International Day Of Women in Statistics and Data Science

This session will welcome attendees, discuss the history of the IDWSDS Conference, and you’ll hear from the representatives from the Caucus for Women in Statistics & Data Science (CWS), International Statistics Institute (ISI) and other professional associations. They will share how their organizations help their members create global connections. We will also share technical information about how to access key features of the Sched App and our Zoom rooms.

Generative data augmentation is widely used to address class imbalance by enriching minority classes with synthetic samples. Existing approaches typically employ a fixed augmentation strength across all classes, ignoring differences in class imbalance, data structure, and generative quality. We propose an adaptive augmentation framework that determines class-specific augmentation strengths using both class proportions and structural information measured through Gini correlation. This strategy allocates more synthetic data to underrepresented and weakly structured classes while limiting augmentation for well-represented classes.

We develop a theoretical framework showing that the resulting excess risk is controlled by a weighted combination of class-conditional Wasserstein discrepancies and Gini-based structural factors. We further establish consistency results and demonstrate that adaptive augmentation provides tighter control of risk distortion than fixed augmentation schemes.

Experiments on imbalanced classification datasets show consistent improvements in minority-class recall and macro-F1 performance, while empirical results closely match theoretical predictions. The proposed framework provides a principled, structure-aware foundation for generative data augmentation.

Statistics provides tools for answering questions that are central to building trustworthy AI: What works? What fails? How certain are we? In this talk, I will discuss my path from studying statistics to conducting research as a data scientist and now working full time in AI safety. I will share examples of how statistical thinking can support this work, including deciding what to measure, uncovering failure modes, quantifying uncertainty, and connecting evidence to real decisions. I hope to show that many paths lead into this field and that statisticians and data scientists have important skills to bring to the challenge of building trustworthy AI.

This session explores how artificial intelligence and data science are driving the next generation of precision medicine, drawing on large-scale biobanks, electronic health records (EHRs), and genomic data from both academia and industry. Dr. Marie Loh (Nanyang Technological University / Genome Institute of Singapore) will discuss translational opportunities from HELIOS-SG100K, a multi-ethnic Asian population cohort with rich phenotypic and molecular data, with examples in cardiometabolic disease and atopic dermatitis. Ms. Chih-Ting Yang (Vanderbilt University) will use the All of Us Research Program to show how biobank data can be translated into longitudinal disease staging, illustrated through a study of cardiovascular-kidney-metabolic syndrome. Dr. Sarah Lotspeich (Wake Forest University) will present an LLM-enhanced, ICD-10-based algorithm for recovering missing EHR data, demonstrating accuracy comparable to expert chart review at scale. Rounding out the session, Dr. Christine Hsiung will bring an industry perspective from the genetics sector, sharing applied research insights from Taiwan Biobank genomic data and translational pipelines. Together, these talks span academic and industry approaches, genomic profiling, longitudinal disease modeling, and AI-driven data quality methods, illustrating how data science is reshaping precision medicine across diverse populations and settings.

An Invited Session Proposal from Statistics Society of Australia (SSA) Women in Statistics Network Supported by the The International Statistical Institute (ISI) Committee on Women in Statistics

Session Title: Breaking barriers

Session Organiser: Alysha De Livera

Session Chair: Ayse Aysin Bilgin

The progression of the talks:

curiosity → identity → contribution → legacy
aspiration → exploration → contribution → impact

Speakers:
1. Jayamini Liyanage, [email protected], La Trobe University, Australia, (PhD Student)
2. Melissa Middleton, PhD, GStat, [email protected], Murdoch Children’s Research Institute, Australia, (Early Career)
3. Alysha De Livera, PhD, AStat, [email protected], Latrobe University, Australia, (Mid Career)
4. Ayse Aysin Bilgin, PhD, [email protected], Macquarie University, Australia Professor (Late Career)

As artificial intelligence becomes deeply integrated into healthcare, addressing inherent biases and ensuring equitable outcomes is paramount. This session explores multidisciplinary strategies for developing inclusive and fair medical AI systems. The panel brings together four experts to discuss critical challenges and actionable solutions. Dr. Heajin Kim (Chair) will highlight the importance of reflecting sex and gender characteristics in medical AI to foster equity. Researcher Jeongyeon Kim will provide a critical analysis of data bias issues within Korea's AI Hub. Dr. Yuseong Chu will review current research trends and methodologies for achieving algorithmic fairness in healthcare. Finally, Dr. Wonjung Park will introduce innovative approaches utilizing synthetic data generation to address dataset imbalances. Together, this session offers a comprehensive roadmap towards trustworthy and inclusive medical AI.

