2026
Student Seminar: Ruochen Wang
2:00–2:30 pm Jones 111
Wednesday, November 4, 2026, at 2:00 PM, in Jones 111, 5747 S. Ellis Avenue
Master’s Thesis Presentation
Ruochen Wang, Department of Statistics, The University of Chicago
“Training Window Length and Partial Pooling for Prescription Forecasting: A Longitudinal Count Study of Empagliflozin Prescribing in England, 2018–2026”
Statistics Colloquium: Michael Sobel
11:30 am–12:30 pm Jones 303
Michael Sobel
Department of Statistics
Columbia University
Title: Measuring Public Institutions and their Agents: Disparities and Agreement, with Application to Asylum Decisions in the U.S. Immigration Courts
Abstract: Public institutions derive their legitimacy from the professionalism and impartiality of their agents. In order to maintain standards of conduct, the performance of an agent is often compared to the performance of her peers. But public institutions typically assign agents to different tasks, and as a result, naive comparisons may reflect the difficulty of the assignment and not an agent’s. To address this concern, we develop measures of disparity and agreement that compare an agent’s performance with the potential performance of randomly selected peers who could have been assigned the same set of tasks. We then give realistic conditions under which disparity is identified from the marginal distribution of each agent’s actions. We also show how agreement, which is generally unidentified when agents are assigned to different tasks, can be bounded or estimated under sensitivity assumptions about the dependence between potential outcomes. Finally, we use a nonparametric Bayesian model to examine applications for asylum filed between 2007 and 2014 in the widely studied and often maligned U .S. immigration courts. Our results do not support the picture of widespread partiality asserted in previous empirical work. At both the judge and court level, we find limited disparity and substantial potential agreement.
Students Seminar: Yating Liu
10:00 am–12:00 pm Jones 111
Tuesday, October 27, 2026, at 10:00 AM, in Jones 111, 5747 S. Ellis Avenue
Dissertation Defense Presentation
Yating Liu, Department of Statistics, The University of Chicago
“Learning from Text to Biological Data: The Trustworthy Inference using Robust Latent Structure”
Statistics Colloquium: Jing Lei
11:30 am–12:30 pm Jones 303
Jing Lei
Department of Statistics & Data Science
Carnegie Mellon University
Title: Evaluating Black-Box Classifiers via Stable Adaptive Two-Sample Inference
Abstract: Black-box machine learning methods predict well but are hard to validate. This talk develops adaptive, sample-efficient inference that uses flexible machine learning tools to test black-box predictions, with conditional distribution testing as the unifying problem. First, a weighted conformal rank-sum statistic turns aggregated conformal p-values into a test whose target, the Neyman-Pearson-Gini divergence, connects rank-sum statistics, ROC curves, and total variation. Second, two-sample U-statistic orthogonalization and cross-fitting remove the bias from estimating density-ratio weights, so the data can be reused without sample splitting. Third, for validating a fitted classifier, a cross-fitted distinguisher with a “safe-side bias” yields valid inference even though the statistic targets a data-dependent lower bound of the divergence. Applications show that classifiers with similar accuracy can differ substantially in goodness of fit.
Statistics Colloquium: Peter McCullagh
11:30 am–12:30 pm Jones 303
Peter McCullagh
Department of Statistics
University of Chicago
Title: Causal Inference Versus Stochastic Models
Abstract: The classical statistical model is a stochastic process, which means a probability distribution on the space of outcomes. Sampling distributions and distributional properties of derived summaries follow from the process. Logical positivism in the sense of von Mises reverses the natural order, beginning with an outcome variable or kollektiv, from which probabilities are defined as limiting frequencies. Causal inference is based on the same frequency definition supplemented by random sampling. These notions of probability are in conflict, even when limiting frequencies exist. Accordingly, all derived concepts such as expectation, independence, random variable, and treatment effect are also in conflict. Broadly speaking, the frequentist concept is adequate for sample surveys and other areas of application such as auditing, where the population is concrete, finite, and accessible for simple random sampling. These conditions are not met in typical scientific research work, where the population is abstract, infinite and only partially accessible. The emphasis in this talk is on stochastic regression models, where random sampling does not arise, treatment assignment is associated with a Markov kernel, and the treatment effect is a group action on the parameter space.
