News and Events
All News
All Events
1. Department Seminar Series
2. Statistics Career Fair
3. Dissertation Defenses
4. Michael Woodroofe Lecture Series
5. Seminars in Other Departments
6. Archived Events
7. Search Events
8. Distinguished Alumni Speaker Series
Departmental Awards

Distinguished Alumni Speaker Series

News and Events
All News
All Events
1. Department Seminar Series
2. Statistics Career Fair
3. Dissertation Defenses
4. Michael Woodroofe Lecture Series
5. Seminars in Other Departments
6. Archived Events
7. Search Events
8. Distinguished Alumni Speaker Series
Departmental Awards

Past Distinguished Alumni Speakers

Daniel Almirall - Associate Professor, Department of Statistics, Institute for Social Research, University of Michigan - March 21, 2025

On the design and analysis of sequentially-randomized trials for clustered adaptive interventions

An adaptive intervention is a sequence of decision rules that guide the provision of intervention at critical decision points based on the evolving needs of individuals. In many health, education and policy settings, adaptive interventions target clusters of individuals (e.g., clinics or schools), with the intent of improving outcomes at the level of individuals. These are called clustered adaptive interventions (cAIs). Clustered, sequential, multiple-assignment, randomized trials (cSMART) can be used to form high-quality clustered adaptive interventions. This non-technical talk introduces clustered adaptive interventions and cSMARTs, and presents current and past research concerning the design and analysis of cSMARTs. This includes recent work with University of Michigan students on finite-sample adjustment methods for the marginal mean comparison of cAIs (with W. Pan), an extension of moderator analyses using ideas from Q-learning (with Y. Song), and a three-level analysis method (with G. Durham)—i.e., repeated measures nested within individuals, nested within clusters. It also includes a new cSMART design for the construction of event-triggered cAIs (with M. Ferlic). Time-permitting, the talk ends by describing current challenges in the design and analysis of multilevel SMARTs, a special type of clustered SMART, whereby nested groups of individuals are randomized sequentially (e.g., a clinic-level randomization followed by a doctor-within-clinic-level randomization) leading to interesting questions concerning causal spillover. Throughout, I draw on multiple examples from implementation science, a relatively new area that concerns the optimization and evaluation of clustered interventions to improve the adoption of evidence-based practices.

Eric Laber - James B. Duke Distinguished Professor, Research Professor of Global Health, Departments of Biostatistics, Bioinformatics and Statistical Science, Duke University - March 22, 2024

Reinforcement Learning for Respondent-Driven Sampling

Respondent-driven sampling (RDS) is a network-based sampling strategy used to study hidden populations for which no sampling frame is available. In each epoch of an RDS study, the current wave of study participants are incentivized to recruit the next wave through their social connections. The success and efficiency of RDS can depend critically on attributes of incentives and the underlying (latent) network structure. We propose a reinforcement learning-based adaptive RDS design to optimize some measure of study utility, e.g., efficiency, treatment dissemination, reach, etc. Our design is based on a branching process approximation to the RDS process, however, our proposed post-study inferential procedures apply to general network models even when the network is not fully identified. Simulation experiments show that the proposed design provides substantial gains in efficiency over static and two-step RDS procedures.

Link to Institution/Research Page

Yevgeniya (Jane) Pinelis - Chief of AI Assurance, Joint Artificial Center United States Department of Defence - October 7, 2022

Being a statistician in the Department of Defense: my journey to AI Assurance

Rigorous independent and objective assessments of AI-enabled systems in the DoD are a necessary step in ensuring that our systems are fielded and used responsibly. However, standards, technologies and processes for performing test and evaluation (T&E) on AI-enabled systems remain largely a research endeavor. This talk will provide an overview of challenges and progress in AI Assurance to facilitate speedy and agile deployment of reliable and effective AI/ML capabilities to the warfighter. It will highlight the latest advancements in understanding the fundamental issues of designing safe and secure of AI systems and in how AI systems should be tested and assured as a means of shaping appropriate use. Additionally, Dr. Pinelis will talk about her work since graduation, which has mainly been in the areas of defense and national security.

Link to Institution/Research Page

Dave Hunter - Professor of Statistics at Penn State University - April 15, 2022

When the Supreme Court declares your CSCAR work unconstitutional: Reflections on data science and society

The 2003 U.S. Supreme Court case known as Gratz v. Bollinger addresses a formula for college admissions that was created by a graduate student in statistics at the University of Michigan for a specific purpose. The method used to create this formula is considered simplistic by modern machine learning standards; yet the debate that ensued, which could not have happened if a more modern approach had been used, illustrates that science and society do not always benefit from machine learning models that achieve the best possible predictive performance. This talk discusses the history of the legal case, the admissions formula and how it was created, and the implications of the debate for how we build predictive models.

Link to Institution/Research Page

Yizao Wang - Associate Professor, Department of Mathematical Sciences, University of Cincinnati -March 19, 2021

Long-range Clustering of Extremes

In this talk I will go over a few recent examples of stationary sequences, of which the extremes form long-range clustering. The phenomena of clustering of extremes have been extensively investigated in the literature since 1980s. However, for most examples the extremal clustering occurs only locally. That is, the locations of extremes within each cluster are bounded and shrink to a single point at the macroscopic level after normalization. For long-range clustering, on the other hand, the locations of extremes within each cluster are unbounded, and they can be further characterized by a random closed set in the scaling limit.

There are two classes of models that recently have been shown to exhibit the phenomena of long-range clustering of extremes. One is the Karlin model, which this talk will focus on. The other is the so-called stable-regenerative model, of which if time permits I will highlight briefly some key features.

Based on joint works with Olivier Durieu (Université de Tours, France) and Gennady Samorodnitsky (Cornell University, USA).

Link to Institution/Research Page

Earl Lawrence - Group Leader, Statistical Sciences at Los Alamos National Laboratory -February 14, 2020

Computer Experiments at Los Alamos National Laboratory: Life on Mars and Really Big Computers

Inference with computationally expensive physics models is a big part of statistics at Los Alamos National Laboratory. The first part of that talk will cover some well-known background on the statistical approach computer experiments. This will take place in the context of ongoing work for ChemCam, an instrument on the Mars rover Curiosity whose goal is to determine whether Mars ever had conditions that could have supported microbial life. ChemCam uses laser-induced breakdown spectroscopy to analyze the chemical composition of Martian soil and rocks. Our goal is to use the resulting spectra and a LANL-developed predictive model to estimate the elemental abundances from surface samples. The second part of the talk will cover new work to address computer experiments from exascale supercomputers. The next generation of supercomputers are expected to have I/O limitations relative to their computing ability: they will simulate more than they can save. This requires changes to our usual post-hoc analysis scheme. To address this, we are developing approaches to in situ statistical inference, statistical modeling that gets done inside simulations as they are running. Our early work considers modeling extremes for climate and space weather.

Link to Institution/Research Page