Gaussian data.
While the traditional higher order orthogonal iteration (HOOI) is proved to be convergent to a set of factor loading matrices, the closeness of them to the true underlying factor loading matrices are in general not established, or only under i.i.d. Gaussian noises (Zhang and Xia, 2018). Under the presence of serial and cross-correlations in the idiosyncratic components and time series variables with only bounded fourth order moments, we propose a pre-averaging method that accumulates information from tensor fibres for better estimating all the factor loading spaces. The estimated directions corresponding to the strongest factors are then used for projecting the data for a potentially improved re-estimation of the factor loading spaces themselves, with theoretical guarantees and rate of convergence spelt out. We also propose a new rank estimation method which utilises correlation information from the projected data, in the same spirit as Fan et.al. (2022) for factor models with independent data. Extensive simulations are performed and compared to other state-of-the-art or traditional alternatives. A set of matrix-valued portfolio return data is also analysed.

ROOM D.251 (2d FLOOR) + ONLINE (TEAMS LINK HERE : TEAMS)

Abstract :
In this talk we consider a simple supervised classification problem for object recognition on grayscale images. There are two possible perspectives to solve this problem. Firstly, one can interpret object recognition as a high-dimensional classification problem, where every pixel is a variable. The task is then to map these pixel values to the conditional class probabilities or the labels. Increasing the dimension makes the problem considerably harder, leading to slow convergence rates due to the curse of dimensionality. Another perspective is to view images as two-dimensional objects. Increasing the number of pixels leads to higher resolution and therefore better performance is expected for large images. Following the second route, we present a new image deformation model, for which we propose and analyze two different classifiers. The first method estimates the image deformation by support alignment. Under a minimal separation condition, it is shown that perfect classification is possible. The second method fits a CNN to the data. We derive a rate for the misclassification error depending on the sample size and the number of pixels $d^2$. Under suitable conditions, this rate is of order $1/\sqrt{d}.$ Because of the setting we have chosen, not only are our methods not affected by the curse of high dimension, but they actually improve with increasing dimension.

09:05 – 09:50 : Pierre Devolder & Oussama Belhouari
"Multi-step valuation methods for hybrid life insurance products in a stochastic interest rate framework"

Abstract
In a complete financial market, financial products are valued with the risk- neutral measure and these products are completely hedgeable. In life insurance, the approach is different as the valuation is based on an insurance premium principle which includes a safety loading. The insurer reduces the risk by pooling a vast number of independent contracts. In our framework, we suggest valuations of a class of products that are dependent on both mortality and finance risk, namely hybrid life products. The aim of this paper is to generalize different valuation operators suggested in the literature into a stochastic interest rate framework. We illustrate our methods with a classical application, namely a Pure Endowment with profit. Several numerical results are presented, and an extensive sensitivity analysis is included.

09:50 – 10:20 : Ensiyeh Nezakati
"Distributed estimation of Covariate-adjusted Gaussian graphical models"

Abstract
With the development of technology, the size of the datasets grows at a high rate, such that in certain situations it is not possible to store all needed datasets in the memory of one single machine. Moreover, in recent frameworks, like federated learning, due to privacy concerns, it may be impossible to collect datasets from different resources on one single central machine. As such, the dataset is partitioned onto a cluster of parallel machines. Distributed statistical approaches, also known as ‘divide and conquer’ approaches, have drawn a lot of attention in the last decade and have been developed for various statistical problems. Moreover, inverse covariance matrix estimation plays an important role in statistical and machine learning framework, especially in the framework of Gaussian graphical modeling. Most current methods for inverse covariance matrix estimation assume that the random vector has zero or constant mean. However, in many real applications, like genomic data analysis, it is often important to adjust for covariate effects on the mean of the random vector to obtain more precise estimates. Our purpose is to propose new, unbalanced distributed estimators for both the mean structure and the inverse covariance matrix for covariate-adjusted Gaussian graphical models. These estimators aggregate all local parallel estimators into the final ones by maximizing the pseudo log-likelihood function which comes from the asymptotic distribution of K debiased estimators. Asymptotic behavior and statistical guarantees of these estimators are provided when the number of parameters, covariates and machines all grow with the sample size. A simulation study and a real data example are used to assess the performance of these estimators.

10:20 – 10:50 : Coffee break 

10:50 – 11:20 : Aigerim Zhuman
"Combination of Control Variates and Adaptive Importance Sampling"

Abstract
Adaptive importance sampling and control variates are two widely used variance reduction techniques associated with Monte Carlo integration. The adaptive importance sampling method is based on updating the sampling policy, the sequence of distributions used to generate the particles. The method of control variates consists of projecting the integrand on the linear space spanned by a vector of auxiliary functions with known expectations, called control variates. We propose to incorporate control variates into the adaptive importance sampling procedure in order to improve the accuracy of Monte Carlo integration. The obtained estimate, called the AISCV estimate, arises as the weighted least squares estimate for the intercept in a multiple linear regression model where control variates are used as explanatory variables. Moreover, we introduce a quadrature rule with adapted quadrature weights which do not depend on the integrand. The latter property is computationally advantageous in case of multiple integrands. Our main result states a concentration inequality for the normalized AISCV integration error. The performance of the AISCV estimate is illustrated on synthetic examples and real-world data for Bayesian linear regression.

11:20 – 11:50 : Stephan Lhaut
"Uniform concentration bounds for frequencies of rare events"             

Abstract
Statistical Learning Theory stressed the problem of finite sample guarantees and motivated the study of concentration inequalities between a risk functional (unknown) to be minimized and its empirical version on the sample (known), holding uniformly over a class of functions of finite complexity. Many results are available, but in a multivariate extreme value context, these bounds may not capture well the rare nature of the encountered events and hence overestimate the difference between the true and empirical risk. In a binary classification framework, we propose specific inequalities designed to handle such events. The derived bounds are explicit, enabling numerical comparisons.

11:50 – 12:20 : Anas Mourahib
"Sparse multivariate Generalized Pareto distributions"

Abstract
Consider a random vector representing risk factors and suppose that we are interested in extreme scenarios. The Peaks Over Thresholds method is widely used in extreme value theory. It uses the fact that asymptotically, exceedances over a high threshold can be modelled using a multivariate Generalized Pareto distribution. In the literature, statistical practice of this method has been discussed only when all risks are large simultaneously. This condition is not realistic for example when some of the risks are nearly independent. To address this point, we construct a parametric model that allows some risks to be large without the other ones. We also compute the density of this model and estimate the parameters using Maximum Likelihood.  

12 :20 – 12:30 : Closing

Résumé:

Les essais cliniques se déroulent généralement sur une longue période. Pour des raisons éthiques, il faut s’assurer que les patients ne courent aucun danger durant toute la période de l’essai.  L’iDMC  (independent Data Monitoring Committee), est un groupe d’experts indépendants externe à l’étude, qui est en charge de vérifier les données au fil de l’étude.  Leur rôle est d’assurer que le rapport risque/bénéfice du médicament testé est toujours en faveur du patient. Ils doivent également vérifier que l’étude se déroule tel que décrit dans le protocole et, à la suite des analyses intérimaires,  aviser le sponsor quant à la continuité de l’étude. 

Workshop by TEAMS here : TEAMS LINK

In this talk, I shall present examples and theory about efficient and inefficient display of data in form of graphs and tables. Participants will learn a systematic approach for discussing and improving of data visualizations.

This is a variation of a slowly evolving annual talk. If you have seen it already, you may not learn a lot of new things but you might discover a new trick or a new perspective.

The talk will be given in room C.115, access via Teams is possible but we can't guarantee good quality.

Some of us will go for a beer after the talk (rdv at 17h30 at the Café des Halles), feel free to join us.

TEAMS + Room C.115

https://teams.microsoft.com/l/meetup-join/19%3a3577dab1932a4a6b8cd661a1b935e9eb%40thread.tacv2/1675261846658?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

Abstract :

The manufacture of vaccines uses a wide range of automated technologies. The manufacturing processes are controlled and adjusted by on-line sensors. The data generated by those sensors can be exploited by process experts as they become accessible to them. During this seminar, a live demonstration of the data analytics applied in real-time at GSK will be shown. Data analytics are applied to transform the data into information and then insights for the end user to accelerate decision making on the shop floor.

Abstract :

The Guidance on Estimands and Sensitivity analyses (ICH E9(R1)) in clinical trials was released in 2019. The need for the estimands framework arose out of the realization that there was a lack of transparency regarding the alignment between clinical trial objectives and study design, data collection and statistical analysis (Mallinckrodt, 2019,2021). The central aim of the estimand is exactly that, to bridge this gap by choosing and specifying beforehand 'what is to be estimated' and ‘how’ to properly evaluate and estimate it. By understanding 'what' is to be estimated, and deciding on 'how' to estimate it, it is easier to foresee and understand events which could hamper that process. More formally, the Guidance denotes these events as "intercurrent events" and defines them as "events occurring after treatment initiation that affect either the interpretation or the existence of measurements associated with the clinical questions of interest (ICH, p.17)". Intercurrent events must be identified upfront, and in conjunction with the population of interest, the treatment (regimen), the endpoint (response/outcome variable) and the population-level summary statistic. These five ingredients are the building blocks required to define an estimand. By making well-informed choices and specifying the five attributes, the trial study team offers a transparent summary regarding 'what' they are estimating, i.e. the estimand. Only once it becomes clear which estimand is of interest can the type of study design, the required data and the statistical analysis methods be specified. In other words, the primary objective drives the estimand choices which in turn inform the decisions regarding the study design, data collection and analysis choices.

This presentation will introduce the different steps on how to define estimands and what the role of the statistician is within the estimands framework. We will cover in depth the notion of intercurrent events and how these can introduce bias into the study. In addition, the different strategies proposed by the Guidance on handling intercurrent events will be introduced and illustrated through examples.

Workshop by TEAMS here : TEAMS

Abstract:
Joint models for longitudinal and time-to-event data are useful in situations where an association exists between a longitudinal marker and an event time. Examples are abundant in cancer trials as well as in degradation studies in reliability applications. In such contexts, separate models for the longitudinal and survival components that do not take into account the dependence produce inefficient results and are prone to bias. Models and methodologies for studying the longitudinal and the time-to-event processes simultaneously have been an active area of research for the past few decades. 

A lion’s share of research has been focused on studies where the underlying event occurrence process is governed by a single source of failure/death. Frequently however, one encounters time-to-event mechanism where the event is triggered at the onset of the earliest of multiple potential risks. Joint inference under such competing risks framework have almost exclusively been investigated through models of cause-specific hazards. 

By contrast, this talk will present some reports on joint models based on latent failure times. We shall carry out our investigation under a Bayesian framework which is naturally suited for joint models that are inherently hierarchical in nature. The methodology will be implemented on a dataset from a cancer clinical trial and will be supplemented by findings from extensive simulations.

TEAMS LINK here : TEAMS

 

Abstract:
The first part of the talk will present GDPR in the nutshell: key definitions, principles, and obligations. Thereafter, the application of different principles will be illustrated with the data cycle viewed from the angle of the Regulation. Finally, practical examples will enable students to better understand the notions of anonymization and de-identification tackling the edge of the scope of the Regulation as well as the worthiness and risks of data manipulation from the perspective of GDPR compliance.

Workshop given also by TEAMS here : TEAMS

Abstract :
Carbon pricing is seen as a crucial tool for achieving the set goals of carbon neutrality in Europe. This paper studies the current dimensions of carbon pricing, focusing on the EU emission trading system and country specific carbon taxes, to better understand their interaction with economic performances in Europe. We analyse and compare the different tools by means of two levels of structural VAR models. For the European wide context of the ETS we apply a monthly smooth transition SVAR model identified by non-Gaussianity and narrative sign restrictions and find heterogeneous effects over two regimes of economic activity. Further, we study effects of individually set carbon taxes in a new version of panel SVAR models which accounts for the joint market development and different economic sectors. Joining the results enables to discuss the effectiveness of current and potential future carbon pricing tools in Europe.

Abstract:
Motivated by crowd-sourcing applications where we want to rank experts according to their abilities, we consider a model where we have partial observations from a bivariate isotonic nXd matrix with an unknown permutation \pi^* acting on its rows. We consider the twin problems of recovering the permutation \pi^* and estimating the unknown matrix. We introduce a polynomial-time procedure achieving the minimax risk for these two problems, this for all possible values of n, d, and all possible sampling efforts. Along the way, we establish that, in some regimes, recovering the unknown permutation \pi^* is considerably simpler than estimating the matrix. 
This is based on a joint work with Alexandra Carpentier (U. Potsdam) and Emmanuel Pilliat (U. Montpellier).

More information (Web site)

We are happy to announce that the 30th Annual Meeting of the Royal Statistical Society of Belgium will take place in Louvain-la-Neuve on October 19 and 20, 2023 ! This edition of the annual meeting is organized by UCLouvain with the support of the Institute of Statistics, Biostatistics and Actuarial Sciences.

Program : Soon available

Abstract :
In clinical research, surrogate (or intermediate) endpoints can be used instead of final endpoints to help speed up the evaluation of a new treatment or intervention. A common example is the use in oncology of the time-to-progression of a cancer as a surrogate of the overall survival. However, before any use putative surrogates must first be statistically assessed to be good surrogate.
In this work we propose the validation of surrogate endpoints using mediation analysis which aims at decomposing the effect of the treatment on the final endpoint into an indirect effect through the surrogate endpoint and a direct effect independent of it. A surrogate endpoint will be validated if most of the treatment effect corresponds to the indirect effect (the treatment therefore mainly operates through the surrogate).
We are particularly interested in the case where the final endpoint is a time-to-event and the surrogate endpoint is either also a time-to-event or a longitudinal biomarker. Joint models have been developed to properly take into account the relashionship between such endpoints. From these models the direct and indirect treatment effects can be derived whose causal interpretations can be established under given assumptions. Moreover, heterogeneous data from meta-analyses or multicenter studies can be used to strengthen the validation process.
We illustrate this approach with two examples in oncology: with a meta-analysis in resectable gastric cancer to evaluate the time-to-relapse as a surrogate for the overall survival and with a multicenter trial in locally advanced prostate cancer to evaluate the evolution over time of the prostate-specific antigen as a surrogate for the disease-free survival.

09:10 – 09:40: Fanny Hoogstoel

“Right to be forgotten”: how to estimate excess risk mortality when no disease-specific registry data are available

Facilitating access to insurance products for people with a perceived aggravated risk has recently become a hot topic, with in particular the adoption of the "right to be forgotten" in France (in 2016) and Belgium (in 2019). New studies based in part on relative survival analysis are currently being carried out to extend the right to be forgotten in Belgium (Soetwey et al. 2020) in order to reach the same level as in France, to include other pathologies and to also include other insurance products.

A fundamental element is therefore access to follow-up data, both for the “general population” and for cohorts suffering from the pathology studied. However, such data are often non-existent, incomplete or unreliable.

Concerning the survival of people suffering from a given pathology for which there is no appropriate registry data available, a possible approach could be to use data from cohorts or registers from other countries and to "adjust" them to the specificities of the country of interest, in the footsteps of the work carried out by Touraine et al. 2020 in particular. In this work, we propose first a simulation algorithm that allows to compare existing methods and second a new proposal to adjust such data via the use of generalized additive models (GAMs). This will be illustrated with an application to an insurance context in the framework of the right to be forgotten.

09:40 – 10:30: Oskar Laverny

Version control for academics

Version controls systems (VCS) are systems that keep track of changes made to a collection of files and folders. The usage of such systems allows for reproducibility of the work done and largely facilitates collaborations. We start by arguing that using such a version control system is a must for any serious work that boils down to text files, which is the case of much of academic workflows. Then we introduce the de-facto standard VCS nowadays, git, through a bottom-up exposition of its internal data model. We end by a review of potential applications to academic workflows, from reproducibility of numerical results to templating of articles and facilitating review processes, showing the usefulness of the concept.

10:30 – 11:00: Coffee break

11:00 – 11:30: Keivan Diakité

Longevity heterogeneity correction and demographic risk sharing in public pension systems

Many pension systems require and continue to need change to maintain long-term sustainability. Most developed countries' policymakers have been pushed to reform their pension systems in order to preserve or re-establish financial sustainability. As longevity heterogeneity is frequently disregarded in the policy chosen, the pursuit of financial sustainability is done at the expense of intra-generational actuarial fairness.

We propose a pension system management system comprising two adaptation mechanisms.

The first dynamic mechanism is integrated directly into the pension formula and corrects the heterogeneity of longevity when it exists between the agents of the pension scheme.

A steering mechanism for both the contribution rate and the mean benefit ratio respects Musgrave's rule, which makes it possible to distribute the demographic risk between working people and retirees.

In order to capture both effects and incorporate a mortality component into the pension formula, we model longevity heterogeneity and ageing in a pension scheme using different multi-population mortality models.

11:30 – 12:00: Jeongjin Lee

Partial Tail Correlations for Extremes

We develop a method for investigating conditional relationships between variables at their extreme levels. We consider a vector space constructed from transformed-linear combinations of independent regularly varying random variables. The notion of partial tail correlation is developed through projection theorem.

We show that the partial tail correlation can be understood as the inner product of prediction errors associated with the best transformed-linear prediction. Using a subset of the inner product space as a modeling framework, we connect partial tail correlation to the inverse of the inner product matrix and show that a zero in this inverse implies a partial tail correlation of zero. We develop a hypothesis test for partial tail correlation of zero and demonstrate the performance in two applications: high nitrogen dioxide levels in Washington DC and extreme river discharges in the upper Danube basin.

12 :00 – 12:30: Patricia Ortega Jimenez

Efron’s monotonicity under given copula structures

Given a multivariate random vector, Efron’s marginal monotonicity (EMM) refers to the stochastic monotonicity of the variables given the value of their sum. The study of EMM considering vectors of dependent variables has many applications. For example, in the context of risk sharing, EMM refers to the study of the comonotonicity (or no-sabotage condition) of a risk sharing rule (see Denuit and Robert (2022)).

Recently, based on the notion of total positivity of the joint density of the vector, Pellerey and Navarro (2022) obtained sufficient conditions for EMM when the monotonicity is in terms of the likelihood ratio order. In this work, new sufficient conditions are provided, based on the properties of the marginals and the copula. Moreover, parametric examples are considered for some of the results included in Pellerey and Navarro (2022).

12 :30 – 12 :45: Closing

This is a joined ISBA statistics and AHIDDA group seminar by

Marie CHION
MRC Biostatistics Unit, University of Cambridge, Cambridge, UK

on

"Accounting for multiple imputation in differential
quantitative proteomics"

Abstract :
Quantitative proteomics using liquid chromatography-mass spectrometry can identify and quantify several thousand proteins in a few hours of analysis. Differential quantitative proteomics analysis compares the measured intensity of proteins between different
conditions to determine those whose abundance varies "significantly". A particularity of these large datasets analysed is that they contain missing values between 5 and 15%.