Refining a Two-Stage Pipeline for Gene-by-Environment Interaction Discovery: Random-Forest Screening Followed by SOIL-Based Prioritization | Mapping Inequality: Women Statisticians and the Settlement House Movement | Precision by Design: A Transferable Framework for Sampling Optimisation in Hierarchical Capture-Recapture|How Biostatistics Research Experience Enriches the Pre-Medical Journey

From its earliest beginnings, Bayesian statistics has been a synthesis of theory, methodology, computation and application. In this presentation, I will reflect on the development of the award-winning Australian Cancer Atlas (https://atlas.cancer.org.au/) and spotlight its Bayesian foundations. I will highlight some of the challenges and proposed solutions to modelling and visualisation of an awkward spatial geography, “filling in” missing covariates, and communicating uncertainty. I will also touch on new research directions inspired by the Atlas: new spatio-temporal models, spatial vulnerability indices, meta-analysis transfer learning, distributed AI and responsible data science. Importantly, these methodological discussions will be complemented by reflections on the impact of the work for patients, health practitioners, cancer support groups and government agencies. Bayesian statistics really can make a difference!
Selected References

Baade P, K Mengersen [2024] Building HOPE through the Australian Cancer Atlas Insight+ MJA 35  //insightplus.mja.com.au/2024/35

J Bon, A Bretherton, K Buchhorn, S Cramb, C Drovandi, C Hassan, A Jenner, H Mayfield, J. McGree, K Mengersen, A Price, R Salomone, E Santos-Fernandez, J Vercelloni & X Wang, [2023] Being Bayesian in the 2020s: opportunities and challenges in the practice of modern applied Bayesian statistics. Philosophical Transactions. Series A, Mathematical, physical, and engineering sciences, 381(2247), Article number: 20220156.

Bretherton A, Bon J, Warne D, Mengersen K, Drovandi C, [2026] A Principled Approach to Bayesian Transfer Learning, Bayesian Analysis. To appear.

Cramb SM, K Mengersen, and PD Baade. [2011] Developing the atlas of cancer in Queensland: methodological issues. International Journal of Health Geographics, 10, p 1-11, 2011

Goodwin S,  T Saunders, J Aitken, P Baade, U Chandrasiri, D Cook, S Cramb, E Duncan, S Kobakian, J. Roberts, K. Mengersen, [2024] Designing the Australian Cancer Atlas: visualizing geostatistical model uncertainty for multiple audiences.  JAMIA. Journal of the American Medical Informatics Association, //doi.org/10.1093/jamia/ocae212

Hassan, C [2024] Structured Models and Algorithms for Sensitive Data. PhD Thesis, Queensland University of Technology, Australia.

Hogg J, J Cameron, S Cramb, P Baade, K Mengersen [2024]  A Two‐stage Bayesian Small Area Estimation Approach for Proportions.  International Statistical Review. V92 I3 455482.

Hyland-Wood B, Snoswell A, Sandeep R, Chun O, Perrin D, Fielt E, Price A, Mengersen K (2024) Response to proposals paper on introducing mandatory guardrails for AI in highrisk settings. Analysis and Policy Observatory, 2024/10/4

Jahan F, Duncan E, Cramb S, Baade P, Mengersen K, [2020] Multivariate Bayesian metaanalysis: Joint modelling of multiple cancer types using summary statistics, International Journal of Health Geographics, 19 (1)