Statistics Colloquium: Kaizheng Wang
11:30 am–12:30 pm Jones 303
Kaizheng Wang
Department of Industrial Engineering and Operations Research/Data Science Institute
Columbia University
Title: Learning to Augment Statistical Inference with Generative Models
Abstract: Generative models can produce seemingly unlimited synthetic data, yet discrepancies between synthetic and real populations can introduce bias and undermine statistical conclusions. For a new task with scarce or no real data, how much synthetic data can be safely used? This talk develops a general framework that leverages historical tasks to calibrate synthetic-data augmentation and quantify the resulting uncertainty. When no real observations are available, the framework adaptively selects the synthetic sample size to achieve nominal average-case coverage. In the application to LLM-generated survey responses, this calibrated size can be interpreted as the number of human respondents the LLM is effectively worth, providing a measure of its simulation fidelity. More generally, when real observations are available, the augmentation scheme is characterized jointly by the number of synthetic observations and the weight assigned to each. We learn a size–weight frontier that identifies the largest synthetic contribution compatible with reliable uncertainty quantification. Together, these results show how generative models can strengthen statistical analysis when real data are limited.
DSI/Statistics Colloquium: Frederic Koehler
11:30 am–12:30 pm DSI 105
Frederic Koehler
DSI/Department of Statistics
University of Chicago
Title: A Least-Squares Perspective on Ellipsoid Fitting
Abstract: Suppose we have n i.i.d. random vectors X_1,\ldots, X_n in d dimensions. How large can n be, as a function of d, such that there exists an ellipsoid interpolating all n points? Saunderson, Parrilo, and Willsky conjectured an explicit answer (n ~ d^2/4) for Gaussian data. Subsequently, many authors made rigorous progress on this question. We study a natural least-squares generalization of ellipsoid fitting, over a general class of data distributions, and solve its high-dimensional limit—- this proves the Gaussian conjecture as a special case. The solution combines in a nice way ideas from high-dimensional probability, statistical learning, and convex geometry + optimization.
Based on a joint work with Youngtak Sohn.
DSI/Statistics Colloquium: Nikolaos Ignatiadis
11:30 am–12:30 pm DSI 105
Nikolaos Ignatiadis
DSI/Department of Statistics
University of Chicago
Title: Borrowing Strength in Multiple Testing with Compound p-values
Abstract: Many large-scale multiple testing problems in high-throughput biology involve thousands of units (e.g., genes or proteins) but only a small number of replicates per unit. A key observation in the microarray literature around the turn of the century was that standard t-tests may be nearly powerless in this setting. This motivated the development of methods such as the B-statistic (Lönnstedt and Speed, 2002), Significance Analysis of Microarrays (SAM; Tusher, Tibshirani and Chu, 2001), and limma (Smyth, 2004), which borrow information across units.
We study this type of borrowing across units through the lens of compound p-values. Unlike bona fide p-values, compound p-values need not be valid for each hypothesis separately, but instead satisfy an average superuniformity property across the null hypotheses. We first explain how compound p-values arise both from a nonparametric version of limma and from SAM. We then use the compound p-value perspective to construct a procedure that achieves the asymptotic power of the oracle Bayes rule in a sparse regime under the hierarchical model underlying the B-statistic and provides finite-sample frequentist control of the false discovery rate without requiring this hierarchical model to hold.
Statistics Colloquium: Xinran Li
11:30 am–12:30 pm Jones 303
Xinran Li
Department of Statistics
University of Chicago
Title: Randomization Inference with Sample Attrition
Abstract: Sample attrition, or equivalently missing outcomes, is a common challenge in randomized controlled experiments, and analyses that simply discard units with missing outcomes may suffer from severe bias. In this talk, we develop computationally efficient methods for randomization inference that remain valid under a broad class of potentially informative missingness mechanisms, allowing a unit’s missingness to depend on its potential outcomes. We pay particular attention to monotone missingness mechanisms, under which outcomes are systematically more likely to be observed under treatment than under control, or vice versa. Such restrictions arise naturally in many empirical settings, especially in the social sciences.
Specifically, we construct valid p-values for testing sharp, bounded, and quantile null hypotheses on treatment effects through a worst-case consideration of the classical Fisher randomization test. By leveraging distribution-free test statistics, these worst-case p-values can be computed efficiently and often admit closed-form solutions. Importantly, by incorporating both potential outcomes and potential missingness indicators into the test statistic, our methods can exploit structural assumptions such as monotone missingness to sharpen inference. Our approach also connects to a range of partial-identification bounds in the literature. We illustrate the proposed methods through simulation studies and empirical applications evaluating the impacts of job training, media censorship, and artificial intelligence.
Student Seminar: Yinjie Wang
4:30–5:00 pm Jones 111
Thursday, July 23, 2026, at 4:30 PM, in Jones 111, 5747 S. Ellis Avenue
Master’s Thesis Presentation
Yinjie Wang, Department of Statistics, The University of Chicago
“Simple and Effective Query-Adaptive Sparse Attention”