One way of dealing with the problem of missing values is to impute them, i.e., to replace them with a value defined by the user or an algorithm. Thus, multiple imputation allows the imputation process to be iterated several times to obtain several complete data sets.
These are then combined before applying conventional statistical tools. However, standard software for statistical analysis of proteomics data uses the average complete dataset and ignores the uncertainty induced by the random imputation process.

Therefore, we present a rigorous method of multiple imputation using Rubin's rules and a variant of the t-moderated test that accounts for the variability arising from both the original dataset and the multiple imputation process.

Résumé:
En statistique, une valeur est dite manquante lorsqu’elle n’est pas observée. Elle peut être due à une non-réponse ou à des problèmes expérimentaux par exemple. Les techniques
courantes d’analyse statistique reposent sur des jeux de données complets. Les appliquer tels quels en présence de données manquantes peut entraîner des résultats fortement
biaisés.

Dans ce séminaire, nous définirons les valeurs manquantes ainsi que les mécanismes qui les régissent. Nous décrirons ensuite les différentes méthodes disponibles pour analyser
les jeux de données qui en contiennent. Nous aborderons notamment les techniques d’imputation. Celles-ci consistent à remplacer une valeur manquante par une valeur
définie par l’utilisateur. Enfin, nous appliquerons ces notions à un contexte biologique particulier : la protéomique quantitative.

TEAMS LINK here :  TEAMS 

Abstract :
We study the problem of regularised maximum-likelihood optimisation in ill-posed generalised linear models with covariates that include subsets that are relevant and that are irrelevant for the response. It is assumed that the source of ill-posedness is a joint low dimensionality of the response and a subset of the relevant covariates in the sense of a latent factor generalised linear model (GLM). In particular, we propose a novel iteratively-reweighted-partial-least-squares (IRPLS) algorithm and show that it is better than any other projection or penalisation based regularisation algorithm. Under regularity assumptions on the latent factor GLM we show that the convergence rate of the IRPLS estimator with high probability is the same as that of the maximum likelihood estimator in our latent factor GLM, which is an oracle achieving an optimal parametric rate. Our findings are confirmed by numerical studies.

Abstract :The Laplace approximation is a popular method for providing mean and variance estimates for a Bayesian posterior. But can we trust these estimates for practical use? One might consider using rate-of-convergence bounds for the Bayesian Central Limit Theorem (BCLT) to provide quality guarantees for the Laplace approximation. But the bounds in existing versions of the BCLT either: require knowing the true data-generating parameter, are asymptotic in the number of samples, do not control the Bayesian posterior mean, or apply only to narrow classes of models. Our work provides the first closed-form, finite-sample quality bounds for the Laplace approximation that simultaneously (1) do not require knowing the true parameter, (2) control posterior means and variances, and (3) apply generally to models that satisfy the conditions of the asymptotic BCLT. In fact, our bounds work even in the presence of misspecification. We compute exact constants in our bounds for a variety of standard models, including logistic regression, and numerically demonstrate their utility. And we provide a framework for analysis of more complex models.

This is joint work with Ryan Giordano (MIT) and Tamara Broderick (MIT).

(Stanford University)

will give a presentation on

“Statistical tools for modeling women physiology - Topic models identify sub-communities in human vaginal microbiota”

(DKV, Prof. invité UCLouvain)

will give a presentation on

“The concordance probability in all its shapes and forms”

(GSK, Prof. invité UMons et UCLouvain)

“Tolerance Intervals in Non-Clinical Biostatistics: Past, Present and Future”

Abstract :

A framework for quantifying model risks in commercial banking is proposed. Model Uncertainty is investigated from different angles with the aim to capture risks stemming from the model itself as well as its interaction with wider frameworks. As a first step we investigate model misspecification by assessing erroneous functional forms, ineffective variable selection, and calibration issues. Then the focus moves to model sensitivity to estimate the impact of either portfolio changes and external shocks. Finally, interactions among various models are scrutinized. Our final goal is to derive a distribution of indicators for summarizing the impact of model uncertainty on synthetic measures like bank’s economic, capital, liquidity ratios, and so on. Governance impacts are summarized in terms of the definition of a comprehensive model appetite framework with corresponding tolerance bands. Ex-ante assessment and continuous monitoring allow for a thorough improvement of the whole model risk management process.

Abstract :

Using two broad use cases, students will be familiarized with the "Machine Learning Use Case Assessment Canvas", a framework for scoping, aligning and understanding the full "Analytics Lifecycle" of analytical projects before they've truly begun. This will give students a taste of practical methodology from start to end of analytical projects and give them a more concrete idea of what are the most important questions to answer for any given real-life project.

Workshop also by TEAMS here : TEAMS

Abstract :

Pediatric cardiac critical care providers are often challenged with the equally important but often conflicting goals of minimizing patients’ exposure to mechanical ventilation and preventing extubation failure. Extubation failures have been associated with adverse outcomes including cardiac arrest and mortality. Reliable measures of extubation readiness, while validated in adult patients, remain elusive in pediatric cardiac critical care. Patients in the cardiac intensive care unit (CICU) have heterogeneous pathophysiology, and failure to breathe without assistance from a ventilator can be the result of primary respiratory or cardiac failure, or a mixed etiology. Physicians and nurses need prediction tools to help with clinical decision making when assessing children in the CICU for extubation readiness. 
This paper develops a prediction tool using large-scale “shallow” data from a clinical registry of over 50 institutions from North America (Pediatric Cardiac Critical Care Consortium: PC4), combined with small-scale “deep” data from CICU monitors and devices at 1 minute intervals. The latter data source allows the opportunity to study physiologic parameters during the key period when patients are evaluated for extubation readiness. We develop a tandem machine learning based approach to combine large-scale, shallow data with small-scale, deep data to improve prediction. The idea is to perform sequential classification: first using widely available covariates for risk stratification and subsequently refining prediction using deep data. Time series models are used to extract features from the deep data. We develop a novel framework that is time and cost-effective, for identifying patient subgroups that would most benefit from a second-stage prediction refinement using the deep data. Final tandem prediction is obtained by combining predictions from both the first and second stage classifiers. Our proposed method yields a classifier with improved prediction accuracy for predicting extubation failure in the CICU.

Workshop given online by TEAMS here : TEAMS

Abstract:

Whether you are a statistician, researcher, data analyst or data scientist, data visualisation and dashboard design skills are important to communicate your findings and complex information in a clear and effective way. In this workshop, we will cover the fundamentals of dashboard design and data visualisation, including topics like choosing appropriate chart types, best practices in terms of visual elements and layout, and common mistakes and pitfalls. We will review the existing visualisation tools and explore the main features and functionalities of Power BI, a widely used tool that allows you to connect to and transform data, create and customise visuals, and share dashboards with others. This workshop will also give you a clearer view of the process and steps to follow when developing a dashboard in Power BI.

Abstract :

Extending rank-based inference to a multivariate setting such as multiple-output regression or MANOVA with unspecified d-dimensional error density has remained an open problem for more than half a century. None of the many solutions proposed so far is enjoying the combination of distribution-freeness and efficiency that makes rank-based inference a successful tool in the univariate setting. A concept of center-outward multivariate ranks and signs based on measure transportation ideas has been introduced recently. Center-outward ranks and signs are not only distribution-free but achieve in dimension d > 1 the (essential) maximal ancillarity property of traditional univariate ranks. 

In the talk, we recall basic ideas behind the rank tests, we introduce the multivariate center-outward ranks and signs based on the measure transport approach and show that fully distribution-free testing procedures based on center-outward ranks can achieve parametric efficiency. We show the asymptotic normality results required in the construction of such tests in multiple-output regression and MANOVA models.

Simulations and an empirical study demonstrate the excellent performance of the proposed procedures.

will give a presentation on

Assigning Topics to Documents by Successive Projections

Abstract:

Topic models provide a useful tool to organize and understand the structure of large corpora of text documents, in particular, to discover hidden thematic structure. Clustering documents from big unstructured corpora into topics is an important task in various fields, such as image analysis, e-commerce, social networks, population genetics. Since the number of topics is typically substantially smaller than the size of the corpus and of the dictionary, the methods of topic modeling can lead to a dramatic dimension reduction. We study the problem of estimating the topic-document matrix, which gives the topics distribution for each document in a given corpus, that is we focus on the clustering aspect of the problem. We introduce an algorithm that we call Successive Projection Overlapping Clustering (SPOC) inspired by the Successive Projection Algorithm for separable matrix factorization. This algorithm is simple to implement and computationally fast. We establish upper bounds on the performance of SPOC algorithm for estimation of topic-document matrix, as well as near matching minimax lower bounds.

 

will give a presentation on

Deep Learning Meets Sparse Regularization
 

Abstract:

Deep learning has been wildly successful in practice and most state-of-the-art artificial intelligence systems are based on neural networks. Lacking, however, is a rigorous mathematical theory that adequately explains the amazing performance of deep neural networks.
In this talk, I present a new mathematical framework that provides the beginning of a deeper understanding of deep learning. This framework precisely characterizes the functional properties of trained neural networks. The key mathematical tools which support this framework include transform-domain sparse regularization, the Radon transform of computed tomography, and approximation theory. This framework explains the effect of weight decay regularization in neural network training, the importance of skip connections and low-rank weight matrices in network architectures, the role of sparsity in neural networks, and explains why neural networks can perform well in high-dimensional problems.
 

Will give a presentation on

Spatial depth for object data

Abstract:

In the first part of the talk we give a brief introduction to the concepts of statistical depth and object data. In summary, statistical depth measures assign high (low) values for points located near (far away from) the bulk of the data distribution, allowing quantifying their centrality/outlyingness. Whereas, object data refers to samples of data where the observations reside in an arbitrary metric space, instead of the space R^p.

In the second part of the talk we describe a novel measure of statistical depth, the metric spatial depth, for object data. This depth measure is shown to have highly interpretable geometric properties, making it appealing in object data analysis where standard descriptive statistics are difficult to compute. The proposed measure reduces to the classical spatial depth in a Euclidean space. In addition to studying its theoretical properties, to provide intuition on the concept, we explicitly compute metric spatial depths in several different metric spaces. Finally, we showcase the practical usefulness of the metric spatial depth in outlier detection, non-convex depth region estimation and classification.

Related preprint: https://arxiv.org/abs/2306.09740

will give a presentation on

E is the new P

Abstract:

How much evidence do the data give us about one hypothesis versus another? The standard way to measure evidence is still the p-value, despite a myriad of problems surrounding it. We present the e-value, a recently popularized notion of evidence which overcomes some of these issues. While e-values were only given a name as recently as 2019, interest in them has since exploded with papers in the Annals, JRSS B, Biometrika and the like - June 2022 saw the first international workshop on e-values, a second one is planned for May 2024.

In simple cases, e-values coincide with Bayes factors. But if the null is composite or nonparametric, or an alternative cannot be explicitly formulated, e-values and Bayes factors become distinct and e-processes can be seen as a generalization of nonnegative supermartingales. Unlike the Bayes factor, e-values always allow for tests with strict frequentist Type-I error control under optional continuation of data collection and combination of data from different sources. E-values are also the basic building blocks of anytime-valid confidence intervals that remain valid under continuous monitoring and optional stopping. In parametric settings they tend to be strictly wider than, hence consistent with Bayesian credible intervals. This led to the development of the e-posterior, an analogue to the Bayesian posterior that *gets wider rather than wrong* if the prior is chosen badly.

This work is based on:

P. Grunwald,. R. de Heide, W. Koolen (2023). Safe Testing. To appear in J. Roy. Stat. Soc., Series B

P. Grunwald (2023) . The E-Posterior. Proc. Phil. Trans. Soc. London Series A, 2023.

Bernadette Govaerts, Catherine Rasse and Christian Ritter (UCLouvain)

Note: although a teams link to this session is provided, we do not recommend attending it remotely.

Learning statistics by playing with physical objects

In an era of so-called artificial intelligence, digital twins and virtual reality, we may easily forget the power of learning by playing with physical objects. Touching things and observing real phenomena allows developing a feeling for otherwise abstract concepts such as probabilities, measurement errors and empirical relationships.

In this workshop, we will visit a wide selection of objects, games, and small experiments which we found helpful in this context.

Our workshop will be divided into three parts:

The first part, from 14h30 to 15h30, will be a seminar. We will demonstrate and analyze what we can learn from rolling dice, playing cards, dropping little balls through a maze of pins, measuring metal pieces, sending steel balls down a ramp to splash into a gutter filled with water, building and testing paper helicopters, etc.

In the second part, from 15h30 to 16h30, while enjoying a prolonged coffee break, you will have the opportunity to play and learn with these artifacts yourself. We shall set up several tables with our games and experiments and invite you to explore.

In the last part, from 16h30 to 17h, we will return to the seminar room, reflect on what we have seen, gather further ideas, and summarize the workshop.

P.S.: After the seminar, some of us might go for a beer at the Café des Halles. Feel free to join us.

-------------------------------------------------------------

https://teams.microsoft.com/l/meetup-join/19%3a181620c51efa44038c5ba68b042af908%40thread.tacv2/1695136284121?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

will give a presentation on

Using Internet Data Sources

Summary:

In this talk, I shall visit a collection of internet sources to find data about important subjects in health, wealth, environment, and life in general. This will include the Gapminder, Eurostat, Statistics Belgium, Healthdata.Org, IRCELINE and Sciensano. Whenever appropriate both direct access and access via a programmable API will be presented.

This is a version of an evolving talk given every year or two.

The presentation will take place in room C-115. For those who cannot come in person, remote participation is possible.

https://teams.microsoft.com/l/meetup-join/19%3ab569c40586524f89b8ae1b69c2dae99c%40thread.tacv2/1697573477553?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

will give a presentation on

Monetizing Data and Statistics: Strategies for Success

Summary:

Through case studies, we will explore strategies for data and statistics monetization.

First, we will delve into the strategy of consulting on other people's data, allowing you to leverage your modeling expertise for valuable insights.

Second, we will uncover the potential of monetizing or selling data, providing a pathway to unlock income through data collection and sales.

Third, we will investigate the application of statistics and data analysis, revealing how data-driven decisions can offer a distinctive competitive edge when entering a new market. These strategies provide viable monetization methods in the field of data and statistics.

Ce sera une présentation en-ligne via le lien:

https://teams.microsoft.com/l/meetup-join/19%3ab569c40586524f89b8ae1b69c2dae99c%40thread.tacv2/1697574006124?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

La présentation sera transmise au C-115.

will give a presentation on

On the topology of higher-order age-dependent random connection models

Abstract:

Preferential attachment is a popular mechanism for generating scale-free networks. While it offers a compelling narrative, the underlying reinforced processes make it difficult to rigorously establish subtle properties. Recently, age-dependent random connection models were proposed as an alternative that is capable of generating similar networks with a mechanism that is amenable to a more refined analysis. In this talk, we analyze the asymptotic behavior of higher-order topological characteristics such as higher-order degree distributions and Betti numbers in large domains. We demonstrate the practical application of the theoretical results to real-world datasets by analyzing scientific collaboration networks based on data from arXiv. This talk is based on joint work with Péter Juhász

will give a presentation on

Statistical analyses of clinical trials in the HIV field in Africa

Abstract:

The Centre for the AIDS Programme of Research – “CAPRISA” – is a research organisation in KwaZulu-Natal, South Africa, conducting high-impact globally relevant and locally responsive research on HIV, TB and SARS-CoV-2 epidemiology, pathogenesis, prevention and treatment. In this presentation, we will cover the findings and impact of HIV research conducted by CAPRISA since its founding in 2002, focusing on randomized clinical trials and other studies designed with the objective of understanding high HIV incidence in young African women. The presentation will cover the various study designs and statistical analysis approaches used in the conduct of this research, and outline some of the challenges faced in analysis as well as implementation.

Teams Link: https://teams.microsoft.com/l/meetup-join/19%3a14db0e0b685e4bbf9d8e02d4f67b84a8%40thread.tacv2/1700145961654?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

will give a presentation on

Identifying risk in clinical trials: central statistical monitoring and beyond
 

Abstract:

Clinical trials are used to assess the safety and efficacy of drugs and medical devices. In large clinical trials, hundreds of sites enroll thousands of patients across multiple countries. A clinical research site may face many challenges during study conduct, including misunderstanding of the study protocol, non-compliance with Good Clinical Practices, use of miscalibrated equipment, and failure to report data in a timely manner.

CluePoints has developed a web platform to analyze data from clinical trials. The platform uses statistical and machine learning algorithms to detect potential risks within study conduct. Risk signals are reported to study teams, allowing them to implement corrective actions and mitigate the impact on trial results. In this presentation, we will demonstrate how statistics and machine learning methods can support risk-based quality management and clinical trial oversight.

Methods range from univariate statistical tests to custom deep-learning architectures designed to interrogate clinical data.

It is not uncommon that organizers, speakers and participants go for a beer at the Café des Halles after the talk. Feel free to join us.

Teams link: https://teams.microsoft.com/l/meetup-join/19%3a14db0e0b685e4bbf9d8e02d4f67b84a8%40thread.tacv2/1700146080782?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

will give a presentation on

Separation in Networks

Network separation measures the likelihood of connections within groups compared to those between groups, operating at both the network and node levels. This concept holds significance across various disciplines such as sociology, management, and social psychology, providing insights into social inclusion, inequalities, discrimination, stereotyping, norms enforcement, and conflicts.

The separation can be assessed across different attributes, including time-invariant factors such as gender and ethnicity, as well as time-variant factors like achievement and smoking status.

This talk aims to showcase the importance of measuring separation in network analysis. It will delve into various metrics, discussing their advantages and disadvantages. Additionally, the session will provide practical insights into conducting these analyses using the R package Netseg by Michal Bojanowski. The presentation will be enriched with real-world applications, drawing examples from the SILNER and EGONET studies.

The seminar will be followed by a coffee break.

