Welcome to MS&E 228 – Applied Causal Inference Powered by ML and AI

Instructor: Vasilis Syrgkanis, Assistant Professor, MS&E
Units: 3
Winter Quarter 2023
Tue, Thu 1:30-2:50PM McCullough 115

Description:
The course will cover fundamentals of modern applied causal inference. Basic principles of causal inference and machine learning and how the two can be combined in practice to deliver causal insights and policy implications in real world datasets, allowing for high-dimensionality and flexible estimation. Lectures will provide foundations of these new methodologies and the course assignments will involve real world data and synthetic data analysis based on these methodologies.

Prerequisites: Basic knowledge of probability and statistics. Recommended: MS&E 226 or equivalent.

Office Hours: (Starting Week 2)

Format

The course will consist of lectures and homework assignments. Lectures will be held in person and will not be recorded. The lectures will cover fundamentals that fuse classical structural equation models (SEMs) and DAGs, with tools for statistical inference based on machine learning (lasso, random forest, deep neural networks) to infer causal parameters and quantify uncertainty. Grading will be primarily based on the weekly homework assignments and secondarily on class participation. There will be a total of 7-8 homeworks, rolled out roughly on a weekly basis, that will involve either mathematical proofs or coding exercises.

Grading

• Homework 90%
• Participation 10%

Course webpages

• Discussion and homework material: canvas
• Submissions: gradescope
• Example Code and Notebooks: Github

Problem sets

There will be a total of 7 or 8 problem sets. Problem sets must be submitted online through Gradescope gradescope. Problem set sheets will also be posted on Canvas.

Depending on their length, the total number of points in each set might vary. Each part of each problem will be graded as follows:

• You will receive zero points if you do not attempt it.
• If you attempt it, but there are either substantial methodological errors or major conceptual misunderstandings of the material, you will receive 2 points.
• If you attempt it and there are no substantial methodological errors or major conceptual misunderstandings of the material, you will receive 3 points.

We expect students who make a reasonable effort, even if not perfect, will receive 3 points. Receiving 2 points is intended to be a sign of significant comprehension issues and should only happen in cases where you haven’t been able to keep up with the material. Since you can get full credit without having been perfect, you should make sure to read the solutions to aid your comprehension of the material.

We encourage discussion of assignments with other classmates, but everybody must turn in their own written solutions in their own words. If you do a substantial subset of the work on your problem set with others, you must document on each assignment the other students that you worked with. You may work in pairs on coding portions of the assignments, but again, please make sure to clarify in your submission who you collaborated with; if you turn in the same code as someone else, without noting the collaboration, it will be considered an Honor Code violation. (See also the section on the Honor Code below.)

Computation

Many of the problem sets, as well as the guided project, will require you to be comfortable carrying out computations on data. Note that the “official” course language will be python: this is what we use for our problem sets, in our lectures and discussion sections, and the language we support in office hours.

We will provide links to python on the website, as well as any datasets that are needed through the course of the quarter. You can find notebooks that will be used throughout the class in python and R in this github repo: https://github.com/CausalAIBook/MetricsMLNotebooks.

As noted above, you are welcome to work in pairs on any computational component in the class, and on the project. In addition, you are welcome to use another language (e.g., R), for the project, but note that the course staff is only responsible to provide technical support for python.

Late day policy

To accommodate unforeseen challenges that may arise during the quarter, you have three late days for the problem sets. Each late day allows you to turn in an assignment up to 24 hours late. (Any fraction of a late day counts as one late day.) You may use multiple late days on the same problem set. Work submitted beyond the allowed late days will not receive credit.

Please note that we have provided the late day policy to help provide flexibility to you in managing your course load during the quarter. If circumstances arise that require further accommodations, we encourage you to contact your academic advisor as well as the Office of Accessible Education (see below) to help make appropriate arrangements. Out of fairness to all students, in the absence of an OAE Academic Accommodation Letter, we will generally be unable to provide accommodations beyond the late day policy above.

Access and accommodations

Stanford and our class are committed to providing equal educational opportunities for disabled students. Disabled students are a valued and essential part of the Stanford community. We welcome you to MS&E 228.

If you experience disability, please register with the Office of Accessible Education (OAE). Professional staff will evaluate your needs, support appropriate and reasonable accommodations, and prepare an Academic Accommodation Letter for faculty (not only for your teaching staff in MS&E 226, but also your other courses). To get started, or to re-initiate services, please visit oae.stanford.edu.

If you already have an Academic Accommodation Letter from OAE, we invite you to share your letter with us. Academic Accommodation Letters should be shared at the earliest possible opportunity so we may partner with you and OAE to identify any barriers to access and inclusion that might be encountered in your experience of this course.

Honor code

Your work on problem sets and the project is governed by the Stanford Honor Code. Any violations of the Honor Code will be referred to the Office of Community Standards for adjudication.

Submitting on Gradescope

As noted above, problem sets and project submissions are submitted and graded through Gradescope. To ensure this process is smooth, there are a few things to keep in mind:

• You are required to tag your answers correctly. The graders will ignore any part of your solution that is not tagged. Note that this means you also have to correctly tag your code. Allow enough time prior to submission to ensure you are able to tag correctly.
• In order to grade code you submitted, we need to be able to copy your code. Make sure this is possible (e.g., do not upload screenshot images of your code). We will deduct points if we cannot check your code.
• If you believe we have made a mistake grading your work, you should submit a regrade request through Gradescope. This sends your request directly to the grader on that particular question. You must submit your regrade request within 14 days of the grades of that particular problem set or project part being published.