Leontyeva Y, Y Huang, S Cramb, J Cameron, P Baade, K Mengersen, et al. [2025] Bayesian Spatial Relative Survival Model to Estimate the Loss in Life Expectancy and Crude Probability of Death for Cancer Patients, Statistics in Medicine 44 (3–4), e10287

Price A, M Rigby, P Fiévez, K Mengersen [2025] A spatial vulnerability index for environmental health Ecological Indicators Vol 178, September 2025, 113793 

Neonatal mortality remains a major global public health challenge, with an estimated 6,200 newborns dying daily, mostly in settings where patient records are scarce. For data-scarce neonatal units, a key question arises: when transition data is limited, does it matter whether Analytical Estimation or Monte Carlo Simulation is used to model patient outcomes?

This study addresses that question using a neonatal dataset of 6,000 daily state transitions across 2,000 patients in Ghana. An absorbing Markov chain with four states: Hospital Admission, Neonatal Intensive Care Unit (NICU), Recovered, and Death was used. Both methods were evaluated across six data levels with 500 independent replications using bias, variance, standard deviation, and mean squared error.

Both methods perform similarly with adequate data and degrade equivalently under data scarcity because they share the same estimated transition matrix. Expected time to absorption is more sensitive to limited data than absorption probabilities, and 500 observed transitions emerge as the minimum reliable threshold. These findings provide an evidence-based data standard for resource-constrained healthcare systems.

Keywords: Absorbing Markov Chain, Monte Carlo Simulation, Analytical Estimation, Absorption Probability, Expected Time to Absorption, Sensitivity Analysis.

Authors: Emmanuella Frimpong, Dr. Irene Kafui Vorsah Amponsah, PhD (Visiting Lecturer at Ohio University)

The rapid deployment of machine learning in healthcare has exposed a fundamental tension between model performance and clinical utility. This talk addresses the urgent need for Explainability (XAI), moving beyond the "black box" paradigm to ensure that AI systems are not only accurate but also transparent, accountable, and trustworthy. In the high-stakes environment of medical diagnostics, understanding an algorithm’s pathway is essential for verifying clinical reliability and meeting the rigorous standards demanded by both clinicians and regulators.
A central theme of this talk is the transition from surface-level performance metrics to deep reproducibility and validation. We will examine how traditional accuracy scores can sometimes be deceptive and how to combat this, compare XAI with true interpretability, and tackle the challenge of the inheritance of bias. This session outlines proactive mitigation strategies, including rigorous data quality checking, subgroup analysis, and continuous fairness assessments throughout the model lifecycle.
Finally, the talk will navigate the evolving regulatory landscape, specifically the EU AI Act.

This session brings together recent advances in statistical methodology for analyzing complex, high-dimensional, and heterogeneous data arising in modern biomedical and genetic research. The presentations cover a broad range of topics, including variable screening for mixed-type and high-dimensional data, integration of mechanistic and statistical models for infectious disease prediction, semiparametric approaches for meta-analysis, and set-based association testing for functional genetic responses. Together, these talks highlight emerging statistical frameworks that improve flexibility, scalability, and interpretability in the analysis of complex data structures.

This session is organized by the European Survey Research Association – Special Interest Group on multi-source statistics.
As survey research increasingly operates in a complex data ecosystem, integrating probability-based surveys with administrative records, digital traces, non-probability samples, and other non-traditional data sources has become essential for improving inference and studying phenomena that cannot be adequately measured by a single data source. In particular, new data sources can provide better data about hard-to-reach populations and minorities, helping to make these groups more visible in research and evidence-based decision-making. This session will discuss cutting-edge methodologies and innovative case studies that integrate diverse data sources in survey research, with a particular focus on their potential to improve the measurement of underrepresented populations.

