alt text

Zijian Guo

Home
About Me
CV
Research
Publications
Research by Topics
Talks
Software
Teaching
Statistical Optimization
Teaching Evaluation
Group Members
Meeting
Meeting Reservation

Research Topics

My research develops statistical foundations for reliable and generalizable inference in modern data settings where classical assumptions often fail. Below is a directory of my main methodological areas and their applications.

Generalizable Multi-source Learning

Distributional Robustness · Minimax Optimization · Federated Learning

Making models reliable when data come from multiple heterogeneous studies, hospitals, or environments. Rather than assuming a single pooled distribution, I define robust targets using minimax objectives that guard against worst-case shifts, accompanied by efficient primal-dual algorithms and valid uncertainty quantification.

Reliable Causal Inference

Instrumental Variables (IV) · Robust Identification · Causal Invariance

Centering on instrumental variables (IV), I develop identification and inference tools that remain meaningful even when instruments are imperfect or invalid. My work also extends to under-identification, non-regular inference, and causal invariance learning to identify relationships that remain stable across different environments.

High-dimensional Uncertainty Quantification

High-dimensional Inference · Multiple Testing · Hidden Confounding

Turning high-dimensional modeling into trustworthy scientific conclusions. I develop inference-first toolkits—including confidence intervals, hypothesis tests, and multiple testing procedures—that remain valid in realistic regimes beyond simple linear models, addressing complications like hidden confounding and endogeneity.

Nonstandard Inference

Uncertainty Quantification · Post-selection Inference · Perturbation Methods

Providing reliable uncertainty quantification in modern problems where standard Wald intervals break down. I develop inference methods based on perturbation, resampling, and stability-driven constructions for scenarios involving optimization-defined targets, data-adaptive selection, or lack of standard regularity conditions.

Applications to Genetics and Health Data

Mendelian Randomization · Multi-center EHRs · Biomarker Discovery

Making large, heterogeneous datasets—such as genetics, proteomics, and multi-center EHRs—actionable for science. Examples include MR-SPI for robust biomarker discovery via valid genetic instruments selection, and SurvMaximin for transporting survival risk prediction models across hospitals without sharing patient-level data.