Teams Link: https://teams.microsoft.com/l/meetup-join/19%3a69bfaa3d04aa49c6a923a9ecf79bfc7c%40thread.tacv2/1701186791594?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

(avec la collaboration de Flore Vancompernolle Vromman, Sylvain Courtain, Pierre Leleux, Constantin de Schaetzen, Eve Beghein et Alexia Kneip, UCLouvain)

will give a presentation on

Le problème de la réduction des discriminations dans les prédictions de modèles de classification supervisée: une formulation maximum entropique de la régression logistique avec contraintes de "parité démographique" (demographic parity)

Abstract:

Les problèmes de discrimination et de ségrégation sont en passe de devenir des préoccupations essentielles dans nos sociétés contemporaines. D'autres part, de nombreuses décisions sont maintenant informées, voire prises, par des modèles ou algorithmes numériques dont les prédictions pourraient être, pour diverses raisons, discriminatoires. Dans ce contexte, nous allons nous intéresser à la classification supervisée et à divers mécanismes qui peuvent réduire cette discrimination (en terme de parité démographique). En particulier, nous décrirons certains de ces mécanismes et présenterons une variante de la régression logistique qui permet de facilement introduire des contraintes d'équité par rapport à une ou plusieurs variables sensibles. Ce modèle peut ensuite être enchâssé dans une procédure stepwise permettant de réduire davantage la discrimination en éliminant les variables qui ont le plus grand impact en terme de discrimination. Les comparaisons expérimentales montrent que, sur les jeux de données étudiés, certaines méthodes réussissent à augmenter de manière significative la parité démographique sans observer d'impact important sur la précision du modèle.

It is not uncommon that organizers, speakers and participants go for a beer at the Café des Halles after the talk. Feel free to join us.

Teams link: https://teams.microsoft.com/l/meetup-join/19%3a69bfaa3d04aa49c6a923a9ecf79bfc7c%40thread.tacv2/1701187136409?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

will give a presentation on

Statistical foundations of topological data analysis

Dates :

• Wednesday December 13, 2023, 10h45-12h45 and 14h-16h ISBA, Room c115
• Thursday December 14, 2023, 10h45-12h45 and 14h-16h ISBA, Room c115

-------------------------------------------------------------------

Extended abstract:

Topological data analysis (TDA) is an emerging field at the interface between algebraic topology and data science. The philosophy behind TDA is to leverage invariants from algebraic topology to gain insights into data sets. While initially, TDA was developed and promoted by mathematicians, it is now applied in a variety of disciplines such as biology, chemistry, and materials science. The key tool in TDA is the persistence diagram, which captures the appearance and disappearance of topological features at multiple scales. The goal of this course is to explain how the tools provided by TDA can be combined with rigorous statistical methods to enable a scientific analysis of complex geometric data.

Lecture 1: The persistence diagram. We discuss the topological foundations of the persistence diagram and discuss its computability. We also present practical use cases with the GUDHI package.

Lecture 2: Distance-to-measure. While the persistence diagram can unearth refined geometric structures, it can be sensitive to the presence of outliers. We discuss the distance-to-measure as one way to make the persistence diagram more robust.

Lecture 3: Point processes. To carry out formal hypothesis testing, it is essential to work with well-defined null hypotheses. We introduce point processes as a flexible model for point clouds.

Lecture 4: Persistence-based summary statistics. We combine the methods from the first three lectures to define statistical tests based on the persistence diagram. These include simulation-based tests that are feasible for moderately large data as well as asymptotic approaches that become exact for large windows.

Registration (free for university members and students): please fill in the following form

will give a presentation on

Statistical foundations of topological data analysis

Dates :

• Wednesday December 13, 2023, 10h45-12h45 and 14h-16h ISBA, Room c115
• Thursday December 14, 2023, 10h45-12h45 and 14h-16h ISBA, Room c115

-------------------------------------------------------------------

Extended abstract:

Topological data analysis (TDA) is an emerging field at the interface between algebraic topology and data science. The philosophy behind TDA is to leverage invariants from algebraic topology to gain insights into data sets. While initially, TDA was developed and promoted by mathematicians, it is now applied in a variety of disciplines such as biology, chemistry, and materials science. The key tool in TDA is the persistence diagram, which captures the appearance and disappearance of topological features at multiple scales. The goal of this course is to explain how the tools provided by TDA can be combined with rigorous statistical methods to enable a scientific analysis of complex geometric data.

Lecture 1: The persistence diagram. We discuss the topological foundations of the persistence diagram and discuss its computability. We also present practical use cases with the GUDHI package.

Lecture 2: Distance-to-measure. While the persistence diagram can unearth refined geometric structures, it can be sensitive to the presence of outliers. We discuss the distance-to-measure as one way to make the persistence diagram more robust.

Lecture 3: Point processes. To carry out formal hypothesis testing, it is essential to work with well-defined null hypotheses. We introduce point processes as a flexible model for point clouds.

Lecture 4: Persistence-based summary statistics. We combine the methods from the first three lectures to define statistical tests based on the persistence diagram. These include simulation-based tests that are feasible for moderately large data as well as asymptotic approaches that become exact for large windows.

Registration (free for university members and students): please fill in the following form

will give a presentation on

Statistically Enhanced Learning for Sports Analytics and Injury Prevention

Abstract:

In this talk, I will describe how one can use probability distributions to model the outcomes of football matches, and how this can be combined with machine learning procedures to predict big tournaments and hereby even outperform bookmakers.
The idea is to enhance machine learning by the creation of information-rich covariates which are statistical estimates. I will further illustrate this approach on handball matches and explain how our approach can serve coaches in their game preparation. I shall conclude with an outlook on how these findings can be translated to sports medicine and, in particular, the estimation of injury risks.

TEAMS: https://teams.microsoft.com/l/meetup-join/19%3a40064c3c302a4341aa9f3056598da798%40thread.tacv2/1705388562254?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

will give a presentation on

Beyond Numbers: An Introduction to the Role of Data Science inside Teams and Federations

This presentation will delve into the current role of data science in practical applications in sports and its future perspectives. I will share insights based on my experiences in the realms of data and sports, having recently completed a Master's degree at ENSAI in France. Two case studies will be showcased: The first case centers around Montpellier Hérault Sport Club, a professional soccer club in France, where I undertook an internship between two seasons. This segment will focus on the available data in elite football, its utilization, and its specific applications within the club. It will discuss the use of event data to assist the coaching staff, the structural aspects, and the various ways data contributes to a professional team, with a particular emphasis on tactical data. The second case takes place at Tennis Australia, where I was a part of the Game Insight Group, a team of researchers and I was working on the 'Win Predictor' project. This project involves predicting outcomes in ATP tennis matches, with a specific goal of providing accurate probabilities for Tennis Australia's broadcasts, such as the Australian Open. This section will explore the feasibility of game prediction, the associated challenges, and how mathematics and modeling can be applied to sports predictions, with its limits.

TEAMS: https://teams.microsoft.com/l/meetup-join/19%3a40064c3c302a4341aa9f3056598da798%40thread.tacv2/1705389020794?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

09:05 - 09:30 : Morine DELHELLE
"Copula based dependent censoring in cure models"

Abstract :
In survival data analysis datasets with both a cure fraction (individuals who will never experience the event of interest) and dependent censoring (loss of follow-up for a reason linked to the event of interest before the occurence of this event) are not scarce and it is important to use an adequate model dealing with these two characteristics if we want to avoid bias in parameters estima-tions or false conclusions in clinical trials. In this presentation I will propose a fully parametric survival mixture cure model that takes possible dependent censoring into account which is based on an unknown copula that describes the relation between the survival and censoring times. So, the advantages of the model are that dependent censoring and the cure fraction are both consi-dered and that the copula is not assumed to be known. Moreover, it allows us to estimate the strength of dependence. The situations with and without covariates will be discussed. In survival data analysis datasets with both a cure fraction (individuals who will never experience the event of interest) and de-pendent censoring (loss of follow-up for a reason linked to the event of interest before the occurence of this event) are not scarce and it is important to use an adequate model dealing with these two characteristics if we want to avoid bias in parameters estimations or false conclusions in clinical trials. In this pre-sentation I will propose a fully parametric survival mixture cure model that takes possible dependent censoring into account which is based on an unknown copula that describes the relation between the survival and censoring times. So, the advantages of the model are that dependent censoring and the cure fraction are both considered and that the copula is not assumed to be known. Moreover, it allows us to estimate the strength of dependence. The situations with and without covariates will be discussed.

09:30 - 09:55 : Jean-loup DUPRET
"A Subdiffusive Stochastic Volatility Jump Model"

Abstract :
Subdiffusions appear as good candidates for modelling illiquidity in financial markets. Existing subdiffusive models of asset prices are indeed able to capture the motionless periods in the quotes of thinly-traded assets. However, they fail at reproducing simultaneously the jumps and the time-varying random vola-tility observed in the price of these assets. The aim of this work is hence to propose a new model of subdiffusive asset prices reproducing the main charac-teristics exhibited in illiquid markets. This is done by considering a stochastic volatility jump model, time-changed by an inverse subordinator. We derive the forward fractional partial differential equations (PDE) governing the probabi-lity density function of the introduced model and we prove that it leads to an arbitrage-free and incomplete market under a suitable change of measure. By proposing a new procedure for estimating the model parameters and using a series expansion for solving numerically the obtained fractional PDE, we are able to price various European-type derivatives on illiquid assets and to de-part from the common Markovian valuation setup. This way, we show that the introduced subdiffusive stochastic volatility jump model yields consistent and reliable results in illiquid markets.

09:55 - 10:20 : Hugues ANNOYE
"Generating administrative data bases with Wasserstein GAN"

Abstract :
In a world increasingly surrounded by data, data privacy and anonymisation are becoming more and more important. Under these circumstances, the need for fake data bases that replicate the characteristics of the population while preserving privacy is arising. In this presentation, we investigate how we can use Wasserstein generative adversarial network (WGAN), developed by [ACB17] in the context of image synthesis, to create administrative data bases and we also adapt it to take weights into account. Administrative data have the spe-cificity of mixing continuous and categorical data, which should be taken into account in the architecture of the WGANs. Then, we present a new method to evaluate the results based on Support Vector Data Description (SVDD) on real data coming from Labour Force Survey (LFS).


10:20 - 10:45 : Coffee break 


10:45 - 11:10 : Lise LEONARD
"High-dimensional regression : Model averaging and inference"

Abstract :
With the advent of technology and the proliferation of data collection me-thods, researchers now have access to vast amounts of data. High-dimensional regression models are designed to handle datasets with more predictors than observations, allowing researchers to leverage the wealth of information avai-lable in these high-dimensional datasets. However, these high-dimensional me-thods, such as the Lasso which is the most used in this context, depend on unknown tuning parameters. The goal of our proposal is to eliminate that difficult choice of the tuning parameter and obtain an estimator that allows inference. The main feature of our procedure is to pool together information from multiple estimators to obtain one single, final estimator. We propose a strategy to aggregate these regression coefficients to reduce the prediction risk of the estimation and to eliminate the tuning parameter. Theoretical results on the distribution and the prediction risk of the method are presented. In particular, we show the normality of the estimator even after the aggregation, which allows for inference for high-dimensional models. The performance of the method is illustrated by numerical simulations and an application on real data.

11:10 -  11:35 : Hortense DOMS
"Bayesian joint model for longitudinal HR-QOL and time-to-event outcomes"

Abstract :
In cancer clinical trials, the traditional endpoints are overall survival (OS) and progression-free survival (PFS). Recently, the use of health-related quality of life (HRQoL) as a major endpoint has become increasingly common. HRQoL data are measured using self-administered questionnaires collected longitudi-nally throughout the follow-up period and are used to assess changes in pa-tients’ perceptions of their physical and mental health over time. To analyse this data, the standard approach in oncology is to calculate a score for each patient at each time point and apply a linear mixed model (LMM) to the patient’s score. However, recent research has shown that using the LMM to analyse HRQoL is not appropriate. A more suitable methodology, known as item response theory (IRT), is therefore emerging. IRT models link patient responses to a latent variable representing the HRQoL dimension studied. Ad-ditionally, missing data in HRQoL questionnaires may result from dropouts due to clinical events. As this form of dropout can be informative, it is essen-tial to take it into account when analysing longitudinal outcomes in order to obtain valid results. In this presentation, we explain how to extend the joint model to the HRQoL longitudinal data framework. A multilevel item response theory model is used for longitudinal data and a proportional cause-specific hazards model is used for survival data. Inference is performed in a Bayesian framework using the Markov chain Monte Carlo algorithm and we apply the proposed model to data from patients with first progression of glioblastoma.

11:35 - 12:00 : Aigerim ZHUMAN
"Speeding up Monte Carlo Integration : Control Neighbors for Optimal Conver-gence"

Abstract :
The method of control variates is a powerful technique that allows to reduce the variance of the Monte Carlo estimate of a multivariate integral by introdu-cing auxiliary functions with known expectations, called control variates. We propose to use nearest neighbor estimates as control variates in order to speed up the convergence rate of the Monte Carlo integration procedure. Our novel estimate, called the Control Neighbor estimate, achieves the optimal conver-gence rate for Lipschitz functions. In addition, a non-asymptotic bound on the probabilistic error of the procedure is obtained via an extension of McDiarmi-d’s inequality for functions with bounded differences on a high probability set. Moreover, several numerical experiments confirm the good performance of the proposed estimate.

Abstract :
The local covariance measure (LCM) quantifies deviations from conditional local independence (CLI) between two time-continuous stochastic processes. It is a function of time defined as the expectation of a suitably defined stochastic integral. Under CLI this integral is a martingale and LCM is constantly 0. The talk will focus on applications in survival analysis where one process is the indicator of death. In this case LCM can be used to test the hypothesis that survival time is conditionally independent of a baseline covariate even in the presence of censoring. More importantly, LCM can be used to define an assumption-lean additive hazard target parameter. The talk will cover aspects of implementation based on machine learning (ML) as well as asymptotic theory in the spirit of double/debiased ML and using cross-fitting techniques.

SEMINAR by Thomas Nagler (LMU Munich) on "An new bootstrap for time series"

Abstract :
Resampling methods such as the bootstrap have proven invaluable in statistics and machine learning. However, the applicability of traditional bootstrap methods is limited when dealing with large streams of dependent data, such as time series or spatially correlated observations. In this paper, we propose a novel bootstrap method that is designed to account for data dependencies and can be executed online, making it particularly suitable for real-time applications. This method is based on an autoregressive sequence of increasingly dependent resampling weights. We prove the theoretical validity of the proposed bootstrap scheme under general conditions. We demonstrate the effectiveness of our approach through extensive simulations and show that it provides reliable uncertainty quantification even in the presence of complex data dependencies. Further extensions to nonstationary time series will be discussed.

(Stanford University)

Conformal Prediction with Conditional Guarantees

Abstract :
We consider the problem of constructing distribution-free prediction sets with finite-sample conditional guarantees. Prior work has shown that it is impossible to provide exact conditional coverage universally in finite samples. Thus, most popular methods only provide marginal coverage over the covariates. We bridge this gap by defining a spectrum of problems that interpolate between marginal and conditional validity. After the reformulation of conditional coverage as coverage over a class of covariate shifts, we show how to simultaneously obtain exact finite sample coverage over all possible shifts when the target class of shifts is finite dimensional. For example, our algorithm outputs intervals with exact coverage over each group from a collection of protected groups. For more flexible, infinite dimensional classes where exact coverage is impossible, we provide a simple procedure for quantifying the gap between the coverage of our algorithm and the target level. Moreover, by tuning a single hyperparameter, we allow the practitioner to control the size of this gap across shifts of interest. Our methods can be easily incorporated into existing split conformal inference pipelines, and thus can be used to quantify the uncertainty of modern black-box algorithms without distributional assumptions. This is joint work with Isaac Gibbs (Stanford) and Emmanuel Candes (Stanford).

(Stanford University)

Election Modeling in 2024: A Conformal Inference Approach

Abstract :
We consider a high-stakes application of statistical inference: uncertainty quantification for election night modeling. In this problem, the analyst observes vote counts from early-reporting jurisdictions, e.g., precincts on the East Coast of the United States, and fits a model to these results that predicts the final outcome in each contested race. Quantifying the error of this prediction is crucial; an overconfident prediction can mislead the public and harm the news provider’s reputation. Over the last four years, we have worked on methods to extend conformal prediction, a popular method for assumption-lean inference, to this setting. Working with election data poses many challenges. For example, the data-generating distribution shifts over time, and spatiotemporal correlation can invalidate standard approaches. Variants of our model have been featured in The Washington Post’s coverage of the 2020 and 2022 national United States elections, and the model introduced in this talk will be used in this fall’s presidential election. This is joint work with Lenny Bronner (The Washington Post) and Emmanuel Candes (Stanford).

Link to TEAMS

Simone Padoan

(Bocconi University)

Abstract : ...

Guy Nason

(Imperial College, London)

invited by Rainer von Sachs

will give a presentation on :

Network Time Series

Abstract :

A network time series is a multivariate time series where the individual series are known to be linked by some underlying network structure. Sometimes this network is known a priori, and sometimes the network has to be created, often inferred from the multivariate series itself. Network time series are becoming increasingly common, long, and collected over a large number of variables. We are often interested in network time series where the network changes over time.

We describe some recent developments in the modeling of network time series via generalized network autoregressive (GNAR) process models. These models use regular autoregressive links between a variable and its past and between a variable and the past of its neighbours. GNAR models are highly parsimonious and, hence, work well for short series or those afflicted by worrying amounts of missing data. For the same reason, they tend not to overfit and often exhibit excellent forecasting performance, especially when compared to alternatives such as vector autoregressive models.

This talk explains the GNAR model and some interesting variants. We introduce some new tools for model selection and exhibit their use on data from different applied domains.

 Guy Nason

Abstract :
In this workshop we take a deep dive into missing values from the standpoint of imputation. In the first part, I start by properly defining the problem of missingness and introducing a view on missing values that is particularly suited for imputation. Focussing on the missing at random (MAR) case, we discuss classical results of ignorability and what a good imputation method needs to bring to the table. Within this discussion, I also introduce some of the most promising imputation methods currently available. Finally, we take a brief look at state-of-the-art research on how to evaluate an imputation method for a given dataset, based on the concept of imputation scores (I-Scores). Throughout I provide examples to illustrate the concepts. If time allows, I will also discuss problems and approaches under the more difficult scenario of missing not at random (MNAR) missingness.
In the second part, we study the concepts introduced in the first part on tangible R examples.

This seminar will be accessible online via Teams

(UCLouvain)

will give a presentation on :

Expensive but Worth it: Live Projects in Statistics, Data Science, and Analytics Courses

Abstract :

Students in statistics, data science, analytics, and related fields study the theory and methodology of data-related topics. Some, but not all, are exposed to experiential learning courses that cover essential parts of the life cycle of practical problem-solving. Experiential learning enables students to convert real-world issues into solvable technical questions and effectively communicate their findings to clients.
We describe several experiential learning course designs in statistics, data science, and analytics curricula. We present findings from interviews with faculty from the U.S., Europe, and the Middle East and surveys of former students. We observe that courses featuring live projects and coaching by experienced faculty have a high career impact, as reported by former participants.  However, such courses are labor-intensive for both instructors and students.  
We give estimates of the required effort to deliver courses with live projects and the perceived benefits and trade-offs of such courses. Overall, we conclude that courses offering live-project experiences, despite being more time-consuming than traditional courses, offer significant benefits for students regarding career impact and skill development, making them worthwhile investments.

