/tags/2026-fall/index.xml 2026 Fall - McGill Statistics Seminars
  • Singular Learning Theory for PPCA and Factor Analysis

    Date: 2026-11-20

    Time: 15:30-16:30 (Montreal time)

    Location: Online via Zoom

    https://mcgill.zoom.us/j/83483745684

    Meeting ID: 834 8374 5684

    Abstract:

    An age-old question in principal components analysis (PCA) is how many principal directions should be retained? Model-based approaches, like probabilistic PCA, provide one way of addressing this question. Information criteria can be used to treat the choice of the number of PCs as a model selection problem. However, there is a subtlety that is often overlooked: the PPCA model, as well as the related factor analysis model, both have singularities, which alter convergence rates that motivate information criteria like the BIC. In this talk, I will provide an overview of singular learning theory, an area that involves using ideas from algebraic geometry to tame the asymptotics of models with singularities. I will also introduce the singular BIC (sBIC) as an information criteria that is able to incorporate asymptotic information at singularities and can be used for model selection. The efficacy of the sBIC as compared to the BIC is demonstrated empirically and theoretically, where it is shown that in high-dimensional settings the sBIC requires a smaller eigengap in order to detect principal component directions.

  • A Latent Semiparametric Mixture Model for Robust Heterogeneity Learning in High Dimensions

    Date: 2026-11-18

    Time: 16:00-17:30 (Montreal time)

    Location: In person, Burnside 1104

    https://mcgill.zoom.us/j/83483745684

    Meeting ID: 834 8374 5684

    Abstract:

    Learning heterogeneity from high-dimensional data has become a fundamental task across many data-driven fields. Cluster analysis is a widely used unsupervised tool for this purpose, and many existing methods have been adapted to high-dimensional settings through sparsity constraints. In contrast to algorithmic approaches, model-based clustering offers clean theoretical guarantees and computational simplicity, but can be restrictive in practice due to its reliance on specific parametric distributions. We propose a novel latent semiparametric mixture model for clustering, in which the observed data follow a Gaussian mixture after latent coordinate-wise monotone transformations. Furthermore, assuming a sparse optimal clustering rule in the latent space, we develop a new method, CESME, tailored for high-dimensional clustering. By allowing unspecified transformations, our framework becomes substantially more flexible and robust than traditional model-based approaches. The clustering error of CESME is shown to approach the optimal clustering error at a rate matching that obtained when the true transformations are known.

  • Beyond Pixels: Functional Representations of Images for Statistical Analysis

    Date: 2026-11-13

    Time: 15:30-16:30 (Montreal time)

    Location: In person, Burnside 1104

    https://mcgill.zoom.us/j/83129965084

    Meeting ID: 831 2996 5084

    Abstract:

    Images are an increasingly important source of information across scientific disciplines, yet their statistical analysis remains challenging due to the high dimensionality and complex spatial structure of pixel-based representations. In this talk, we explore a functional perspective on image analysis, where images and the objects they contain are represented as continuous functions rather than discrete collections of pixels. We first introduce a framework for representing and aligning object contours, enabling statistical shape analysis and classification. We then extend this approach to multiple interacting contours, preserving their relative geometric relationships. Finally, we discuss the representation of image colors through smooth surfaces, modeled as multidimensional functions. Together, these approaches illustrate how images can be analyzed as collections of objects defined by their shapes, textures, and colors.

  • Causal Inference in Clinical Trials with Intercurrent Events: From Treatment Noncompliance to Additional Medication Use

    Date: 2026-11-06

    Time: 15:30-16:30 (Montreal time)

    Location: Online via Zoom

    https://mcgill.zoom.us/j/83483745684

    Meeting ID: 834 8374 5684

    Abstract:

    Randomized clinical trials are widely regarded as the gold standard for causal inference. However, intercurrent events (ICEs) occurring after treatment initiation can complicate both the definition and estimation of treatment effects. Common examples include treatment noncompliance, where the treatment received differs from the treatment assigned, and the use of additional medications.

