Introduction

During this short course, we will introduce a platform, which explores advanced causal inference strategies designed for complex clinical trial efficacy analyses, addressing key challenges such as imperfect randomization, death truncation, missing data, surrogate outcomes, and real-world data with unmeasured confounders. The platform has successfully demonstrated in hospitals nationwide, exemplified how modern technology bridges statistical theory and practice, enhanced research precision and supporting informed decision-making in both regulatory and clinical contexts, and facilitated several high-quality publications.

Short Course: July 4, 2025, 13:00 - 17:00

Short Course Instructors

Xiao-Hua Zhou, Distinguished Chair Professor, Peking University
Haoxuan Li, Ph.D. Candidate, Peking University
Chunyuan Zheng, Ph.D. Student, Peking University

Short Course Outline

● The Background of Causality
● Mathematical Foundation for Causal Inference
● Randomized Experiments
● Violations to Randomization
  ■ Non-compliance
  ■ Missing Data
  ■ Truncation by Death
  ■ Other Types of Intercurrent Events
● The Identification and Estimation of Causal Estimands
  ■ Complier Average Causal Effect
  ■ Defier Average Causal Effect
  ■ Survivor Average Causal Effect
● Unmeasured Confounders
● Instrumental Variables
● Application Example of Causality: Recommender Systems and Large Model

● Causality for Large Models
  ■ Hallucination in Large Models – From a Spurious Correlation Perspective
    ◆ Single Modality
    ◆ Multiple Modality
    ◆ Understanding and Mitigating Hallucination
  ■ Understanding and Enhancing Training Pipeline of Large Models
  ■ Causality and Safety, Bias, and Explainability in Large Models

● Large Models for Causal Discovery
  ■ Review of Causal Discovery Algorithms
  ■ Large Models as Knowledge-Based Methods
    ◆ What Can Large Models Do?
    ◆ Query-Based Pairwise Causal Edge Inference
  ■ Large Models Assist Traditional Causal Discovery Pipelines
    ◆ Pre-Discovery: Ordering & Extracting Hidden Variables
    ◆ Post-Discovery: Orientation
  ■ Hybrid Large Models and Traditional Causal Discovery Pipelines
  ■ Benchmark

● Large Models for Causal Inference
  ■ Evaluation and Benchmark for Causal Inference
  ■ Relaxing Important Causal Inference Assumptions with Large Models
    ◆ Positivity Assumption
    ◆ Stable Unit Treatment Value Assumption
    ◆ Unconfoundedness Assumption

● Causal Inference Methods and Software for Randomized and Real-World Data
  ■ Randomized Control Trail
    ◆ Perfect Randomized Control Trail
    ◆ Trail with Compliance
    ◆ Trail with Truncation
    ◆ Trail with Missing Outcome
  ■ Observational Study
    ◆ Study with Unmeasured Confounders
    ◆ Study with Truncation