Exceptionally, this seminar will be accessible online via Teams

(RTBF  - journaliste pour Décrypte, la cellule de datajournalisme et de fact-checking de la RTBF)

Enhancing Data Literacy in Journalism – What we learned with the Décrypte Team at RTBF

Abstract :
Are journalists proficient data analysts? This workshop delves into the critical examination of journalists' roles in an era where "data journalism" is not just a buzzword but a necessity. The Covid-19 pandemic underscored the essential need for data literacy, revealing that journalistic expertise does not always align with analytical proficiency, especially among those with traditional literary backgrounds. As our societies face an overwhelming influx of information, audiences are increasingly seeking not just news but clarity and understanding, pressing journalists to transcend their conventional roles and embrace data-driven storytelling.

In the first part we will discuss the operational methodologies of the Décrypte team, the data and fact-checking unit at RTBF. This exploration will shed light on the critical decision-making processes behind topic selection, defining a journalistic angle, and identifying the pivotal stages where the confluence of journalism and data analytics yields the most significant benefits. Attendees will gain an understanding of how these disciplines, when synergized, enhance the clarity, reliability, and impact of news stories.

In a second part, I will showcase a few exemplary cases of data journalism executed by the Décrypte team. Special emphasis will be placed on the "scrollytelling" format, a compelling blend of scrolling and storytelling designed to elucidate complex topics through a gradual, educational framework. Participants will be treated to a series of case studies that highlight the transformative power of data in storytelling, offering a blueprint for merging analytical depth with narrative engagement.

Link to TEAMS

• Programme
• Inscription

Vendredi 17 mai 2024
"DÉPENDANCE/AUTONOMIE: QUELLES COUVERTURES POUR QUELS PILIERS?"

- Karel Van den Bosch (Bureau fédéral du Plan)
“La protection sociale en matière de soins de longue durée en Belgique”

- Sophie Lebichot (Ethias)
- Dave Tombeux (Ethias)

La présentation aura lieu en français et en anglais

Efrat Muller

will give a presentation on :

Integration of Multi-omics Data

Abstract:

The human gut microbiome, and its metabolic activity in particular, have been implicated in a wide range of disease states, including metabolic disorders, inflammatory bowel diseases, and colorectal cancer. This growing appreciation for the impact of the gut microbiome’s metabolism on human health has given rise to studies that generate both microbiome and metabolome high-throughput data from human gut microbiome samples. Truly integrated analysis of both omic datasets, however, remains a challenging task.
In this talk, I will introduce the domain of human microbiome research, present the concept of multi-omic studies (e.g., combining metagenomics with metabolomics), and briefly discuss the common practices and open challenges in multi-omic data integration. I'll then present my latest research project, in which we developed a framework for identifying “multi-omic modules” that capture both cross-omic associations and associations with disease simultaneously, based on extensions of Canonical Correlation Analysis.

Teams link

Mathilde Papillon

will give a presentation on :

Topological Learning

Abstract:

The natural world is full of complex systems characterized by intricate relations between their components: from social interactions between individuals in a social network to electrostatic interactions between atoms in a protein. Topological Deep Learning (TDL) provides a comprehensive framework to process and extract knowledge from data associated with these systems. TDL has demonstrated theoretical and practical advantages that hold the promise of breaking ground in the applied sciences and beyond. However, the rapid growth of the TDL literature for relational systems has also led to a lack of unification in notation and language across message-passing architectures. This presents a real obstacle to building upon existing works and deploying TNNs to new real-world problems. In this talk, I provide a turn-key, unified mathematical and graphical notation for TDL and apply it to the existing body of literature. I will shed light on current open questions in the field and exciting opportunities for future development.

Teams link

Didier Schwab

(Université Grenoble Alpes)

invited by Eugen Pircalabelu

will give a presentation on :

De FlauBERT à Pantagruel : Recherches et Questions sur les Modèles de Langue Écrits, Oraux, Pictographiques et Multimodaux en Français, leur Évaluation et leurs Biais

Abstract :

Depuis l’apparition des modèles de langues contextuels tels que ELMO (Peters et al., 2018) ou BERT (Devlin et al., 2019), les représentations informatisées de textes sont devenues notablement plus précises et permettent ainsi de mieux réaliser différentes tâches typiques du traitement automatique du langage. Cette révolution s’est poursuivie avec l’apparition des larges modèles de langues, associés à la génération du langage (tels que GPT, Llama, Gemini, etc.).

Cette présentation vise à faire le point sur les différentes recherches menées depuis 2019 sur les grands modèles de langue en français. Nous y aborderons les modèles écrits, oraux, multimodaux, ainsi que ceux appliqués à des domaines spécifiques (juridique, biomédical). Nous examinerons les modèles de langue masqués (de type BERT, pour la représentation du sens) et les modèles de langue causaux (type GPT, pour la génération de texte), leurs différences, ainsi que leurs complémentarités.

 Didier Schwab

Program:

09:00 – 09:05 : Opening

Abstracts for YRD September 20, 2024:

09:05 - 09:25 : Anas Mourahib
Title: Statistics of Extremes

Abstract:
On September 7, 2024, a flood tragically claimed the lives of at least 18 people in southern Morocco. To prevent such disasters in the future, we aim to construct a dike that is 1 meter higher than the ``once per 10000 year'' sea level. The challenge is that we have less than 100 years of data, which is similar to estimating the likelihood of an event that has not yet occurred. This is where Extreme Value Theory (EVT), a branch of statistics developed since the 1970s, comes into play. In this presentation, we will introduce two well-established main methods in EVT: the Block Maxima approach and the Threshold Exceedances approach. A brief application will follow to demonstrate how these methods can help in designing the dike.

09:25 - 09:50 : Stephane Lhaut
Title: Testing parametric models for the angular measure for bivariate extremes

Abstract:
The angular measure on the unit sphere characterizes the first-order dependence structure of the components of a random vector in extreme regions and is defined in terms of standardized margins. Its statistical recovery is an important step in learning problems involving observations far away from the center. In this paper, we consider the goodness-of- fit problem which consists of testing the adequacy of the extremal dependence structure of a bivariate random sample to a given parametric model. The proposed test statistic consists of a weighted L1-Wasserstein distance between a purely non-parametric, rank-based, estimator of the true angular measure obtained by maximizing a Euclidean likelihood, and an estimated version of the angular measure under the postulated parametric model. The asymptotic distribution of the test statistic is derived and used to obtain critical values for the proposed testing procedure via a parametric bootstrap. Consistency of the bootstrap algorithm is proved. A simulation study illustrates the finite-sample performance of the test for two popular models: the logistic and the Hüsler-Reiss models.

09:50 - 10:15 : Benjamin Deketelaere
Title: Quantile Regression with a Censored Covariate

Abstract:
In the last few years, there has been growing interest in regression with censored covariates. This is of interest e.g. when studying the symptom trajectory of a neurodegenerative disease. We are interested in studying quantile regression models with covariates that are either left, right or interval censored.  Quantile regression offers many advantages over more classical mean regression, like e.g. its ability to study the whole distribution as opposed to the center of the distribution. For instance, in a study on the factors that influence high blood pressure, quantile regression allows to focus on the factors that are important for individuals in the upper tail of the blood pressure distribution. We propose a linear quantile regression model, and propose a two-stage estimation procedure of the regression coefficients, in which both steps are based on maximum likelihood estimation. The first step consists in modeling the distribution of the censored covariate given the other covariates, whereas in the second step the quantile regression coefficients are estimated. To do this, we propose to use families of enriched exponential and enriched Laplace distributions, respectively, both of which use Laguerre polynomial expansions to make the families sufficiently rich and flexible. We investigate the finite sample performance of the proposed method by means of extensive simulations.  The developed methodology is also used to study the National Health and Nutrition Examination Survey data on the factors influencing high blood pressure.

10:15 - 10:40 : Coffee break 

10:40 - 11:15 : Patricia Ortega-Jiménez
Title: Comparisons of VaR and CoVaR in terms of the value of the conditional variable.

Abstract:
Let us consider a random vector (X, Y ). Given a risk level v ∈ [0, 1], the most extended risk measure is the Value at Risk, V aRv(Y ) = F−1(v), which represents the maximum expected loss. However, the V aRv (Y ) measures the risk of the single institution without accounting for interactions with other risks. A dependence-adjusted version of the Value at Risk is the Co- Value at risk, CoV aRv,u(Y |X), which stands for V aRv(Y |X = V aRu(X)) for the risk levels v ∈ [0,1] and u ∈ [0,1]. Our goal is to find the values of the institution X that lead to the CoV aR being greater than the V aR of Y , as relying solely on the V aR may not be sufficient to face financial losses. We compare these two measures in terms of the risk-level of the conditional variable, u. For v ∈ (0, 1), under regularity conditions and a positive dependence structure, there exists a unique cut point uv such that CoV aRv,uv (Y |X) ≥ V aRv(Y ) if and only if u ≥ uv. We will see that this value uv only depends on the dependence structure of the vector. In addition we will discuss sufficient conditions and implications of the existence of an upper bound u∗ ∈ (0, 1) such that.  uv ≤u∗ forall v∈(0,1).
Several examples of copulas with bounded and unbounded cut points are analyzed and a non-parametric estimator is provided. The presented results are mainly based on the recent paper: Ortega-Jiménez, P., Pellerey, F., Sordo, M. A., and Suárez-Llorens, A. (2024). Probability equivalent level for CoVaR and VaR. Insurance: Mathematics and Economics, 115, 22-35.

11:15 - 11:40 : Oussama Belhouari
Title: The Three-step method in a dynamic setting

Abstract:
A crucial issue in a dynamic framework, is how risk valuations at different times are interrelated. In this regard, the notion of time consistency was widely introduced and discussed in the literature. A time-consistent dynamic valuation is a pricing method according to which a product that will be, in almost all states of nature, cheaper than another one at a future date should already be cheaper today. Common actuarial premium principles are not time consistent. To this end, we link the latter with an iterated property. This paper aims at constructing a time-consistent, dynamic version of the Three-step method introduced in [Deelstra and Hieber, 2020], employing a backward iteration scheme. The backward scheme is exemplified in a dual-iteration approach using a classical application, specifically a Pure Endowment. Furthermore, we explore the continuous-time limit of the backward scheme, incorporating profit-sharing into the Pure Endowment to investigate a hybrid life payoff. Our analysis reveals that, due to time consistency, the price of the actuarial component in the Three-step method undergoes a substantial increase. To address this, and in accordance with [Devolder and Lebègue, 2016], we present a reduced time-consistent variant by decreasing the safety loads in each iterative step of the backward scheme.

11:40 - 12:15 : Luc Boone
Title: Application of inverse probability of censoring weighting (IPCW) in open-label cancer clinical trials with centrally reviewed endpoints: Illustrating and extending the methodology

Abstract:
Blinding of investigators (and patients) in randomized controlled trials is not always feasible. For subjective endpoints or outcome measures, this is especially problematic considering that assessment bias cannot be ruled out. In oncology, the endpoint progression-free survival (PFS) is commonly used and is defined as the time from randomization until tumor progression or death, whichever occurs first.  
Tumor progression assessment is mostly done through radiographic images, which is prone to assessment bias in open-label (non-blinded) clinical trials.
Blinded independent central review (BICR) of endpoints such as progression-free survival is carried out in open-label cancer trials to mitigate assessment bias of unblinded local investigators. In many cases, BICR takes places retrospectively. Patients for whom progressive disease was not confirmed by BICR are commonly censored at the last time of (local investigator) assessment, in the absence of follow-up radiographic assessment. The censoring for such unconfirmed progressions is generally assumed to be independent when estimating survival curves and hazard ratios, which is questionable. Rather, the censoring in this case might be dependent and violate the assumption of independent censoring, and bias the commonly used estimators.
The goals of the research project are to: First, illustrate and study the application of inverse probability of censoring weighting (IPCW) to adjust for dependent censoring (Robins, 1993) in the context of open-label cancer clinical trials with centrally reviewed subjective endpoints; Second, propose and study the extension of the methodology of IPCW in the described setting through the use of joint models for longitudinal and time-to-event data; And lastly, offer recommendations to applied clinical trial statisticians on when and how to implement the illustrated and proposed methods.
-    Robins JM. Information recovery and bias adjustment in proportional hazards regression analysis of randomized trials using surrogate markers. Proceedings of the Biopharmaceutical Section, American Statistical Association 1993; 24–33.

14:30 : Philippe Naveau (C.115)

Laboratoire des Sciences du Climat et de l'Environnement (LSCE)
Invited by Eugene PIRCALABELU

will give a presentation on :

Statistical modelling of records in a changing climate

Abstract:

The increase of  recent climate record frequencies raises many statistical and climatological questions. Statistically,  the literature on modelling  record changes in a non-stationary context is sparse and, in this context,  I will present different case studies that blind multivariate extreme value theory and counterfactual theory.  Another aspect will be to determine how this non-stationary allows to improve the forecasting of record frequencies. If time allowed, the choice of climatological explanatory variables (co-variates) to understand unprecedented heatwaves will be touched upon. All statistical techniques will be illustrated by cases studies based either on climate model outputs or weather stations recordings. Concerning the temporal periods, these examples will treat past, present and future yearly time scales.

Sébastien Laurent

IAE Aix-Marseille

Asymptotics for penalized QMLEs of time series regressions

We examine a linear regression model applied to the components of a time series, aiming to identify time-varying, constant as well as zero conditional beta coefficients. To address the non-identifiability of parameters when a conditional beta is constant, we employ a lasso-type estimator. This penalized estimator simplifies the model by shrinking the estimates when the beta is constant. Given that the model accommodates conditional heteroskedasticity and the relevant regressors are unknown, the total number of parameters to estimate can be quite large. To manage this complexity, we propose a multistep estimator that first captures the dynamics of the regressors before estimating the dynamics of the betas. This strategy breaks down a high-dimensional optimization problem into several lower-dimensional ones. Since we avoid making strict parametric assumptions about the innovation distributions, we use Quasi-Maximum Likelihood (QML) estimators. The non-Markovian nature of the global model means that standard convex optimization results cannot be applied. Nevertheless, we analyze the asymptotic distribution of the multistep lasso estimator and its adaptive version, deriving bounds on the maximum value of the penalty term. We also propose a nonlinear coordinate-wise descent algorithm, which is demonstrated to find stationary points of the objective function. The finite-sample properties of these estimators are further explored through a Monte Carlo simulation and illustrated with an application to financial data.

The latest forecasting innovations, showing how data can be transformed into valuable predictive insights

Note: This talk will be given in presence in room C-115.  Online access is possible but the quality cannot be guaranteed.

Speaker:
Thomas Christ, Chief data scientist at the company prognostica, Würzburg, Germany.

Summary:
In today’s fast-paced world, data-driven forecasting is crucial for staying competitive. This session explores the latest forecasting innovations, showing how data can be transformed into valuable predictive insights.

14h30 Part 1
We’ll begin with the basics: How do advanced methods—from traditional statistics to machine learning and foundation models—reveal hidden patterns in historical data? Using a demand forecasting example, we’ll discuss building accurate, actionable models and addressing practical challenges, like handling intermittent demand, managing data hierarchies, and integrating external sources. Key success factors such as model explainability, user engagement, and adaptability will also be covered to ensure forecasts are both precise and trusted by decision-makers.

15h30 Coffee break

16h00 Part 2
The session’s second part will bring theory to life with interactive examples, like predicting solar flare activity or tracking urban bike traffic. Join us to learn how to turn forecasting theory into practical insights, gaining tools and strategies to stay ahead in this rapidly evolving field.

At 17h30, we'll probably go to have a beer at the Crêperie Bretonne. Feel free to join us.
 

  TEAMS 

Thomas Nagler

LMU Munich

The surprising effect of reshuffling the data during hyperparameter tuning

Vlad Stefan Barbu

Laboratory of Mathematics Raphaël Salem, University of Rouen – Normandy, France

Hypothesis testing based on weighted divergences

Abstract :
In this presentation, we first focus on a class of hypothesis tests for assessing the goodness-of-fit of a distribution, as well as on a class of homogeneity tests between two samples, in an iid framework. These tests are built on specific divergence measures, called weighted divergences, that allow us to focus more on a specific part of the distribution's support, while retaining also some information about the rest of the support. This methodology yields tests that are often more powerful than traditional tests, with comparable Type I error rates. We also present the associated asymptotic theory, along with corresponding Monte Carlo simulations, to examine the performance of the proposed tests. Additionally, we provide some elements for calculating test power. Extensions to Markovian and semi-Markovian cases will also be discussed.

Quentin Guibert

Maître de conférence à Paris-Dauphine (CEREMADE)

An extrapolation of temperature effects based on time series data in France

Speakers:
Profs Johan Segers (KUL) and Anna Kiriliouk (UNamur)

Dates:
Thursday October 24, 2024, 9h30-12h30 and 13h30-15h (Room C.115)
Friday October 25, 2024, 9h30-12h30 (Room C.045) and 13h30-15h (Room C.115)

Where:
ISBA, UCLouvain, 20 Voie du Roman Pays, Louvain-la-Neuve

Schedule (for both days):

9h30-11h Theory
11h-11h30 Coffee break
11h30-12h30 Theory
12h30-13h30 Lunch break (organized in the cafeteria of ISBA)
13h30-15h Exercises in R

Table of contents:
 
Day 1 – Univariate extremes

• Block maxima and the GEV
• Threshold exceedances and the GPD
• Temporal dependence and non-stationarity

 
Day 2 – Multivariate and spatial extremes

• Max-stable processes
• Multivariate threshold exceedances
• Clustering and dimension reduction

Registration before October 10, 2024 (for free for members of EDT, but mandatory) via:
URL GOOGLE FORM

Speakers:
Profs Johan Segers (KUL) and Anna Kiriliouk (UNamur)

Dates:
Thursday October 24, 2024, 9h30-12h30 and 13h30-15h (Room C.115)
Friday October 25, 2024, 9h30-12h30 (Room C.045) and 13h30-15h (Room C.115)

Where:
ISBA, UCLouvain, 20 Voie du Roman Pays, Louvain-la-Neuve

Schedule (for both days):

9h30-11h Theory
11h-11h30 Coffee break
11h30-12h30 Theory
12h30-13h30 Lunch break (organized in the cafeteria of ISBA)
13h30-15h Exercises in R

Table of contents:
 
Day 1 – Univariate extremes

• Block maxima and the GEV
• Threshold exceedances and the GPD
• Temporal dependence and non-stationarity

 
Day 2 – Multivariate and spatial extremes

• Max-stable processes
• Multivariate threshold exceedances
• Clustering and dimension reduction

Registration before October 10, 2024 (for free for members of EDT, but mandatory) via:
URL GOOGLE FORM

Institut de Duve, UCLouvain
Invited by Laura Symul

will give a presentation on :

Batch effect detection and visual quality control with CytoMDS, a Bioconductor package for low dimensional representation of distances between cytometry samples

Abstract:

Quality Control (QC) of samples is an essential preliminary step in cytometry data analysis. Notably, identification of potential batch effects and sample outliers is paramount, to avoid mistaking these effects for true biological signal in downstream analyses. However, this task can prove to be delicate and tedious, especially for datasets with many samples. Here, we present CytoMDS, a Bioconductor package implementing a dedicated method for low dimensional representation of cytometry samples composed of marker expressions for up to millions of single cells. This method combines Earth Mover’s Distance (EMD) [1] for assessing dissimilarities between multidimensional distributions, and Multi Dimensional Scaling (MDS) [2] for low dimensional projection of distances. Some additional visual tools, both for projection quality diagnosis and for user interpretation of the projection axes, are also provided in the package. We demonstrate the strengths and advantages of CytoMDS for QC of cytometry data on real biological datasets, revealing the presence of low quality samples, batch effects and biological signal between sample groups.

