Causal Inference with Modern Machine Learning Methods
Doctoral-level introduction to causal inference at the intersection of machine learning, focusing on theoretical foundations and recent research developments.
Course Description
Across disciplines, causal inference is a cornerstone of science, engineering, economics, and public policy. In medicine, we would like to understand how a patient might have responded if we offered a different treatment. In engineering system design and optimization, we would like to understand how the system would behave if we made different design choices. In public policy, we are constantly asking if different taxes, laws, regulations, or programs might improve (or hurt) society at large. Correctly answering such questions can help us make more informed and better decisions. Data are frequently collected and analyzed in the process of seeking statistically quantified answers to these questions. In contemporary applications, these data are decidedly large-scale, complex, and high-dimensional. These call for an urgent need in designing modern statistical/machine learning methods for causal inference.
Recently, an exciting set of tools at the intersection of causal inference and machine learning has emerged to tackle these types of questions in these settings. This course is a doctoral-level introduction to these tools. Our emphasis is primarily on studying the statistical/machine learning tools to formally analyze these types of methods, with the goal of empowering students to have an in-depth understanding of cutting-edge research in this area and contribute their own new (theoretically justified) methods to the field.
The first 2.5 weeks of the course will focus on basic concepts and methods in causal inference at the level of the Imbens and Rubin (2015) book. For the rest of the semester, we will move to recent developments for causal inference using modern machine learning methods and the content will primarily be based on recent research papers.
Learning Objectives
Upon successful completion of this course, students will be able to:
- Explain the core concepts and challenges in causal inference under the potential outcomes framework;
- Apply the most recent developments of modern statistical/machine learning methods to some core causal inference problems;
- Demonstrate and improve the ability to develop and justify the statistical/machine learning methods with mathematical rigor when applying/adapting them to causal inference questions.
Related Posts
Project Updates
-
Causal Inference Overview
Understanding SUTVA and ignorability, fundamental assumptions in causal inference, what they mean, and what happens when they are violated.
-
Causal Assumption: SUTVA and Ignorability
Understanding SUTVA and ignorability, fundamental assumptions in causal inference, what they mean, and what happens when they are violated.
-
Causal Inference vs Causal Estimation
A deep dive into confidence intervals—what they are, how to interpret them, some caveats, and oft-encountered issues in online experimentation.
-
Violation of SUTVA in A/B Testing: network interference
A summary of the Lyft Engineering blog post 'Interference Across a Network' detailing how naive A/B testing can bias effect estimates in ridesharing.
-
Switchback Experiments
An overview of switchback (time-split) experiments: what they are, why they are used to solve network interference, and their trade-offs.
-
Synthetic Control and Experimentation Culture
When standard experiments fail: utilizing Synthetic Controls, managing experimentation culture, and understanding various treatment effects.
-
Case Study: Causal Effect of ETA Reduction
A practical case study on measuring the causal effect of reducing Estimated Time of Arrival (ETA) in a ridesharing marketplace.
-
Case Study: Causal Effect of ETA Reduction
A practical case study on measuring the causal effect of reducing Estimated Time of Arrival (ETA) in a ridesharing marketplace.
-
Paper review: Statistical Challenges in Online Controlled Experiments
A review of Larsen et al. (2024) on the statistical landscape and challenges of A/B testing in large-scale online environments.
-
Case Study: Diagnosing and Addressing a Metric Drop
An end-to-end framework for investigating MAU drops, targeting at-risk users, and making data-driven 'ship' decisions.
-
Quick and Dirty Sample Size Calculation
-
A/B Testing Metrics
A comprehensive guide to selecting and evaluating metrics in A/B testing and online experimentation.
-
Summary: Challenges in Experimentation (Lyft)
A summary of the Lyft Engineering blog post 'Challenges in Experimentation' by John Kirn.
-
Paper review: Performance Guarantees for Individualized Treatment Rules
A review of Qian and Murphy (2011) on formulating individualized treatment rules via conditional outcome maximization with performance guarantees.