Time: Sunday, June 14, 1:30 PM – 5:30 PM

Location: FUSE 1327

Format: In-person

Instructors: Xin Tu, UC San Diego; Jinyuan Liu, Vanderbilt University

Category: Methodology

Target Audience: All levels of (bio)statisticians and data scientists are welcome. The course covers both fundamental and more advanced topics in semiparametric models, accompanied by diverse real-world applications.

Prerequisites: Knowledge of statistical inference and a basic understanding of large sample theory.

Computer and Software Requirements: Basic knowledge of R programming.

Course Description

This short course gives biostatisticians and data scientists an engaging overview of semiparametric modeling through real-world applications with complex structures, such as high-throughput sequencing and network data. Both classical and cutting-edge semiparametric techniques are explored, highlighting their roles in balancing robustness, flexibility, and efficiency with minimal assumptions.

The foundation of statistical inference relies on models with explicit or implicit assumptions about the underlying data-generating process. Often, these models are characterized by finite-dimensional parameters and have only limited robustness in practice. This has motivated the advancement of semiparametric modeling, which blends finite-dimensional parameters of interest with infinite-dimensional nuisance parameters. Such flexibility has led to emerging applications in many research disciplines, especially in causal inference, missing data, survival, and survey studies.

This short course is divided into two halves. The first half introduces the fundamental concepts of semiparametric models and outlines their roles in robust inference with and without missing data. The second half discusses recent advances and applications, including settings that scale up to high-dimensional microbiome data and HIV viral genetic linkage networks, while also scaling down to inference problems involving outliers and small sample sizes.

Teaching Plan / Outline

• This half-day short course is broken into two sessions.
• First half (1 hour 45 minutes): Basic concepts of semiparametric models, including asymptotic linearity and efficiency, influence functions, Neyman orthogonality, U-statistics, and their asymptotic properties.
• Building on this foundation, the course emphasizes recent advances and applications in the second half, including the use of Neyman orthogonality for missing data problems, leading to a class of augmented inverse probability weighted (AIPW) estimators.
• Causal inference and feature aggregation in modeling high-dimensional data: applications to the human microbiome (45 minutes): This section discusses causal inference with both high-dimensional responses and confounders, introduces a feature aggregation framework that bypasses regularization bias, and uses efficient influence functions at multiple levels to characterize causal effects for high-dimensional data.
• Semiparametric modeling of HIV viral genetic linkage networks and highly skewed data (30 minutes): This section presents applications of semiparametric models to highly skewed data with outliers and to linkage probability estimation in network analysis, using functional response models for between-subject attributes.
• Inference under small and moderate sample sizes (30 minutes): This section covers recent developments in Monte Carlo simulation-based score tests, with a focus on sign-flipping score tests that perform well for both parametric and semiparametric models in small samples.

Instructor Bio(s)

Dr. Xin Tu (UC San Diego) is a Professor of Biostatistics at the Herbert Wertheim School of Public Health and Human Longevity Science at UC San Diego. He is also Co-Director of the UCSD Clinical and Translational Research Institute Biostatistics, Epidemiology and Research Design Core, as well as the Stein Institute for Research on Aging Biostatistics Core. He has 30 years of teaching experience and has coauthored over 320 peer-reviewed publications, along with two textbooks and two edited volumes, in areas including U-statistics, categorical data analysis, clinical trials, and social network analysis.

His methodological research spans semiparametric models for longitudinal data with informative missing follow-up, causal inference, and high-throughput data. He has served as lead biostatistician for many studies involving longitudinal data and complex modeling challenges, including doubly robust estimators, structural equation models, and structural mean models.

Dr. Jinyuan Liu (Vanderbilt University) is an Assistant Professor of Biostatistics and Psychiatry & Behavioral Sciences at Vanderbilt University. Her research focuses on effective dimension reduction and efficient integrative modeling of high-dimensional data arising from microbiome studies, imaging, wearable devices, behavior, and psychiatric science, with an emphasis on deriving causal and mediation insights from such complex data.