References
[1] Haidong Yi and Natalie Stanley. 2022. “CytoEMD: Detecting and Visualizing between-Sample Variation in Relation to Phenotype with Earth Mover’s Distance.” In Proceedings of the 13th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics, 1–14. BCB ’22 28. New York, NY, USA: Association for Computing Machinery.
[2] Jan de Leeuw and Patrick Mair. 2009. “Multidimensional Scaling Using Majorization: SMACOF in R.” Journal of Statistical Software 31 (3): 1–30.

Communicating about Statistics

Speakers:

Christian Ritter and Laura Symul, UCLouvain

Summary: 

In this seminar, we will discuss how to communicate efficiently about statistics via graphics, tabulation, presentation, and writing.

In the first half, from 14h30 to 15h30, we will focus on visual communication through graphics and tables. We will distinguish between using graphics and tables for ourselves and preparing graphics and tables for others. We will introduce the three laws of communication, show how perception and brain function are related to decoding graphics, tables, and displays, and give guidance for creating efficient and beautiful displays of quantitative information.

In the second half, from 16h to 17h, we will discuss written communication and oral presentations. Here, we will distinguish between communication in scientific and professional contexts. In the scientific context, there are journal articles and conference contributions such as posters and talks. In the professional context, there are technical reports and short presentations to project teams and management. Written and broadcast media are a special case. 

Reading materials:

1. Edward Tufte: The Visual Display of Quantitative Information, Envisioning Information, and Seeing with Fresh Eyes
2. Stephen Kosslyn: Elements of Graph Design, Graph Design for Eye and Mind
3. A. S. C. Ehrenberg (1986) ‘Reading a Table: An Example’, Journal of the Royal Statistical Society. Series C (Applied Statistics), Vol. 35, No. 3, pp. 237-244.
4. A. S. C. Ehrenberg (1981) ‘The Problem of Numeracy’, The American Statistician, Volume 35, Issue2, pp. 67-71.
5. Jean-Luc Doumont (2009) ‘Trees, Maps, and Theorems’, available from the authors at www.principiae.be.
6. Jean-Luc Doumont (2013) “Creating Effective Slides”, talk given at CTLStanford https://youtu.be/meBXuTIPJQk
7. Uri Alon (2009) “ How to give a good talk” https://www.cell.com/molecular-cell/pdf/S1097-2765(09)00742-4.pdf
8. Joshua Schimel. “Writing Science: How to Write Papers That Get Cited and Proposals That Get Funded”

  TEAMS 

Xavier Milhaud

Associate Professor at the university Aix-Marseille (Institute of mathematics)

Impact of heat waves on mortality: extension of parametric longevity models to account for global warming

Abstract :
Global warming significantly impacts human health, with intense heat waves causing spikes in mortality, particularly among the elderly. Recently, longevity models have started to incorporate a temperature-related component to account for climate change. We propose an extension of recent works, which specifically models the impact of heat waves. Our approach complements existing longevity models, demonstrates better robustness indicators, and can forecast future mortality based on various climate scenarios. In this regard, we present the impacts on french mortality of three Representative Concentration Pathway (RCP) scenarios from the last IPCC.

ONLINE SEMINAR

Fabian Mies (TU Delft) 

Projection inference for high-dimensional covariance matrices 

Abstract: 
Analyzing large samples of high-dimensional data under dependence is a challenging statistical problem as long time series may have change points, most importantly in the mean and the marginal covariances, for which one needs valid tests. Inference for large covariance matrices is especially difficult due to noise accumulation, resulting in singular estimates and poor power of related tests. The singularity of the sample covariance matrix in high dimensions can be overcome by considering a linear combination with a regular, more structured target matrix. This approach is known as shrinkage, and the target matrix is typically of diagonal form. In this paper, we consider covariance shrinkage towards structured nonparametric estimators of the bandable or Toeplitz type, respectively, aiming at improved estimation accuracy and statistical power of tests even under nonstationarity. We derive feasible Gaussian approximation results for bilinear projections of the shrinkage estimators which are valid under nonstationarity and dependence. These approximations especially enable us to formulate a statistical test for structural breaks in the marginal covariance structure of high-dimensional time series without restrictions on the dimension, and which is robust against nonstationarity of nuisance parameters.

Joshua Loftus (London School of Economics (LSE)) 

Causal interpretability for human-centered data science

Abstract: 
Tools for interpretable machine learning or explainable artificial intelligence can be used to audit algorithms for fairness or other desired properties. In a "black-box" setting--one without access to the algorithm's internal structure--an auditor can only use model-agnostic methods based on varying inputs while observing differences in outputs. These include popular interpretability tools like Shapley values and Partial Dependence Plots. But such methods have important limitations that can impact audits with consequences for outcomes such as fairness. In high-stakes applications, it may be worth the effort to use tools that can incorporate background information and be tailored for specific use-cases. We introduce promising ways to do this using the mathematics of causality, with Causal Dependence Plots serving as an example.

Jeroen Rombouts (ESSEC Business School) 

Modeling Higher Moments and Risk Premia for S&P 500 Returns

Abstract: 
We study the impact of additional option pricing model factors on the level, term structure and conditional properties of index return moments and their risk premia. Higher moments are more informative than model-implied equity premia, variances, and variance risk premia for assessing model performance. Based on estimates from a joint option and return likelihood obtained using novel estimation techniques, we relate these model properties to differences in option and return fit. Including three stochastic volatility factors greatly improves option fit. The resulting time series of skewness and kurtosis better match non-parametric benchmarks and the model generates larger skewness and kurtosis risk premia, but it struggles to match the term structure of higher moments. Return jumps improve the modeling of the term structure of skewness and kurtosis and generate larger and more variable skewness and kurtosis risk premia at short horizons, but do not improve option fit in the presence of three stochastic volatility factors.

Authors: Arnaud Dufays, Kris Jacobs and Jeroen Rombouts

Anastassia Negrouk (MyData-Trust) 

Health Data, AI, and Privacy: Striking the Right Balance for Research and Innovation. 

Abstract: 
The global rise of digital health technologies and artificial intelligence (AI) has unlocked immense potential for advancing medical research, personalized care, and public health initiatives. However, these developments also pose significant challenges in the realm of data protection, particularly concerning the sensitive nature of health data.

The secondary use of health data—such as in medical research, AI model training, and healthcare analytics—offers valuable insights but raises privacy risks and ethical concerns. Regulatory frameworks like the General Data Protection Regulation (GDPR) mandate strict controls on the handling of personal health information, complicating international data sharing and analysis efforts. One of the key solutions to these challenges is the anonymization of health data or the use of synthetic data. While anonymization seeks to remove identifying features, ensuring complete privacy without introducing biases or compromising data utility remains difficult. Balancing the benefits of secondary health data use with the need for privacy, accuracy, and compliance remains a critical issue requiring global collaboration and innovative approaches to data governance and AI ethics.

Anastassia will try to explore these different elements while providing practical insights and recommendations for addressing the challenges associated with data protection, privacy, and the secondary use of health data, considering the role of anonymization and synthetic data including for AI-driven innovation.
 

Online access via teams is available but the quality cannot be guaranteed

TEAMS : https://teams.microsoft.com/l/meetup-join/19%3ad71dd7a4c2ba40119f66d2f0d10ae55a%40thread.tacv2/1739269433948?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

Alexander Munteanu 

\ell_p Sensitivity Sampling: Optimal bounds and an Application to Poisson pth-Root-Link Models 

Abstract: 
Sensitivity sampling is a general purpose technique for importance subsampling that is very popular for the construction of \ell_p subspace embeddings. These methods are important building blocks with broad applications in Machine Learning, Computational Statistics, and Computer Science.

Although other subsampling distributions have been shown to achieve smallest possible sample size for constructing \ell_p subspace embeddings, existing analyses of sensitivity sampling fall behind. However, sensitivity sampling is conceptionally and computationally
simpler than other methods and performs equally well or often better in practice. This motivates to reconsider the complexity of constructing \ell_p subspace embeddings via sensitivity sampling.

We first prove that sensitivity sampling is indeed suboptimal in the worst case. However, we introduce a new variation that samples proportional to a mixture of \ell_p and \ell_2 sensitivities. This \ell_2 augmentation technique allows us to obtain a provably optimal
subsample size. As an application, we show how sensitivity sampling can be used to approximate Poisson regression with pth-root-link.

The talk is based on the following two publications (also available on arXiv):

* Alexander Munteanu, Simon Omlor.
Optimal bounds for \ell_p sensitivity sampling via \ell_2 augmentation.
International Conference on Machine Learning (ICML), 2024.
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2406.00328&data=05%7C02%7Ceugen.pircalabelu%40uclouvain.be%7C943f388491f841e2ab1408dd4683bccc%7C7ab090d4fa2e4ecfbc7c4127b4d582ec%7C1%7C0%7C638744253815359406%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=Zak%2BklIlL7DZeNveBJG2BrT%2BVpQv%2F1nx%2BEPGZcvbmc0%3D&reserved=0

* Han Cheng Lie, Alexander Munteanu.
Data subsampling for Poisson regression with pth-root-link.
Advances in Neural Information Processing Systems (NeurIPS), 2024.
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2410.22872&data=05%7C02%7Ceugen.pircalabelu%40uclouvain.be%7C943f388491f841e2ab1408dd4683bccc%7C7ab090d4fa2e4ecfbc7c4127b4d582ec%7C1%7C0%7C638744253815379684%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=JtXte4tm2MqjT1E4M3Q80k7ar00QN5GtbFmNtMtGT24%3D&reserved=0

Christian Ritter, UCLouvain 

Statistical Programming using Prompts in ChatGPT, Claude, Mistral, and DeepSeek 

Abstract: 
This is an open workshop on the topic on experiences with prompt writing to create scripts for statistical analyses in R and Python.
The open session will be animated by Christian Ritter, UCLouvain who will start by presenting a few recent examples in teaching and consulting work using R and Python. This includes building shiny applications, accessing Eurostat data sources, developing time series analyses, and carrying out simple statistical analyses directly in a chat. The tests will be run in ChatGPT 4o, Claude 3.5 Sonnet, Mistral, and Deepseek R1. 
After this introductory part, participants can tell about their own experiences or show results of their own tests. If you would like to show some of your own experiments, please contact christian.ritter@uclouvain.be to facilitate the setup.
 

Online access via teams is available but the quality cannot be guaranteed.

Link: https://teams.microsoft.com/l/meetup-join/19%3ad71dd7a4c2ba40119f66d2f0d10ae55a%40thread.tacv2/1739286566263?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d

Marie-Pier Côté (l’université Laval) 

Un modèle hiérarchique flexible pour les réclamations d'assurance combinant gradient boosting et copules 

Résumé : 
Nous proposons un modèle hiérarchique pour les réclamations en assurance de dommage qui affine les méthodes traditionnelles en tenant compte de la dépendance entre les occurrences de paiement, avec une distribution multinomiale, et entre les montants de paiement, avec des copules. Nous effectuons une prévision qui dépend de variables explicatives en utilisant XGBoost, un algorithme performant de gradient boosting, ce qui nous permet d'améliorer la performance prédictive par rapport aux modèles linéaires généralisés usuels. La construction et l'ajustement du modèle sont illustrés sur des données réelles d’assurance automobile provenant d'une grande compagnie d'assurance canadienne. L'utilisation de XGBoost est bien adaptée à ces données volumineuses contenant un grand nombre d'assurés et de covariables. Les méthodes d’inférence basées sur les rangs pour les copules sont standards, mais leur validité lorsque les distributions marginales sont obtenues par gradient boosting n'a pas été démontrée dans la littérature antérieure. Nous réalisons donc une étude de simulation pour évaluer la performance des méthodes basées sur les rangs des résidus pour l’estimation de la structure de dépendance. Nous montrons quelques applications de notre modèle. Dans une comparaison avec des modèles de référence, nous concluons que les composantes de dépendance de notre modèle améliorent la segmentation et reproduisent mieux le comportement aléatoire global.

Lien TEAMS

15h25 - 15h35 - Pause

• 15h35 - 16h15 - Valentin Patilea, ENSAI, Rennes
  Title: From the observed variables to the latent world: a general approach for survival analysis
• 16h15 - 16h25 – Conclusions et Présentation du livre, Jean-Pierre Florens, Professeur émérite, TSE, Toulouse School of Economics
• 16h25 - 16h30 – Accueil au drink et quelques mots en l’honneur de Jean-Marie Rolin, notre collègue et ami - Léopold Simar, Professeur émérite à et fondateur de l’Institut de statistique (STAT, devenu ISBA-LSBA-SMCS), UCLouvain et Michel Mouchart, Professeur émérite à l’Institut de statistique (STAT, devenu ISBA-LSBA-SMCS), UCLouvain
• 16h30 - Drink
   

Link to registration 

Joelle Desterbecq (UCLouvain) 

Open Science et gestion des données de la recherche : objectifs, principes et ressources

Résumé: 
Vous souhaitez en savoir plus sur l’Open Science ? Vous démarrez un nouveau projet et vous devez établir un plan de gestion des données ? Vous souhaitez utiliser l’outil DMP Online ? Vous souhaitez publier vos données de recherche en mode « Open » ou selon les principes « FAIR » ? Ce workshop s’adresse à vous. Il abordera les points suivants :

• Les objectifs et les différentes dimensions de l’Open Science 
• La gestion des données de la recherche (RDM) et les principes « FAIR »
• Etablir un Data Management Plan
• La publication de données de la recherche en Open Data
• Les outils et ressources de l’UCLouvain en soutien à l’Open Science 
   

Exceptionnellement, ce séminaire sera accessible uniquement aux membres (étudiants, assistants, etc.) de l’UCLouvain

Prof. Yves Moreau (KULeuven) 

will give a presentation on : 

"Human genetic research in the age of mass surveillance". 

Abstract : 

Human genetic research is advancing rapidly, facilitated by the availability of large-scale genomic data sets and the development of high-throughput sequencing technologies. At the same time, the rise of mass surveillance has led to concerns about the potential misuse of genetic data, particularly in relation to privacy and discrimination. This has raised important ethical questions about the collection, storage, and use of genetic data. I will discuss several cases of misuse of forensic DNA profiling and the involvement of different actors including companies, researchers, and scientific journals. I will discuss the risks that such misuses represent for the public trust in human genetics and the need to enforce ethical frameworks and guidelines that ensure the responsible use of genetic data. Researchers are thus faced with the challenge of balancing the benefits of genetic research with the need to protect individual privacy and autonomy. Human genetic research must acknowledge that we live in a time where the fear of mass surveillance is a legitimate concern. While the potential benefits of genomic research are vast, researchers must navigate complex ethical and legal landscapes to ensure that the collection and use of genetic data is carried out in a responsible and ethical manner. They must also preserve the trust society places in healthcare providers.

See also :

Doctorats honoris causa - Programme 27-28 février 2025

Morning sessions : 
09h15 - Introduction
09h30 - Peter Bühlmann (ETH Zürich): "Borrowing Strength from Others (A Lesson from Large-Scale Intensive Care Data)" 
09h55 - Johannes Lederer (University of Hamburg): "Sparse Score Matching: From Classical Statistics to Generative Models"
10h20 - Lutz Dümbgen (University of Bern): "Nonparametric Smoothing of Directional and Axial Data" 
10h45 - Coffee break
11h15 - Qiwei Yao (London School of Economics): "Cointegration Between Two Intrinsically Stationary Spatial Processes" 
11h40 - François Roueff (Institut Polytechnique de Paris, Telecom Paris): "Some recent advances in variational inference"

12h15 - Lunch

Afternoon session : 
14h00 - Guy Nason (Imperial College London): "Locally Stationary Wavelet Processes: A Celebration of the Contributions of Professor Rainer von Sachs"
14h25 - Michael H. Neumann (Friedrich-Schiller-Universität Jena): "Wavelet thresholding for polynomial-tailed noise" 
14h50 - Hernando Ombao (King Abdullah University of Science and Technology): "Modeling Dependence in Time Series Analysis: A Journey" 
15h15 - Coffee break
15h45 - Véronique Delouille (Royal Observatory of Belgium): "A journey in nonparametric statistics with Rainer, from design-adapted wavelet to characterization of uncertainty in the International Sunspot Number"
- Sophie Mathieu and Joris Chau (Open Analytics): "Chemometric applications of Gaussian processes in the frequency domain" 
- Hilmar Böhm: "Rainer’s Deepest Secrets, Revealed by His Personal Psychoanalyst"

17h15 - Drink and social activity

---
September 11, 2025
AULA MAGNA (Foyer royal, 1st floor)
Place Raymond Lemaire 1
1348 Louvain-la-Neuve Belgium

---

REGISTRATIONS ARE CLOSED

Hotel

In case you need accommodation, we suggest Martins hotel, in the city center of Louvain-la-Neuve (2 minutes away from the venue).  
Other options are available in the area of the campus or outside the campus. Commuting to Louvain-la-Neuve is then possible via public transport or hotel shuttles. 