The STRATOS (STRengthening Analytical Thinking for Observational Studies) Initiative is an international collaboration of statistical and methodological experts dedicated to improving the design, conduct, analysis, and reporting of biomedical and observational research. Through accessible guidance documents, recommendations, and educational resources, STRATOS seeks to bridge methodological advances and statistical practice.
Presentations illustrate diverse approaches to improving research quality, including the development of international recommendations for analyzing patient-reported outcomes in oncology trials, innovative strategies for teaching Initial Data Analysis (IDA), consensus-based development of guidance for statistical analysis plans in observational studies, and principled approaches to defining estimands and interpreting quality-of-life outcomes. These examples demonstrate how statistical thinking can improve the transparency, reproducibility, relevance, and interpretability of scientific research.
Together, these talks demonstrate the impact of international collaboration, interdisciplinary research, translating complex statistical concepts into practical tools for researchers. The session also highlights the important leadership role of women statisticians in shaping methodological standards, educational initiatives, and research practices.

Medical decisions—ranging from diagnosis to treatment selection—are inherently uncertain. Clinicians often rely on heuristic, experience-driven processes to integrate heterogeneous data. In this context, Predicted Individual Treatment Effects (PITE) offer a principled statistical framework to quantify how much a specific patient benefits from one treatment over another.

Advances in computational systems now enable machines to identify complex patterns within large datasets, facilitating a shift toward data-driven, individualized healthcare. This work explores using PITE to support clinical decision-making across diverse diseases and contexts. We address critical questions: which AI methods suit specific clinical datasets, how PITE should be validated, and how outcome complexity affects tool reliability.

We demonstrate PITE-based models in various disease settings, each posing unique methodological challenges. Our results show that even under real-world conditions—such as missing data—predictive models maintain interpretability and generate estimates that support clinicians. Notably, our findings highlight that internal validation is insufficient; external validation is essential for robust predictions.

Ultimately, effective PITE-based support requires more than modeling. It demands an adaptive, continuously learning system integrating data management, modeling strategies, regulatory-grade interpretability, and ongoing validation to translate evidence into precise.

Open science is pivotal in advancing medical research by promoting accessibility and collaboration among researchers globally, and biostatisticians play a critical role in supporting these efforts. Shared datasets and software code, particularly those developed using time-intensive algorithms, are fundamental for fostering replicability and improving research efficiency for other scientists. Open access publications accompanied by publicly available data and code also help reduce disparities in knowledge access by making research materials available to researchers who might otherwise lack access. This presentation will explore data sharing principles and experiences, highlighting two recent studies: one on smoking behaviors in the United States and another on the association between caregiving and psychological well-being among Florida college students. The resulting datasets were shared under specific terms of use via Harvard Dataverse to facilitate further medical research. In addition, a subset of the tobacco use data and sample code were shared through the Resources Portal of the Teaching of Statistics in the Health Sciences Section of the American Statistical Association to support the teaching of survey methods. The presentation is intended for students, researchers, and practitioners interested in data sharing and open science.

Classical statistical models are built upon an assumption of short-range dependence — an assumption that fails dramatically when confronted with the complexity of real-world aviation datasets. Such datasets routinely exhibit long-range memory, non-stationarity, and heavy-tailed distributions that render conventional approaches inadequate. In this work, we propose a novel fractional statistical framework that draws on advanced operator theory to directly address these challenges, offering both rigorous theoretical guarantees and compelling empirical improvements over established baselines.
We construct a family of Toeplitz-type estimators grounded in the theory of α-fractional Bergman spaces, establish their theoretical properties, validate the framework on a large-scale aviation dataset comprising over 500,000 UAE flight records, and demonstrate prediction error reductions of 23–31% over ARIMA and 14–18% over LSTM-based approaches.

Modern biomedical data science increasingly relies on datasets with many predictors, limited sample sizes, and complex correlation structures. In these settings, classical regression and standard variable-selection methods may lead to unstable models, overfitting, or conclusions that are difficult to interpret. Shrinkage and penalized estimation provide a principled framework for improving reliability, but their success depends on how multicollinearity and dependence among predictors are handled. This talk discusses reliable variable selection for biomedical data science, moving from classical shrinkage ideas to modern interpretable learning. The presentation will introduce the motivation behind ridge-type methods, LASSO-based procedures, elastic net estimation, and adaptive penalization, with emphasis on approaches that use correlation information to improve model stability and interpretability. Motivated by biomedical applications such as molecular subtype identification and classification problems, the talk will show how statistically grounded regularization can support both prediction and scientific interpretation. The broader message is that reliable biomedical data science requires methods that are not only accurate, but also stable, transparent, reproducible, and interpretable for domain experts.

