Jane-Ling Wang is Distinguished Research Professor in Statistics at the University of California, Davis. She obtained her PhD in Statistics from the University of California, Berkeley, in 1982 under the guidance of Jack Kiefer and Lucien Le Cam, and joined the Department of Statistics and Actuarial Science at the University of Iowa as an Assistant Professor in 1982. From there she moved to Davis in 1984 as an Assistant Professor. Except for a one-year appointment in 1987–88 at the Wharton School of the University of Pennsylvania, she has been a faculty member at Davis ever since.
Her research interests include survival analysis, functional data analysis and machine learning, where her emphasis is nonparametric and semiparametric approaches. She also enjoys collaborations with domain scientists and has long-term collaborations with biologists working to study aging and longevity, and with neuroscientists on brain imaging. Recently, she became interested in research on child development and, being an avid hiker (though not a climber) herself, in the analysis of climbing data for Mount Everest.
Jane-Ling is a Fellow of the American Association for the Advancement of Science (since 2011) and was elected an Academician of Academia Sinica in 2022. In 2016, she received the Noether Senior Research Award and in 2020 the Humboldt Research Award. She has served as the co-editor of Statistica Sinica (2002–05) and the Journal of American Statistical Association, Theory and Methods (2020–22), and since 2015, serves as a statistics editor for Science.
This 2026 IMS Grace Wahba Award Lecture will be given at the Joint Statistical Meetings (in Boston, August 1–6, 2026).
Statistical Learning for Complex Data
Contemporary statistical research is increasingly influenced by developments in AI. Many statisticians are integrating AI-driven methodologies into their research. In this lecture, Jane-Ling will briefly relate her own journey from more traditional nonparametrics to AI-influenced statistical research in the areas of functional data analysis and survival analysis and will reflect on challenges and opportunities.
A key feature of functional data is their infinite-dimensional nature, which presents challenges when harnessing neural networks and incorporating functional covariates in statistical models. A key question is how to effectively encode functional data inputs for deep neural networks (DNNs). Established methods implement dimension reduction via pre-selected basis expansions, which may be suboptimal. In recent work, Jane-Ling and co-authors proposed an adaptive architecture incorporating a basis-layer, in which hidden units act as data-driven basis functions constructed via micro neural networks. This end-to-end approach enables targeted representation of relevant features information, improving dimension reduction and outperforming other DNNs in classification and regression tasks.
A limitation of this approach is that it requires either fully or intensively observed functional data on a common time grid that is the same across all subjects. Consequently, it cannot accommodate traditional longitudinal data that are characterized by sparse and irregular observations. Such data are the subject of an active research area under the rubric of “Sparse Functional Data”. A major challenge is the irregular nature of such data, which means that they cannot be represented as vectors. It turns out that transformers, originally developed for sequence modeling in large language models, provide a natural solution for this challenge due to their ability to accommodate sequences of varying length. Jane-Ling will discuss her previous and current work on how transformers can be employed for effective imputation of irregularly and sparsely observed functional data and nonparametric regression for such sparse functional data.
Survival data present a different set of challenges, primarily due to censoring and other forms of incompleteness, which complicate both theoretical analysis and algorithmic implementation.
Jane-Ling will discuss how these challenges can be addressed, enabling procedures such as conformal prediction and hypothesis testing for complex black-box models.