• Martins hotel
  Rue de l’Hocaille 1
  1348 Louvain-la-Neuve, Belgique
  Phone: +32 (0)10 77 20 20
  Website : Martins hotel
   
• Hotel Ibis Styles Meeting Center
  Boulevard de Lauzelle, 61
  B - 1348 Louvain-la-Neuve, Belgique
  Phone: +32 (0)10 53 90 00 (booking) +32 (0) 10 45 07 51 (Reception)
  Website : Ibis Styles
   
• Best Western Wavre Hotel (with shuttle to LLN)
  B - Avenue Lavoisier 12 - 1300 Wavre, Belgique
  Phone: +32 (0)10 88 74 30
  Website : http://www.bestwesternwavre.com
   

Jens Robben (University of Amsterdam)

The short-term association between environmental variables and mortality: evidence from Europe

Abstract: 
In this workshop, we study the short-term association between environmental factors, i.e., weather and air pollution characteristics, and weekly mortality rates using fine-grained, publicly available data. Hereto, we develop a mortality modeling framework where a baseline model describes a region-specific, seasonal trend observed within the historical weekly mortality rates. Using a machine learning algorithm, we then explain deviations from this baseline using features constructed from environmental data that capture anomalies and extreme events. 

• Session 1 (2:30 p.m. - 3:30 p.m) - C.115: We provide the technical details of our proposed modeling framework, and apply it to European NUTS 3 regions (Nomenclature of Territorial Units for Statistics, level 3). Our findings highlight that temperature-related features are most influential in explaining mortality deviations from the baseline over short time periods. Furthermore, we find that environmental features prove particularly beneficial in southern regions for explaining elevated levels of mortality, and we observe evidence of a harvesting effect related to heat waves.
   
• Session 2 (4 p.m. - 5 p.m.) - C.045 / Salle Gauss: Through a hands-on case study in R, participants will implement a simplified version of our modeling framework. Through various code examples and illustrations, we demonstrate data processing steps, calibrate both the baseline and machine learning models, and extract key model insights.

Material 
https://jensrobben.github.io/Workshop-LLN/

Teams 
Link to part one (14h30)
https://teams.microsoft.com/l/meetup-join/19%3a4d3e563eb1444f6485c9568d83ef1e48%40thread.tacv2/1744643398615?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d
Link to part two: (16h)
https://teams.microsoft.com/l/meetup-join/19%3a4d3e563eb1444f6485c9568d83ef1e48%40thread.tacv2/1744643731025?context=%7b%22Tid%22%3a%227ab090d4-fa2e-4ecf-bc7c-4127b4d582ec%22%2c%22Oid%22%3a%22b05c410a-2e16-4d73-b134-e0adf3e0d016%22%7d
 

Thomas Delclite(StatBel)

Méthodologie des enquêtes en statistique publique : de l’échantillonnage à l’estimation de la variance 

Abstract: 
Ce séminaire présente les principales étapes méthodologiques mises en œuvre dans la production d’enquêtes statistiques à Statbel. Nous aborderons les principes du tirage d’échantillons dans des plans complexes, le calcul des poids d’extrapolation, le traitement de la non-réponse, ainsi que les méthodes d’estimation de la variance des estimateurs. L’objectif est de montrer comment ces éléments s’articulent pour garantir la qualité, la représentativité et la robustesse des statistiques produites dans un cadre institutionnel. Des exemples concrets issus de trois enquêtes seront mobilisés pour illustrer ces concepts : l’enquête sur les revenus et les conditions de vie (SILC), l’enquête sur les voyages (Travel Survey), et l’enquête sur l’usage des TIC par les ménages (ICT-HH). 

TEAMS

Anna Malinovskaya (Nala Earth)

Beyond Tables: Unlocking the World of Spatial Data

Abstract: 
This presentation explores how geospatial data is different beyond traditional data structures. We will examine the growing importance of spatial analytics in research and industry applications, discuss various types of geospatial data and review helpful tools. The talk addresses both theoretical foundations and practical implementations, making it accessible to newcomers while offering valuable insights for experienced data practitioners. Participants will gain perspective on how spatial dimensions can enhance analytical capabilities and uncover main challenges on the way of mastering it. 

Examples from current practice will be given within the limits of confidentiality.

TEAMS
 

José Miguel Flores-Contró (HEC Lausanne / Azenes AG)

Subsidizing Inclusive Insurance to Reduce Poverty

Abstract:
In this article, we assess the benefits of coordination and partnerships between governments and private insurers, and provide further evidence for microinsurance products as powerful and cost-e˙ective tools for achieving poverty reduction. To explore these ideas, we model the capital of a household from a ruin-theoretic perspective to measure the impact of microinsurance on poverty dynamics and the governmental cost of social protection. We analyze the model under four frameworks: uninsured, insured (without subsidies), insured with subsidized constant premiums and insured with subsidized flexible premiums. Although insurance alone (without subsidies) may not be suÿcient to reduce the likelihood of falling into the area of poverty for specific groups of households, since premium payments constrain their capital growth, our analysis suggests that subsidized schemes can provide maximum social benefits while reducing governmental costs.

Gery Geenens (UNSW Sydney) 

Universal Copulas 

Abstract: 
Copulas have emerged over the last decades as primary statistical tools for modelling dependence between random variables. A copula is classically understood as a cumulative distribution function on the unit hypercube with standard uniform margins – we refer to such distributions as “Sklar’s copulas”, owing to their central role in the decomposition of multivariate distributions established by the celebrated Sklar's theorem.

A standard argument in favour of copula models is that they separate the dependence structure (encoded by the copula) from the marginal behaviour of individual components. However, this interpretation holds only in the continuous case: outside it, copulas lose their “margin-free” nature, rendering Sklar’s construction unsuitable for modelling dependence between non-continuous variables.

In this work, we argue that the notion of a copula need not be confined to Sklar’s framework. We propose an alternative definition -- universal copulas -- based on a more precise characterization of dependence. This new definition agrees with Sklar’s copulas in the continuous case, but yields distinct and more suitable constructions in discrete or mixed settings. Universal copulas retain key properties such as margin-freeness, making them sound and effective beyond the continuous realm. We illustrate their use through examples involving discrete variables and mixed pairs, such as one continuous and one Bernoulli variable.
 

Matteo Fontana (Royal Holloway, University of London) 

Multi-Output Conformal Regression 

Abstract: 
This presentation offers a unified comparative study of conformal prediction methods for multi-output regression. While conformal prediction provides robust, distribution-free coverage guarantees, its application to multi-output problems remains challenging due to output dependencies and computational costs. We structure several methods, including different conformity scores, into a common framework to highlight their methodological differences and connections. Finally, we evaluate their empirical performance on a variety of real-world tabular datasets.

Clement Berenfeld 

(Institut national de recherche en sciences et technologies du numérique) 

Will give a presentation on : 

A causal framework for reliable membership inference attack evaluation

Abstract :  
Quantifying memorization is central for assessing privacy risks in machine learning. The standard tool for this purpose, membership inference attacks (MIAs), traditionally re-quires multi-run evaluations (repeated retraining) that are computationally prohibitive for modern large-scale models. This has led to the adoption of one-run methods (training once with a randomized subset of points) and zero-run methods (evaluating models “as-is”), though their statistical soundness remains unclear. We address this by reframing MIA evaluation as a causal inference problem, defining memorization as the causal effect of a data point’s inclusion in the training set. Our work reveals systematic issues: one-run regimes introduce interferences between jointly inserted points, and zero-run regimes introduce confounding from non-random membership assignments. We formalize these challenges by introducing a new interference model for treatment effect estimation, derive causal counterparts to standard MIA evaluation metrics, and propose estimators that are provably consistent by leveraging learning-theory properties.

Sylvain Sardy (Université de Genève) 

The Pivotal Information Criterion 

Abstract: 
The Bayesian and Akaike information criteria aim at finding a good balance between under- and over-fitting. They are extensively used everyday by practitioners. Yet we content they suffer from three afflictions: their inherent (best subset) discrete optimization is infeasible beyond moderate dimension, their need for estimation of a nuisance parameter makes them inefficient in high dimension, and their penalty parameter λ = log n and λ = 2 are too small for feature detection. We alleviate these issues with the pivotal information criterion.

• 9:00 - Leon Rofagha 
  “A discrimination measure under coarsening”

  Abstract: “Abstract. How does one evaluate the quality of a regression model’s predictions \(f(X_1),\ldots,f(X_n)\) under coarsening of the responses \(Y_1,\ldots,Y_n\)? A coarsening of a random variable is a random set containing said random variable. We construct a simple discrimination measure estimating the concordance index in this setup, prove its root-\(n\) asymptotic normality under an i.i.d.-sum condition for the estimated distribution function of \(Y\mid f(X_i)\), and illustrate its performance through experiments by specialising the coarsening mechanism to case-\(K\) interval-censoring.”

• 9:30 - Charlotte Jamotton 
  “A multi-criteria fair Gaussian regressor for insurance premium”

  Abstract: “This article investigates how multiple fairness criteria, such as demographic parity and proxy discrimination mitigation, can be embedded within a Gaussian Process Regression (GPR) framework. We propose a single Bayesian non-parametric model that incorporates fairness interventions through kernel design and objective function modifications. Specifically, we alter the kernel to control similarity structure (e.g., to mitigate omitted variable bias) and extend the objective beyond predictive accuracy to include fairness constraints such as enforcing independence between the premium and sensitive variables. This modified GPR architecture allows us to jointly enforce multiple fairness definitions, spanning both group and individual-level criteria, within a single model. We empirically explore trade-offs with actuarial fairness, and how different fairness criteria interact when combined. The results highlight the importance of adopting a multi-criteria, context-aware approach to fairness in insurance pricing." 

• 10:00 - Guillaume Deside 
  “Errors-in-Variables Bayesian Model of Glycemic Response to Lifestyle Factors”

  Abstract: “In recent years, the volume of data collected in clinical research has increased dramatically. This trend is driven, in part, by the proliferation of wearable medical devices and sensors, which enable both healthy and ill individuals to gather extensive longitudinal data about their health and lifestyle in situ (i.e., in the context of people’s daily lives rather than in artificial clinical environments) (Paranjape 2020). Despite their clinical potential, in situ data present statistical challenges including high missing data proportions, inaccuracies, noisier measurements than traditional clinical data, and uncertainties in event timing and individual response heterogeneity (Salathe 2024). For this presentation, we focused on the example of explaining blood glucose levels, as measured with continuous glucose monitoring (CGM) wearable sensors, using self-reported nutrient intakes and lifestyle and physiological factors. The first and simplest approach considered was to discretize the observed time-series and use linear models or random forests (RF) to roughly assess the importance of the different inputs and the shape of the glycemic responses to different factors. This approach has two important limitations. First, discretization leads to information loss that may compromise inference on parameters related to rapid glucose responses. Second, it does not account for potential timing errors in input factors, which are common in real-world health monitoring data. For example, someone may forget to report a meal or report it several hours after they actually consumed that food. Consequently, we turned to a different framework, building on Zhang et al.’s interpretable error-in-variable Bayesian model (Zhang et al 2021). Specifically, we further developed and generalized that model to include factors that were absent in the initial model and identified as important by the linear and RF models, and to account for a larger diversity of potential glycemic responses. In this presentation, I will present the results obtained from the different approaches when applied to longitudinal data from the Food&You study (Héritier et al 2023), highlight their strengths and limitations, agreements and divergences, and discuss future developments."

• 10:30 - Break

• 10:45 - Mirco Lescart 
  "A flexible sub-asymptotic model for threshold exceedances"

  Abstract : Extreme rainfall, financial crises, and structural failures often involve several variables becoming extreme at once. Standard extreme-value models capture this joint behavior only in the very far tails and usually treat dependence in a rigid way, focusing either on asymptotic dependence or on asymptotic independence. Only a few recent models allow for more flexibility between the two. I present a new sub-asymptotic model that provides a gradual transition between dependence regimes, combining ideas from generalized Pareto theory and flexible mixture representations. Estimation relies on modern likelihood-free inference through the Neural Bayes Estimator, supported by simulation studies. Finally, I illustrate how the model reveals different dependence patterns in Belgian rainfall extremes, from strong joint behavior to near-independence.

• 11:15 - Kamal Gasser  
  “Heatwave attribution over Europe”

  Abstract: “We are investigating if the rise in global mean surface temperature caused by hu-man forcing has resulted in a change in temperature distribution, hence increas-ing the frequency of heatwaves. Heatwaves manifest themselves in both spatial and temporal dimensions ; nevertheless, the majority of attribution studies con-centrate on individual locations and analyze heatwaves by calculating specific indices to capture their spatiotemporal dependency. However, this technique seems inadequate since indices do not accurately represent the real events. We aim to address this gap by using advanced statistical models and concentrating our attribution analysis on the whole of Europe. We use a clustering technique using a distance measure suitable for extremes to pinpoint spatial regions where concurrent severe temperatures occur simultaneously. Then, we formulate a flex-ible non-stationary space-time extreme value model that accommodates diverse forms of asymptotic dependency. Lastly, we use our approach to climate model outputs to estimate return period of extreme event under both a factual and counterfactual world." 

• 11:45 - Edouard Motte 
  “Signature methods in mathematical finance”

  Abstract: “The signature of a path consists of the sequence of its iterated integrals. Path-signature theory has recently emerged as a powerful tool in the fields of machine learning and mathematical finance, mainly due to its universal linearization property, according to which, provided sufficient regularity, any path functional can be expressed as a linear combination of signature elements. In this talk, I will introduce the theory of signatures and discuss some applications in mathematical finance. More specifically, I will show how signatures can be used to solve the (non-linear) stochastic control problem of hedging path-dependent options in the presence of market frictions.

Laurent Franckx (Bureau du Plan)

Scrappage models for cars: multi state processes versus duration models

Abstract: 
I illustrate how a Markov multi-state process can be used to model car scrappage decisions in car stock models, as an alternative to simple duration models. In contrast to duration models, multi-state processes take into account that cars often change hands between owners who exhibit quite different scrappage behaviour.

Using a real-world numeric example, I show that using a simple duration model can lead to a serious misrepresentation of the size, the age structure and the ownership structure of a car stock. In turn, this leads to a low accuracy in the estimation of CO2 emissions from the car stock and of the tax receipts based on car ownership.

Simulating the car stock with a multi-state process leads to more accurate estimates of all key parameters of the car stock.

This seminar will be given in presence in room C-115, 20 voie du Roman Pays, Louvain-la-Neuve. Online participation is also possible but we can't guarantee high quality.

Room C.115 or TEAMS
Lien: https://teams.microsoft.com/meet/3374883247628?p=npNiQH7UVz9I3N0O44
 

Estelle Medous (IGN, France)

Many-to-One indirect sampling with application to the French postal traffic estimation

Abstract: 
In probabilistic surveys, when there is no sampling frame for the target population, a solution is to find a frame population linked in some way to the target population and use indirect sampling. The sampling weights can be determined using the generalized weight share method (GWSM). However, this method cannot be applied when some of the links between the frame population and the sample in the target population are missing or difficult to retrieve exhaustively.

A solution to avoid this issue is to consider an intermediate population linked in some way to both the frame and target populations and use a double indirect sampling. Then the GWSM can be used twice, first between the frame and intermediate populations and then between the intermediate and target populations. As illustrated with the French postal traffic survey, this double indirect sampling appears to be deteriorating the precision of estimators in some situations. Using mathematical derivations, it is possible to highlight the magnitude of the loss of precision in practical situations similar to the French postal context. The talk will start with an introduction to Survey sampling theory.

It will be followed by a presentation of GWSM and double GWSM as well as the theoretical results. Finally, results will be illustrated through Monte Carlo simulations tailored to the French postal traffic estimation.
This seminar will be given online via Teams. It will also be streamed to room C-115, 20 voie du Roman Pays, Louvain-la-Neuve.

Lien: https://teams.microsoft.com/meet/3117965410214?p=lfIkeiZT6eZ8wUG6jn
 

14h30:

Bruno Deprez (U Antwerpen)

"Network-based  anti-money laundering"

Abstract:
Network analytics has proven to be essential for anti-money laundering (AML), since money laundering schemes span many transactions between multiple parties. One of the main challenges in the field is that fraudsters try to evade detection by constantly adapting their laundering methods. This means that financial institutions need to fine-tune and update their internal AML models on a regular basis. However, computing power is the main bottleneck since banks typically need to monitor millions of transactions. Therefore, continual (graph) learning is becoming an essential tool in having an effective AML system. This talk will discuss some recent work by giving an overview of the current methods applied for network-based AML and it will dive into the key aspect of continual graph learning. The topics and ideas will be illustrated on an AML use case.

15h30: Coffee Break

16h00:

Olivier Caelen (UCLouvain)

"Machine Learning for Payment Fraud Detection"

Abstract:
Payment fraud remains a major challenge for financial institutions, requiring models that can learn from vast and complex transaction streams. This presentation will explore how machine learning can be applied to detect fraudulent credit and debit card payments. After introducing the context of card fraud and the operational constraints of (near) real-time detection, we will examine the main methodological challenges like concept drift, sampling bias, class imbalance, and others... We will then conclude with a discussion on how techniques such as clustering and graph mining can be use to build more robust features, improve detection accuracy and reduce false positives.
 

Oliver Dukes (UGent) 

Nonparametric tests of treatment effect heterogeneity for policy-makers

Abstract: 
Recent work has focused on nonparametric estimation of conditional treatment effects, but inference has remained relatively unexplored. We propose a class of nonparametric tests for both quantitative and qualitative treatment effect heterogeneity. The tests can incorporate a variety of structured assumptions on the conditional average treatment effect, allows for both continuous and discrete covariates and does not require sample splitting. Furthermore, we show how the tests are tailored to detect alternatives where the population impact of adopting a personalised decision rule differs from using a rule that discards covariates. The proposal is thus relevant for guiding treatment policies. The utility of the proposal is borne out in simulation studies and a re-analysis of an AIDS clinical trial. This is joint work with Mats Stensrud, Riccardo Brioschi and Aaron Hudson.

Cedric Taverne (GSK) 

(joint work with Mario Becerra Contreras, Mathias Schmitz, Oluwafemi Daniel Olusoji, all GSK)

Borrowing through the evolving vaccine preclinical research plan: Prior elicitation in absence of recurrent designs or controls

Summary:
The preclinical research plan in vaccine development consists of a project-specific mixture of small sample size (n<10 per group) animal in-vivo experiments and human in-vitro experiments (often with 2 to 3 donors) aimed at identifying one vaccine candidate that demonstrates the best immunogenicity while generating some but limited reactogenicity. These experiments consist of discrete screening designs, dose response designs, challenge studies, and potency tests, just to name a few.
The vaccine candidate formulations evolve largely across the preclinical plan, making it challenging to borrow from previous studies. Additionally, positive controls do not exist in most of the vaccine projects, whereas negative controls are mostly invariant non-measurable responses (lower than the bioassay technical limit) and are discarded in all statistical models. These vaccine specificities make prior elicitation and Bayesian borrowing strategies quite challenging to implement in preclinical studies as compared to clinical phases.
Therefore, we have experimented with borrowing model coefficients rather than group locations and scales, of which we found limited echo in the literature. Using case studies, we will illustrate the borrowing of model effects such as the antigen dose slope or the interaction between vaccine components. Posterior distributions of these parameters from previous studies have been aggregated using meta-analytic-predictive priors and other methods. Sensitivity analysis of those borrowings on model predictions, decision making, and animal use reduction (aligned with GSK 3Rs commitment on animal studies) will be discussed.