Despite advances in data science and statistical training across Africa, a structural disconnect persists between the development of technical talent and its integration into national and regional policy processes. It is particularly pronounced for early-career African women statisticians, who face intersecting institutional barriers,UN Women and PARIS21 shows that women occupy only 15% of chief statistician positions in Sub-Saharan Africa, reflecting patriarchal institutional cultures, limited senior-level mentorship and rigid career progression pathways.

This abstract proposes an actionable policy framework to transition young African women statisticians from technical implementers to strategic policy influencers. Using a comparative case-study methodology through the Young African Statisticians Association network, we examine institutional barriers and entry pathways across selected National Statistical Offices in East and West Africa.

The framework advances three strategic pillars: institutional quotas and fast-track leadership pathways for young women in state-led data initiatives; coordinated mentorship systems linking global statisticians with local networks to institutionalize peer and senior sponsorship and gender-responsive data governance through dedicated advisory positions for young women statisticians in ministerial policy processes.

Embedding young women statisticians in Africa's governance is vital for equitable, evidence-based development.

Privacy-preserving record linkage (PPRL) has a rich methodological literature on encoding, matching, and cryptographic guarantees — but comparatively little guidance exists on what happens after the match: how do researchers actually access, query, and analyze the linked data that results? This presentation addresses that gap directly, walking through real-world architectures used by statistical agencies, such as the U.S. Census Bureau's linkage key infrastructure and UK Trusted Research Environments, to control researcher access to linked data. We’ll cover how to handle potential uncertainty when analyzing linked data, such as producing confidence tiers, adjusting thresholds, and a novel approach to clerical review within the PPRL system. We’ll also walk through concrete querying structures such as linkage maps and secure enclaves and will explore examples of queries and their outputs. Attendees will leave with a clearer picture of the access-and-analysis landscape for linked data, practical considerations for designing their own linkage projects, and a better sense of how privacy shapes analysis.

At a time when data drives decisions across science, industry, and society, the voices shaping this landscape matter more than ever. This invited session brings together distinguished women researchers whose work reflects the power of diversity, collaboration, and intellectual curiosity in statistics and data science.

Through their journeys, the speakers will illustrate how meaningful connections—across disciplines, institutions, and cultures—emerge not only from shared research goals but also from openness, mentorship, and resilience. Their experiences reveal how collaborative environments can spark innovation, amplify impact, and create pathways for future generations of women in the field.

More than a showcase of research, this session is an invitation to reflect on how we build scientific communities that are inclusive, dynamic, and forward-looking. By fostering dialogue and embracing diverse perspectives, we strengthen our collective ability to tackle complex challenges and shape a more connected and equitable future for data science.

This session celebrates the role of women as catalysts for change, bridging ideas and people to advance knowledge and inspire the next generation.

This session is sponsored by the Portuguese Statistical Society an by the Portuguese Country Representative of CWS.

This session presents novel computational and methodological advancements in modern statistical inference, focusing on complex data structures such as multivariate, longitudinal, and high-dimensional survival data. The presentations span both frequentist and Bayesian paradigms, featuring innovative strategies to overcome long-standing computational bottlenecks commonly encountered in such investigations. The session opens with a new framework for multivariate order-restricted statistical inference under a general grass ordering, applying a pairwise optimization criterion and introducing an efficient vector projection-based algorithm to estimate the order-restricted mean matrix. It then transitions into the Bayesian realm with two complementary presentations on longitudinal quantile regression. These back-to-back talks first present a robust partially collapsed Gibbs sampler to conquer convergence failures in high dimensions and then showcase a Bayesian adaptive LASSO framework that utilizes clustering-based shrinkage for simultaneous optimal quantile estimation and sparse variable selection. The session concludes by extending advanced continuous shrinkage priors to survival analysis, presenting a hybrid minorize-maximize algorithm for high-dimensional proportional hazards models using the horseshoe+ prior. Together, these talks highlight recent developments of highly efficient algorithms to extract reliable inferences from high-dimensional and complicated datasets.