    In the first part of this talk, I will focus on treatment noncompliance in HIV clinical trials. Following the principal stratum strategy, we define the target estimand within a latent subpopulation called always-compliers. We propose an instrumented hidden Markov model to characterize longitudinal compliance trajectories and event hazards conditional on the current compliance state. We establish nonparametric identification results under the proposed model, develop a maximum likelihood estimator, and extend the framework to accommodate two-stage sampling designs.

  • Applying DOE Principles to Statistical Software Testing

    Date: 2026-10-30

    Time: 15:30-16:30 (Montreal time)

    Location: In person, Burnside 1104

    https://mcgill.zoom.us/j/82510837324

    Meeting ID: 825 1083 7324

    Abstract:

    Testing statistical software presents unique challenges. The space of possible inputs, options, and workflows can be enormous, making exhaustive testing impractical. At the same time, software developers often have limited time and resources for formal testing activities.

    This talk explores how principles from Design of Experiments (DOE) can be applied to the software test case selection problem. By viewing test cases as experimental runs and software inputs as factors, DOE provides a structured framework for constructing efficient test suites that maximize information while minimizing effort. The discussion highlights connections between experimental design and software testing, demonstrating how statistical thinking can improve test coverage, identify defects more effectively, and increase confidence in software reliability. More broadly, the talk illustrates how core ideas from DOE can be successfully applied in nontraditional settings.

  • On Local Inference for Functionals of Ill-Posed Inverse Problems

    Date: 2026-10-23

    Time: 15:30-16:30 (Montreal time)

    Location: Online via Zoom

    https://mcgill.zoom.us/j/83173904340

    Meeting ID: 831 7390 4340

    Abstract:

    Targets of inference in nonparametric instrumental variables, proximal causal inference, and shadow-variable models are often continuous linear functionals of solutions to a pair of ill-posed inverse problems. Recent work has shown that, under minimal assumptions, ill-posedness does not preclude pointwise root-n inference. We show that pointwise guarantees need not translate into locally reliable inference. In a baseline model sufficient for a Neyman orthogonal estimating equation to exist, the target is Hellinger discontinuous, which rules out locally uniformly valid inference. A stronger model that uniformly bounds the minimum-norm solution to one of the inverse problems restores continuity. However, pathwise differentiability then holds only if that solution is unique at the true data-generating process. When it fails, Wald-type confidence intervals from standard procedures undercover asymptotically under some local alternatives. Pathwise differentiability is also restored, even without uniqueness, under a uniform bound on a beta-source norm. We relate these regimes to very weak, weak, and strong identification in linear IV models.

  • Avoiding double dipping through data thinning and data fission

    Date: 2026-09-25

    Time: 15:30-16:30 (Montreal time)

    Location: In person, Burnside 1104

    https://mcgill.zoom.us/j/89492447923

    Meeting ID: 894 9244 7923

    Abstract:

    While classical statistical methods are designed for testing hypotheses about pre-specified models, the reality of modern science is that analysts often explore their data before coming up with models and hypotheses of interest. We refer to the practice of using the same data to generate and then test a hypothesis as double dipping. Classical statistical hypothesis tests will fail to control the Type 1 error rate in settings that involve double dipping. Often, we avoid double dipping by splitting our observations into a training set and a test set. While this sample splitting approach is straightforward and easy to understand, it is generally inapplicable in unsupervised settings. Motivated by unsupervised problems that arise in the analysis of single-cell RNA sequencing data, we first propose data thinning, an alternative to sample splitting that splits each observation in a dataset into two independent pieces. We show that this method provides an elegant solution to our motivating problems under distributional assumptions. We then show how the data fission framework of Leiner et al. (2025) can be fruitfully applied to circumvent the distributional assumptions of data thinning. We conclude by discussing the promise and the challenges of data fission in the context of inference after clustering and inference after variable selection in logistic regression.