Bio:
All authors are statisticians within the Vaccines Research and Translational Statistics team of GSK. They support the development of new vaccines from the first research stages until the vaccine projects are fully transferred to their clinical colleagues. Using Bayesian statistics, they explore disruptive ways to value the data collected and accelerate the development of new vaccines.

Teams: 
https://teams.microsoft.com/meet/3848793947750?p=ihUxXPCvvHUmVDP1NH
 

Swati Chandna 

(Birkbeck, University of London) 

Will give a presentation on : 

Understanding structure in networks observed with covariates via profile least squares

Abstract: 
Network data across various domains (e.g. social, political, biological) are commonly observed with additional node- and/or edge- specific attributes that may explain the intensity of interactions between nodes within the network. For example, in the study of international relations, attributes such as conflict intensity and volume of trade are observed for each country pair, and it is important to understand the extent to which these covariates explain the formation of military alliances between countries (nodes). We study the model where probabilities of edge formation between any two nodes in the network are given by the sum of a linear edge-covariate term and a residual term to model the remaining heterogeneity from unobserved factors. We approach estimation of the model via profile least squares and show how it leads to a simple algorithm to estimate the linear covariate term and the residual structure that is truly latent in the presence of observed covariates. Our framework lends itself naturally to a bootstrap procedure which is used to draw inference on model parameters, such as to determine significance of the homophily parameter or covariates in explaining the underlying network structure. We illustrate the usefulness of our methodology through simulations and application to four real network datasets.  

Jeroen Kerkhof (VUB & UGent)

Socio-economic mortality curves: the Belgian case

Abstract:
In this paper, we create up-to-date socio-economic mortality tables for Belgium. These socio-economic mortality tables provide superior out-of-sample predictions for mortality compared to the commonly used pure gender-based models. We use the socio-economic model for the valuation of socio-economic heterogeneous retirement schemes and we find economically significant differences with the current market practice of using pure gender-based models. In addition to valuation differences due to deviations from the overall population, we also quantify a “convexity” effect for pension funds representative of the general population. This follows from the fact that people with higher (lower) pension payments are more likely to live longer (shorter) than average. In order to assess the impact of socio-economic determinants on longevity, we utilize detailed micro-data for the entire Belgian population obtained from the Belgian statistical office spanning 15–30 years, depending on the variable. We analyze the mortality rates of sub-populations with diverse socio-economic characteristics, using the Li–Lee model. For all socio-economic variables considered, our analysis reveals significant variations (reaching up to ) in survival probabilities for the retirement ages (65+) across distinct socio-economic sub-populations. Furthermore, our analysis indicates that, unlike the diminishing trend observed in the gender gap over time, the impact of socio-economic differences on longevity remains quite stable over the examined period.

Lisa Arnalot (doctorante à l’université de Toulouse) 

Microbiota dynamics in dairy cows around calving and its associations with health

Abstract: 
The transition period (± 3 weeks around calving) is critical for the health and productivity of dairy cows. During this phase, significant physiological and metabolic changes occur, including reduced feed intake and an increased energy demand for lactogenesis. This renders cows more susceptible to metabolic and infectious diseases. The role of the gut microbiota in shaping immunity in the gut and systemically through host–microbiota interactions has been investigated in human biology, but little is known about this in dairy cattle. 
My primary objective was to enhance the description and understanding of the faecal microbiota composition in dairy cows, and its relationship with general health and immunity during the transition period. To do so, several datasets were analysed, relying on a large diversity of analytical methods. First, a thorough meta-analysis of existing faecal microbiota data of lactating Holstein cows was conducted, including more than 2,000 samples. Notwithstanding the challenges posed by inconsistent data quality and methodological heterogeneity, three microbial profiles were established. The findings of this study provide a robust baseline for future comparative research. Second, an experiment was conducted on 25 commercial dairy farms in Brittany (France) to investigate changes in 411 dairy cows during the periparturient period. Shifts in microbial composition were observed within the short timeframe surrounding calving. Finally, we studied the relationship between microbiota and immune parameters, with a particular focus on the cytokine expression and inflammatory response during the transition period. Four immunological profiles were established during this period. Despite the lack of a clear link between these four immune profiles and microbial composition, we found that the variability in cytokine expression could be partly (<10%) explained by the presence of several taxa. Although these associations are exploratory in nature, they have generated specific hypotheses linking host physiology during the periparturient period to microbial ecology. Altogether, this work is part of more global efforts to improve animal welfare, enhance farm profitability, and promote sustainable agriculture from a One Health perspective, which combines animal, environmental, and human health.

TEAMS : Lisa Arnalot (Université de Toulouse): Microbiota dynamics in dairy cows around calving and its associations with health | Meeting-Join | Microsoft Teams
 

Maria Lanzerath

(W.L. Gore & Associates, Germany)

Will give a presentation on : 

From Nose to Tail - a  DOE in Chemical Production

The session will go about an application of DOE to gain fundamental understanding of a chemical process. This will speed up and simplify future DOE for the upscaling of a product idea into the mass production of a commercial product.

Agenda
•    Explain the use case—polymerization process in three production steps with a focus on two of the three steps.
•    Outline the objectives of R&D and process engineering
•    Develop the statistical model.
•    Select the DOE.
•    Show selected results and analyses. 
•    Discuss the value of the whole effort

TEAMS: 
https://teams.microsoft.com/meet/3297473143348?p=UjjxdXH7NJL0YHaxZT
 

SMCS and friends

Organized and moderated by Christian Ritter and Catherine Rasse

The following seminar will be given in presence in room C-115 and in the cafeteria. Online participation is not possible.
It will start with the coffee break right after the talk by Maria Lanzerath. 

Playful Learning: Design of Experiments

Abstract :  
Books, lectures, and videos are excellent for understanding the theory of experimental design and learning which approaches are optimal under specific conditions. But hands-on experience brings these concepts to life in ways that passive learning cannot.
Join us for an interactive exploration of experimental design through carefully crafted games. We'll provide coffee and cookies while you work through challenges at different stations, each highlighting key principles of DOE.
After you've had time to explore, we'll gather briefly in the seminar room to discuss the concepts embedded in each game and connect them to experimental design theory. Then it's back to the games—armed with new insights.
Come ready to think, play, and discover what makes a good experiment good. Don't hesitate to share your own games and experiences. 
 

9:00 Mathias Dah Fienon 
Title : "Dynamic risk-parity portfolio with independent components"

Abstract:
Portfolio diversification involves investing in many different securities and types of assets in order to mitigate risk and maximize returns. One approach to achieve such diversification is the factor-risk-parity model. It assumes that the portfolio weights of uncorrelated factors are equal, up to sign. One criticism of factor-risk-parity is that any orthogonal rotation of these factors will also yield uncorrelated factors, the choice of which is arbitrary. However, if the factors are non-Gaussian, then there is only one rotation that gives independent components, or that minimizes a mutual information criterion. We propose a new methodology that searches for maximally independent factors in a dynamic approach by updating the information from the past using a score-driven model. This fits into a general framework for modelling time varying parameters in time series analysis.
 

9:40 Mengxue Li
Title: "Learning population and individual structure in dynamic networks with degree heterogeneity"

Abstract:
Dynamic networks provide a powerful framework for characterizing time-varying functional connectivity in neuroimaging studies. In practice, such networks are typically collected from multiple subjects across time and exhibit both temporal dynamics and subject-specific heterogeneity. Brain functional connectivity networks also contain hub nodes, defined as highly connected regions that play critical roles in understanding brain functional connectivity. In this talk, we propose a mixed-effect dynamic stochastic block model with degree heterogeneity, which simultaneously disentangles the population connectivity structure from individual variability and recovers the trajectories of hub regions through time-varying degree parameters. We develop an efficient local approximate estimation procedure and evaluate its performance through extensive simulations and a case study of dynamic functional connectivity from the Human Connectome Project.

 
10:10 Gabriel Bailly
Title : "Satterthwaite Approximation and Gaussian Time Series"

Abstract
Satterthwaite (1941, 1946) proposed a very simple approximation to the distribution of linear combinations of Chi-squared random variables. It can be used in univariate time series analysis to approximate the distribution of the sample variance and the periodogram of Gaussian time series; we provide Wasserstein bounds and rates of convergence of the approximation towards the true distribution. Similarly, Tan & Gupta (1983) proposed an approximation to the distribution of linear combinations of Wishart random matrices. This, however, has not yet been applied to the framework of multivariate time series: we take advantage of a special case of the matrix normal distribution to propose a feasible approximation to the distribution of the sample covariance matrix of Gaussian time series.
 

11:00 José Miguel Flores Contro 
Title : "Poverty Trapping: A Ruin Theory Perspective"

Abstract
Trapping refers to the event when a household falls into the area of poverty. Households that live or fall into the area of poverty are said to be in a poverty trap, where a poverty trap is a state of poverty from which it is difficult to escape without external help. Ruin theory, on the other hand, studies stochastic processes and their fluctuations, with its classical application being the modelling of an insurance company’s surplus over time. Since the seminal works of Lundberg
(1903) and Cramér (1930), ruin theory has remained a fundamental area of research in actuarial science. The canonical model in this setting is the Cramér–Lundberg risk process, which has been extensively studied and generalised in the literature. This talk introduces the fundamental principles of ruin theory and examines their application to poverty trapping. In particular, we consider a risk process with deterministic growth and multiplicative jumps, as introduced in Kovacevic and Pflug (2011), to model household capital, incorporating both exponential growth and losses proportional to the current level of capital. Within this framework, we derive closedform expressions for trapping (ruin) probabilities and for the Gerber–Shiu expected discounted penalty function in specific cases, and we discuss how these results relate to the role of policy interventions in the stochastic evolution of household capital. These results illustrate how the mathematical tools of ruin theory provide a stochastic framework for the analysis of poverty dynamics. 

11:40 Robert Paulus
Title : "Adaptive regionalization for extreme precipitation: A neural network-weighted independence likelihood approach" 

Abstract
Recent European floods underscore the high cost of underestimating extreme rainfall. Designing resilient infrastructure depends on estimating precipitation return levels: rainfall amounts associated with very rare events. The challenge is that observational records contain only a small number of extremes, so fitting an extreme-value model separately at each site often yields unstable estimates and very wide confidence intervals. Pooling data across space can reduce uncertainty, but it may also introduce substantial bias when nearby locations do not share the same extreme-behavior patterns.
We propose an adaptive pooling strategy that learns how much information to borrow from surrounding sites. For each target location, we fit an extreme-value model using a neural network–weighted independence likelihood, where the network assigns weights to neighboring observations based on distributional similarity. Extensive simulation studies show a clear bias–variance tradeoff: as the network effectively shrinks the sample size from broad pooling toward purely local fitting, return-level errors first decrease and then increase again, indicating that performance is best at an intermediate level of pooling.

Gianluca Bontempi 

(ULB) 

Will give a presentation on : 

Ethics2vec: aligning automatic agents and human preferences

Abstract :  
Though intelligent agents are supposed to improve human experience (or make it more efficient), it is hard, from a human perspective, to grasp the ethical values that are explicitly or implicitly embedded in an agent's behaviour.
This is the well-known alignment problem, which refers to the challenge of designing AI systems that align with human values, goals, and preferences.
This problem is particularly challenging because most human ethical considerations involve \emph{incommensurable} (i.e., non-measurable and/or incomparable) values and criteria. Consider, for instance, a medical agent prescribing a treatment to a cancerous patient. 
How could it take into account (and/or weigh) incommensurable aspects like the value of a human life and the cost of the treatment?

Now, the alignment between human and artificial values is possible only if we define a common space where a metric can be defined and used. This paper proposes extending the conventional Anything2vec approach to ethics, which has been successful in numerous similar and hard-to-quantify domains (ranging from natural language processing to recommendation systems and graph analysis).

This talk proposes a way to map an automatic agent decision-making (or control law) strategy to a multivariate vector representation, which can be used to compare and assess the alignment with human values. The rationale is that if an automatic agent implements a decision-making strategy, this strategy is optimal with respect to some loss function. At the same time, if the human accepts adhering to the agent strategy,  this implicitly means that such an agent strategy is also optimal with respect to a weighted sum of human criteria. By making such an assumption, it is possible to recover some constraints on the weights of the human criteria that the adoption of the agent strategy implies.

The Ethics2Vec method is first introduced in the case of an automatic agent performing binary decision-making. Then, a vectorisation of an automatic control law (like in the case of a self-driving car)  is discussed to show how the approach can be extended to automatic control settings.

EOS workshop

Registration

Program

8h30 : Yevhen Havrylenko, Lausanne University(45 min)
Synthetic tabular data for ratemaking: deep generative models vs imputation-based methods

Actuarial ratemaking depends on high-quality data, yet access to such data is often limited by the cost of obtaining new data, privacy concerns, and competitive sensitivity. In this talk, we explore synthetic-data generation as a potential solution to these issues. In addition to generative methods previously studied in the actuarial literature, we explore and benchmark fundamentally different class of approaches, which is based on Multiple Imputation by Chained Equations (MICE).

In a comparative study using an open-source dataset, we evaluate MICE-based models against other generative models like Variational Autoencoders and Conditional Tabular Generative Adversarial Networks. We assess each model on three dimensions. First, we assess data fidelity, namely how well the synthetic data preserves marginal distributions and multivariate covariate relationships. Second, we evaluate data utility, measuring the consistency of Generalized Linear Models (GLMs) trained on synthetic versus real data. Third, we assess the ease of use of each method by analyzing its implementation complexity, and the need for customization and finetuning. Furthermore, we investigate the impact of generically augmenting the original data with synthetic data on the performance of GLMs for predicting claim counts. Our results highlight the potential of MICE-based methods in creating high-fidelity tabular data while offering lower implementation complexity compared to deep generative models.

This research talk is based on a joint work with Meelis Käärik and Artur Tuttar, both affiliated with the University of Tartu. The pre-print version of our paper is available at https://arxiv.org/abs/2509.02171

9h15 : Paul Wilsens, KULeuven  (30 min)
Claim dynamics of weather-related insurance claims

Weather events account for a substantial share of property-related insurance claims, making careful modelling of such claims crucial for prudent portfolio management. Highly granular open-source weather-related data are increasingly available at short notice, enabling new opportunities for improved claim development and reserving. In this paper, we provide a methodology to handle the claim dynamics of weather-related property insurance claims, i.e. the occurrence and reporting processes, and apply it to a real insurance portfolio. Specifically, we apply an EM-XGBoost framework, incorporating publicly available weather-related covariates such as precipitation and wind speed, enriched with date- and policyholder-specific information. Weather covariates are used directly to model the occurrence of claims, as well as indirectly to construct different reporting regimes for claims under varying weather conditions. Using data from multiple public sources, we assess the use of observational versus reanalysis weather data for both claim occurrence and reporting. Additionally, we examine the impact of temporal and spatial aggregation on model accuracy for IBNR reserving. This allows us to identify aggregation levels that still yield reliable estimates of occurrence and reporting dates while providing insights at both the individual and portfolio levels.

9h45 : José Miguel Flores-Contro, UCLouvain (30 min)
Linear Risk-Sharing of Losses at Occurrence Time in the Compound Poisson Surplus Model

Denuit and Robert (2023) study a conditional-mean risk-sharing scheme within the classical Cramér–Lundberg risk process, where losses are allocated among pool participants at the time of occurrence. They show that this scheme reduces the infinite-time ruin probability of each participant, highlighting the benefits of risk sharing. However, the conditional-mean approach is not easily interpretable and may be difficult to implement in practice, particularly in contexts where transparency is essential, such as low-income communities using community-based insurance. Motivated by these considerations, we propose a linear risk-sharing rule in which each participant covers a fixed fraction of losses incurred by the pool. We show that when individual losses belong to the same scale-family or share a common mean, pooling under this rule also reduces the infinite-time ruin probability for each participant. This provides a more intuitive and practical risk-mitigation mechanism. Numerical examples illustrate and support our findings.

coffee break (30 min)

10h45 : Mathias Lindholm, Stockholm University (45 min)
Discrimination, fairness and how this can be measured 

We will start by going through different notions of fairness in relation to proxy-discrimination. This includes both classical definitions from the algorithmic fairness literature, but also more recent contributions more directly targeting the insurance pricing setting. A natural follow up question relating to defining fairness and discrimination is how this should be measured. Apart from presenting recent contributions on this we will also discuss ongoing work in this direction.

11h30 : Freek Holvoet, KULeuven (30 min)
Deep learning methods for loss simulation in a multi-peril insurance portfolio

In a portfolio where each policyholder is exposed to losses under multiple perils, calculating a value-at-risk requires an understanding of the dependency structure among perils. In this work, we investigate deep learning architectures to model the joint distribution of multi-peril losses. We explore two frameworks: a direct estimation approach using conditional variational autoencoders, and a flexible two-step approach that decouples the marginal distributions from the dependency structure. For the marginals, we utilize neural networks based on the Tweedie compound Poisson distribution, paying special attention to the mixed discrete-continuous nature of insurance losses and the probability mass at zero. For the dependency structure, we investigate a normalizing flow model called RealNVP, which allows for exact density estimation via invertible transformations of pseudo-residuals. We validate these methodologies using synthetic data to benchmark their ability to recover the underlying joint density and aggregate risk measures against known ground truth.

12h00 : Charlotte Jamotton, UCLouvain (30 min)
A Multi-Criteria Fair Gaussian Regressor for Insurance Premium

This article studies how multiple notions of fairness can be incorporated into a single Bayesian non-parametric regression framework for insurance pricing, with a focus on claim frequency modeling under a log-link. We consider a Generalized Gaussian Process Regression (GGPR) model for count data with risk exposure and introduce fairness interventions in its architecture. Specifically, we address notions of individual fairness by altering the kernel structure to control the similarity between policies (e.g., to mitigate omitted variable bias). We also address group-level fairness by enforcing demographic parity through linear constraints affecting the posterior. This modified GGPR architecture allows us to jointly enforce multiple fairness definitions, spanning both group and individual-level criteria, within a single probabilistic model. We empirically explore trade-offs with actuarial fairness, and how different fairness criteria interact when combined. The results highlight the importance of adopting a multi-criteria, context-aware approach to fairness in insurance pricing.