This session brings together four applied studies demonstrating how modern statistical methods can address diverse scientific questions using complex data from Costa Rica. The presentations cover Bayesian hierarchical modeling, penalized regression, machine learning, and generalized additive models, with applications to infectious disease surveillance, biotechnology awareness, and animal production. Despite spanning different disciplines, the talks share a common emphasis on developing interpretable and robust statistical approaches to better understand complex systems through real-world applications in the Costa Rican context.

This invited session showcases recent advances in statistical methodology for analyzing increasingly complex, high-dimensional, and structured data arising in biomedical research, imaging, surveys, and longitudinal studies. The presentations introduce innovative approaches to sufficient dimension reduction, scalable spatial topic modeling for multiplexed imaging, Bayesian small area estimation under informative sampling, and Bayesian semiparametric quantile mixed-effects models for longitudinal data with missing responses. Together, these talks highlight modern statistical techniques that combine methodological rigor with computational efficiency to address emerging challenges in data science, while demonstrating broad applicability across diverse scientific disciplines.

The junior section of the International Society of Bayesian Statistics (j-ISBA) is excited to present a session showcasing the work of outstanding early-career women researchers from the United States, Italy, and Germany. This session highlights recent advances in Bayesian methodology for analyzing complex relational and structured data, with a focus on flexible probabilistic modeling, scalable computation, and impactful real-world applications.

The four talks highlight modern Bayesian approaches for the modeling of networks and high-dimensional relational data. They span generative models for graph-structured data grounded in classical Bayesian principles of exchangeability, Bayesian nonparametric models for interaction networks with evolving community structure, latent variable models that account for uncertainty and reporting bias in social network data, and hierarchical tensor factorization methods for large-scale sparse relational arrays. Across these contributions, speakers develop computationally efficient inferential strategies using tools including variational inference, hybrid Monte Carlo algorithms, and scalable latent variable methods. These approaches enable Bayesian analysis in settings involving millions of observations while maintaining interpretable probabilistic models.

Together, these talks demonstrate the innovative methodological, theoretical, and applied contributions of early-career women Bayesian statisticians.

This session highlights how modern biostatistics tackles a shared challenge across very different problems: making valid, efficient inference when the data we have is incomplete, limited, imbalanced, or imperfectly comparable to the question we want to answer. Talks in this session explore how statisticians "borrow strength" – from historical or external data sources, from related but heterogeneous datasets, from carefully constructed causal designs, and from generative models that learn the structure of missingness itself – to push beyond the limitations of any single dataset. Together, these four talks, spanning clinical trials, large-scale survey methodology, high-dimensional prediction, and observational causal inference, showcase the breadth of ideas at the intersection of biostatistics, machine learning, and causal reasoning, and the methodological creativity required to turn imperfect data into trustworthy conclusions.

Survival analysis plays a fundamental role in health research, especially when the objective is to understand not only the time until an event occurs, but also the possibility that a proportion of individuals may never experience that event. This
situation is particularly relevant in cancer studies, where long-term survivors may be considered cured or no longer susceptible to recurrence or death related to the disease.

In this talk, I will present statistical modeling approaches based on cure models, with emphasis on their application to uterine cancer data. The discussion is motivated by my work on regression modeling for cure factors using the reparametrized defective generalized Gompertz distribution. This framework allows the estimation of both survival behavior and cure proportions, providing a flexible tool for analyzing time-to-event data in the presence of a cured fraction.

The presentation will highlight how cure models can contribute to women’s health research by identifying factors associated with long-term survival and by offering interpretable measures for clinical and epidemiological studies. By connecting statistical theory, survival analysis, and real health data, this work illustrates how statistical science can support evidence-based understanding of cancer outcomes and strengthen global connections between data science and public health.

In our closing session we will celebrate the 24 hours we spent together in this conference! We will hear from you about sessions that were impactful and engaging, your “ah ha” moments, and lessons learned from our speakers.