Dimitri Konen 

(University of Cambridge) 

Will give a presentation on : 

Inference in high-dimensional nonlinear and non-Gaussian models

Abstract :  
We consider the problem of performing statistical inference on the states of a nonlinear dynamical system, whose dynamics are governed by a nonlinear PDE, and corrupted by noise. These states further depend on an unknown initial condition which we model as a Gaussian random field, resulting in a priori random trajectories. Given discrete measurements of the system corrupted by generic (non-necessarily Gaussian) noise, one then wishes to update the prior trajectory to the best posterior update on the states of the dynamical system. This can be regarded as a problem of Bayesian inference in an infinite-dimensional regression context where the regression function is a solution to the PDE governing the dynamics and where the posterior distribution arises from a Gaussian process prior on the initial condition. The focus will be on so-called `dissipative’ systems for which, from a PDE perspective, inference on the system is expected to be poorly achieved. We will explain, however, why this is in fact good news for the statistician and that posterior trajectories concentrate at root(N)-rate around the ground truth trajectory of the system in a strong (supremum) sense and are asymptotically Gaussian with optimal covariance given by the Fisher information of the model. We will then explain how this allows one to construct uniform-in-time-and-space confidence bands for the true trajectory, and will briefly discuss computational hardness of running an MCMC in this setting.

13/03/2026 - 14:30 - ISBA C115 - 
 

Raphael Huser 

(King Abdullah University of Science and Technology) 

Will give a presentation on : 

Neural Parameter Estimation with Incomplete Spatial Data

Abstract: 
Advancements in artificial intelligence (AI) and deep learning have led to neural networks being used to generate lightning-speed answers to complex questions, to paint like Monet, or to write like Proust. Leveraging their computational speed and flexibility, neural networks are also being used to facilitate fast, likelihood-free statistical inference. However, it is not straightforward to use neural networks with data that for various reasons are incomplete, which precludes their use in many applications. In the first part of the talk, I will describe a novel neural Bayes inference approach that is based on the Monte Carlo expectation-maximization (EM) algorithm. The proposed neural EM approach is likelihood-free, substantially faster than the conventional EM algorithm when conditional simulation from the model is feasible (as it does not require numerical optimization at each iteration), and more statistically efficient than a simpler alternative “masking approach”. The proposed inference approach will be illustrated using simulated data and an application to Arctic sea-ice data, analyzed using a hidden Potts model with an intractable likelihood function. If time allows, in the second part of the talk, I will then explain how score-based diffusion models can be used to perform fast and accurate conditional simulation from spatial models when this is theoretically infeasible or computationally expensive, such as with spatial extremes models. 

Antje Christensen

(Project Director at Novo Nordisk)

Will give a presentation on : 

How to solve real life problems with small data, and bring your decision makers along on the journey

As in any other process based industry, issues and opportunities arise from process parameters and process variation in the pharmaceutical industry. Data are abundant, but sometimes it is the thinly populated regions of the data space that drive our understanding of the process and its issues forward.

We start by exploring an example of an issue that was identified in a pharmaceutical API production process. Different types of data, both observational and experimental, were used to solve the issue and anchor the solution. We will explore the specific example, and generalise to a problem solving flow with special focus on solution testing.

In the second part of the talk, we will explore typical questions that decision makers may have to the analyst at the various stages of the problem solving flow. We will discuss their background in decision analysis, and how you can address them.

The session is intended to be interactive, and I am looking forward to your perspectives and suggestions for other paths that could have been explored, both in problem solving and communication.

https://teams.microsoft.com/meet/36517061766449?p=lvKdmC262fg1lGIhQZ

Christian Ritter

(UCLouvain)

Will give a presentation on : 

See and Show (2026 Edition)

This is a new instance of a seminar given about once a year about perception and visualization related to statistics.  This seminar is evolving slowly. If you attended the last edition you may not learn many new things. The new topics include dashboard design and the ability of AI tools to interpret charts and tables.

Charts and tables are some of the main ways with which we communicate statistical information. This communication is therefore essential if we want to make effective contributions. Making effective charts and tables requires (conscious or intuitive) knowledge of perception and of the workings of the mind.

In this talk, I start with a few examples of charts and tables which we will discuss together. This leads to the first chapter on perception and mental processing. The second short chapter will summarize basic rules of effective communication. On this basis, we will develop useful principles and techniques for reading, preparing, and improving visual communication in statistics. 

Teams : https://teams.microsoft.com/meet/39296103347533?p=7eMWGV4rzLQItYSFJb

François Portier 

(ENSAI, CREST) 

Will give a short course on :  

Measure concentration and non-asymptotic statistics

Tuesday May, 12th, 2026, 9h30-12h30 and 13h30-15h (Room C115)
Wednesday May, 13th, 2026, 9h30-12h30 and 13h30-15h (Room C115)

May we kindly ask you to register for the event at the following link:
https://forms.office.com/pages/responsepage.aspx?id=1JCwei76z068fEEntNWC7A0x3GIiKy5JqzLWN70hq_1UMzhNWE42V01QWERNT1lFWVI3SUs4WTdHRS4u&route=shorturl 
 

Chair International Francqui Professorship 

Prof. Ian McKeague | Lecture : "The Tukey-Dyer method"

Abstract:  
As well as coining the word software, John Tukey famously remarked that the best thing about being a statistician is that you get to play in everyone’s backyard.  Tukey's point was that being a statistician gives the freedom to pursue an interest in almost any field.  In this two-part talk, I will argue for what I call the Tukey-Dyer method: the inspired choice of "a slight corner on something" as the starting point for creative work, even when you are not an expert.  The first part of the talk will discuss how the method can be useful in searching for novel ideas in statistics, whether A.I. can play a role, and how the method has helped start various projects I have been involved with over my career.  The second part of the talk will focus on recent work in the areas of functional data analysis and post-selection inference in which the Tukey-Dyer method has had an influence.

REGISTRATION   (deadline for registration April 5th)

When: 10:00-12:30 on April 16th.

Where: ‘Doyen’ Building (Auditorium Doyen 21), Place des Doyens, 1348 Ottignies-Louvain-la-Neuve. The easiest access is by entering via Place Rabelais. The room is on the 2nd floor.

More details about associated events are available at: https://feb.kuleuven.be/research/decision-sciences-and-information-management/events/FrancquiMcKeague

For all other information and specific requests, please reach out to: nancy.guillaume@uclouvain.be - eugen.pircalabelu@uclouvain.be 

For the local organizing committee at UCLouvain,
Prof. Catherine Legrand
Prof. Eugen Pircalabelu

Wayne Jones

(Principal Statistician, Shell UK)

Will give a presentation on : 

GroundWater Spatiotemporal Data Analysis Tool (GWSDAT).

GWSDAT is an open‑source, user‑friendly tool for visualising and interpreting groundwater monitoring data in space and time, helping practitioners move beyond spreadsheets to statistically robust insight. The presentation shows how GWSDAT supports better decisions across the project lifecycle—early leak detection, trend assessment, monitoring optimisation, and evidence‑based site closure—using real case studies. I’ll also highlight why GWSDAT has become a globally recognised standard, bridging advanced statistics and practical decision‑making for regulators and industry alike.

Link to software: https://gwsdat.net/home/ 

Online with transmission to C-115, Teams link: https://teams.microsoft.com/meet/32438786122672?p=Lh594kGgOqtZqemjC7

Bart Meuleman

(ReSPOND – Department of Sociology, KU Leuven)

Will give a presentation on : 

Measurement equivalence: Between Statistical Dogmatism and Anything Goes

Debates about measurement equivalence have recently intensified across disciplines that rely on cross-group or cross-cultural comparisons. Critics argue that the widespread use of measurement invariance testing has become overly restrictive, discouraging comparative research and imposing unrealistic statistical requirements. Some even call for a paradigm shift away from invariance testing altogether. In this talk, I revisit these critiques and argue that the debate is often framed as a false dichotomy between strict statistical dogmatism and an “anything goes” approach to comparison.

Therefore, I return to the conceptual foundations of (comparative) measurement and make the auxiliary measurement assumptions of invariance testing clear. Based on this, I argue that many of these criticisms rest on misunderstandings or overlook existing methodological developments. At the same time, invariance testing should not be applied mechanically, nor should failed tests automatically preclude comparative analysis. Finally, I outline two productive ways forward: treating deviations from invariance as potentially meaningful information about cross-group differences, and evaluating how consequential such deviations are for the substantive comparisons of interest.
 

16:00 in person in room C-115 with possibility to follow online, teams link: https://teams.microsoft.com/meet/35752674800961?p=6fHZS4j0mZhjCx5OVz 

EXALT Workshop - Learning spatio-temporal climate extremes

WEBSITE : https://exalt-project.github.io/workshop-2026/

Welcome to the EXALT project!

The EXtreme weather Attribution at mid- and high-Latitudes using advanced statistical Techniques (EXALT) project is an interdisciplinary effort that involves three research institutes: the Namur Institute for Complex Systems (naXys) at UNamur, the Earth and Climate Center (part of ELI, the Earth and Life Institute) at UCLouvain, and the institute of Statistics, Biostatistics and Actuarial Sciences (part of LIDAM, the Louvain Institute of Data Analysis and Modeling in economics and statistics) at UCLouvain. The PhD grants are financed through the Action de Recherche Concertée (ARC) called EXALT - EXtreme weather Attribution at mid- and high-Latitudes using advanced statistical Techniques, under the supervision of Profs. Francesco Ragone, François Massonnet, Johan Segers and Anna Kiriliouk (UCLouvain).

The overall objective of EXALT is to introduce a novel comprehensive framework for extreme event attribution and to apply it to the most influential events in the mid-high latitudes, namely heatwaves and compound drought-heat events over Europe, floods in southern and northern Europe, and sea ice extent over the Antarctic. Our attribution framework will (i) introduce non-stationarity in the extreme-value models, (ii) be developed with a multivariate mindset, and (iii) deliver reliable statistics thanks to a meticulous choice of study region (via clustering methods) and climate data (augmented through rare event algorithms). Our proposed methods aim to better constrain the probabilities of occurrence of past and future extreme events and to arrive at robust attribution statements that do not understate (nor overstate) the further risks that society will face. 

Registration

Registration is free but mandatory; you can register by sending an email to anna.kiriliouk@uclouvain.be before May 1st. Please note that places are limited and will be assigned on a first come first serve basis.

                   

Nicolas Schtickzelle

(UCLouvain)

Will give a presentation on : 

P‑values and effect sizes: why “simple” statistical quantities aren’t always simple

P‑values and effect sizes are among the most widely used—and sometimes misused—quantities in statistical analysis. The p‑value often appears as the sole metric in weak statistical reports, while effect sizes are increasingly promoted to complement or replace p‑values by quantifying the magnitude of an effect rather than its mere statistical significance. Although these concepts seem straightforward to scientists with a solid statistical background, their practical implementation can reveal surprising subtleties.

In this seminar, we explore two situations where the gap between the conceptual definitions of these quantities and their real‑world computation becomes especially important. First, we examine how common statistical software computes p‑values and how these implementations can quietly diverge from textbook definitions—sometimes with the intention of reducing misinterpretation, but at the cost of obscuring key conceptual ideas. Second, we investigate how the treatment of “absent” predictors in linear models influences effect‑size estimation in multimodel averaging, leading to potentially large differences depending on the chosen parameterization and interpretation.

Through these two examples, the seminar highlights why even basic statistical quantities deserve a closer look, and how understanding the nuances behind them can improve both statistical practice and critical interpretation.

Teams Link: https://teams.microsoft.com/meet/349902929081522?p=DQKvIriuXSAtcyWEcn

Benoit Berac

(European Investment Bank)

Will give a presentation on : 

Introduction to the Implementation and Conservative Adjustments in the Estimation of Regulatory PD in Low Default Portfolios

This presentation provides a brief introduction to how regulatory probability of default (PD) is estimated for low default portfolios, with a focus on the practical implementation of internal ratings, their mapping to PD curves, and the application of conservative adjustments. Emphasis will be placed on prudential requirements, the associated model risk, and calculation practices rather than theoretical modelling, with examples drawn from regulation and wholesale exposures in banks.

About the speaker:  Benoit Berac received his master's degree in statistics at UCLouvain in 2014. He has 20 years of experience in banking regulation, specializing in the implementation of CRR calculation projects and the integration of risk model outputs within banking computation systems.

Teams Link: https://teams.microsoft.com/meet/34910478870806?p=vyE6a6ikzLyEbDsMdW 

IMAL Workshop - Imperfect Data : From Mathematical Foundations to Applications in Life Sciences.

The workshop will be held at Martin’s Agora City Resort in Louvain-la-Neuve (Rue de l’Hocaille 1-3, 1348 Louvain-la-Neuve) and a tentative schedule, as well as the list of speakers and title of their presentation is listed below:
 

9h30  Welcome words/Introduction

9h40  - 10h20 Rajen Shah (University of Cambridge) : "Hunt and test for assessing the fit of semiparametric regression models".
10h20 - 11h00 Lise Léonard (UCLouvain) : "Statistical Inference and Model Averaging in High-Dimensional Regression".

Pause

11h20 - 12h00 Gaëtan Louvet (ULB/UNamur) : "Robustness of Tukey depth median"
12h00 - 12h40 Jeong Min Jeon (Seoul National University) :  "Local Fréchet regression with spherical predictors".

Lunch break

14h00 - 14h40 Cristian Preda (Université Lille) : "Categorical functional data analysis with application to clustering patient trajectories after discharge from an acute geriatric unit".
14h40 - 15h20 Hortense Doms (UCLouvain) : "Joint modeling of longitudinal HRQoL data accounting for the risk of competing dropouts".

Pause

15h40 - 16h20 Agathe Guilloux (INRIA) : "Dynamic survival prediction with time-dependent covariates observed at irregular frequency".
16h20 - 17h00 Negera Wakgari Deresa (KU Leuven) : "On an extension of the Cox model for time-dependent covariates under dependent censoring with unknown association".

17h00  Closing

17h10  Drinks and social time

Registration to the event (morning and/or afternoon sessions, lunch and/or drinks) is free, but mandatory using the form : IMAL Workshop – Remplir le formulaire . (The deadline for registration is May 20th at 17pm.)
Parking : For external participants travelling by car, a QR entry code will be provided for parking nearby. Please let us know if you need one.

More information about the ARC IMAL project is available here.
For further inquiries please reach out to:
Eugen Pircalabelu (eugen.pircalabelu@uclouvain.be)
Nancy Guillaume (nancy.guillaume@uclouvain.be)

Looking forward to welcoming you all at the event!
On behalf of the IMAL team,
Prof. Catherine Legrand
Prof. Germain Van Bever
Prof. Eugen Pircalabelu

Eugen Pircalabelu

(UCLouvain)

Will give a presentation on : 

An exploratory analysis for unraveling landmarks and understanding the human spinal cord anatomy

I will present in this talk a case-study that involves spinal cord measurements. Understanding

the segmental anatomy of the human spinal cord remains challenging due to its inter-individual variability and the lack of methodological consistency across previous studies. Although recent imaging studies enhanced the segmental localization, they do not cover the entire length of the spinal cord and require expert guidance. In this cadaveric study, my collaborators’ team dissected 40 spinal cords and measured the length, thickness, and width of 29 segments (C2 toS5).

With the help of basic tools from multivariate analysis (such as Procustes shape analysis and

Distance correlation) and functional data analysis we try to model and understand characteristics of the human spinal cord. Our analysis highlights reliable anatomical landmarks, in particular the L4 and C6 segments while (mildly) enhancing our understanding of the spinal cord anatomy and unveiling some aspects of its complexity.

Teams Link: https://teams.microsoft.com/meet/31784337248719?p=QijzzNOiBllAau3gQG 

Erwin Dreesen

(KULeuven)

Will give a presentation on : 

Why Precision Dosing Fails (and How Statistics Could Fix It)

Model-informed precision dosing uses statistical models to individualize therapy, but real-world performance often suffers from model mismatch, prior miscalibration, and overconfidence in single-model forecasts. This talk shows how to make MIPD more reliable using practical tools from applied statistics: controlling prior informativeness (including prior flattening), robust Bayesian updating, and multimodel/ensemble approaches to address model uncertainty. I will illustrate these ideas with case studies across different disease areas and drugs, and I will discuss how to validate performance using calibration and prediction-focused metrics that matter in clinical practice.

Teams Link: https://teams.microsoft.com/meet/325074107501257?p=T7gDATOIXe5s01DMsU 

Kris Sankaran 

(University of Wisconsin - Madison) 

Will give a presentation on : 

New Diagnostics for Dimensionality Reduction of Genomic Data  

Abstract: 
Dimensionality reduction helps organize high-dimensional genomics data into manageable low-dimensional representations, like differentiation trajectories in single cell data and community profiles in metagenomic sequencing. Such reductions are powerful but sensitive to hyperparameters and prone to misinterpretation. This talk addresses two risks in dimensionality reduction. First, we consider how to choose the number of topics K in topic modeling, a method from population genetics and language modeling, now common in microbiome analysis. While broadly useful, topic models require users to specify the number of topics K, which governs the resolution of learned topics. We discuss a new technique, topic alignment, for comparing topics across models with different resolutions. Simulation studies show that this approach distinguishes between true and spurious topics. Second, we examine the distortions introduced by nonlinear dimensionality reduction methods, like t-SNE and UMAP, when applied to single cell data. For example, these methods can introduce spurious clusters and fail to preserve cell density. We adopt the RMetric algorithm from manifold learning to measure local distortions. We also develop visualizations to explore these distortions. Representing cells as deformed ellipses highlights changes in local geometry, and an interactive interface selectively reveals evidence for distortion without overwhelming the viewer. Through case studies on simulated and real data, we find that the visualizations can flag fragmented neighborhoods, support hyperparameter tuning, and enable method selection. Topic alignment and distortion visualization are available as software packages, with case studies in online vignettes: https://go.wisc.edu/7h58r9, https://go.wisc.edu/ss5ts9.

Kris Sankaran 

(University of Wisconsin - Madison) 

Will give a presentation on : 

A Brief Introduction to Shapley Values  

Abstract: 

This workshop will introduce Shapley values, a game-theoretic concept adapted to explain predictions from black-box machine learning models. We study the method from statistical and geometric perspectives, then apply language from causal inference to discuss subtleties in its interpretation. Since naive implementations become computationally intractable as samples or features grow, we review approximation methods like IME and KernelSHAP. We close with practical implementations in R (shapviz) and Python (shap), motivated by interpretable machine learning problems from social science and developmental biology. Material will be accessible to researchers with experience in linear regression and supervised machine learning. The lecture is based off notes for a recently developed undergraduate course on interpretable machine learning (https://github.com/krisrs1128/stat479_notes).