Course communications: Ed Discussion

As noted above, we will use Ed Discussion to manage course announcements and a discussion forum. Ed Discussion will be available through Canvas.

Please use Ed Discussion for all course-related communication with us. We will aim to respond to questions in a 24-48 hour period, except of course for those of an urgent nature (e.g., typos on problem sets or lecture notes, clarifying course logistics, etc.). We encourage students to respond to each other, particularly during this “waiting period” before course staff answers! Among other things this waiting period means you should not wait until the last day before a problem set is due to message us; we may not be able to respond in time.

Office hours

You are encouraged to attend office hours to ask questions of a technical nature. Office hours will be held in-person. Our goal is to foster an inclusive environment for additional learning support during office hours.

If you are having difficulty, we urge you to seek help from the TAs or the instructor as soon as possible. The material builds on itself and a solid understanding of the foundations is necessary for the rest of the course. Remember, the teaching team is here to serve as a resource for any questions you may have.

Inclusion

It is our intent that students from a diverse set of backgrounds and with a wide range of perspectives be well served by this course; that students’ learning needs are being addressed both in and out of class; and that the diversity that students bring to this class becomes a resource, strength, and benefit to all of us. We try to present materials and lead discussions that are respectful of the diversity in our student population. Your suggestions are encouraged and appreciated; please let us know of ways to improve the effectiveness of the course for you personally or for other students or student groups.

Unfortunately, incidents of bias, discrimination, or intolerance do occur, whether intentional or unintentional. They can contribute to creating an unwelcoming environment for individuals and groups in the classroom. Please speak out if such an event occurs and we will do our best to handle it accordingly. You can reach out to the teaching team directly or bring the concerns to the University’s Diversity and Access Office and Acts of Intolerance to Student Affairs.

Suggested textbooks

Course lecture notes will be posted online in the form of powerpoint presentations. In addition accompanying chapters of a course textbook will be released on Canvas. Any feedback on slides and the course textbook will be greatly appreciated as we iterate over the material. These are the primary material for the course. In addition, you may find it helpful to have the following textbooks on hand. I’m not requiring them, because they are available online, and we will not be linearly working our way through any of them.

• Hernan MA, Robins JM, Causal Inference: What If
• Scott Cunningham, Causal Inference: The Mixtape
• Kiciman, Sharma, Causal Reasoning: Fundamentals and Machine Learning Applications
• Shai Shalev-Shwartz and Shai Ben-David, Understanding Machine Learning: From Theory to Algorithms
• Wainwright, High Dimensional Statistics: A Non-Asymptotic Viewpoint

  Other references

Here is an assortment of other books that you may find useful to consult but which are not available online.

• Judea Pearl, Causality
• Pearl, Glymour, Jewell, Causal Inference in Statistics
• Pearl, Mackenzie, The Book of WHY: the new science of cause and effect
• Peters, Janzing, Schölkopf, Elements of Causal Inference
• Matt Taddy, Business Data Science
• Kohavi, Tang, and Xu, Trustworthy Online Controlled Experiments
• Imbens and Rubin, Causal Inference for Statistics, Social, and Biomedical Sciences
• Angrist and Pischke, Mostly Harmless Econometrics

Course Plan

Lecture 1: Introduction; case studies; importance of causality; importance of handling high dimensional data/flexible modeling;

Experiments and causality

Lecture 2: Causality via Experiments; Potential Outcomes framework; Two means estimate and confidence interval/asymptotic distribution; limitations of trials; what if we have pre-treatment co-variates: precision and heterogeneity

Inference with linear models

Lecture 3: Basics of statistical inference in linear models; confidence intervals for p « n; simultaneous confidence bands; interpretation of coefficient as partialling out; inference on ATE from trials via regression; Revisiting the role of covariates in randomized trials: precision and heterogeneity: variance characterization and comparisons

Lecture 4: High dimensional methods and prediction; regularization; lasso; elasticnet;

Lecture 5: Inference in high-dimensional methods; double lasso; partialling out; intro to Neyman orthogonality

Observational data, causality, DAGs

Lecture 6: Causality in observational data; confounding; conditional ignorability; identification by conditioning; identification via propensity scores

Lecture 7: Structural equations models and DAGs; basics of DAGs; conditional ignorability in DAGs; Good and Bad controls

Lecture 8: General DAGs and Counterfactuals; SWIGs; D-separation; Interventions; Re-visting identification by conditioning

Lecture 9: Graphical criteria for valid adjustment sets; Good and Bad Controls

ML estimation of non-linear models

Lecture 10: Modern methods for non-linear prediction: trees and forests; neural networks; feature engineering; some guarantees

Lecture 11: Ensembling; stacking; auto-ML

Statistical inference with non-linear models

Lecture 12: DML for PLR and fully non-linear for ATE; Generic debiased ML framework

Unobserved Confounding

Lecture 13: Omitted variable bias; Instrumental variables; LATE; proximal inference

Lecture 14: Inference in PL IV and non-linear IV models; inference with weak instruments; DML with weak identification

Heterogeneous Effects and Policy Learning

Lecture 15: CATE methods; meta learners; neural network methods; policy learning

Lecture 16: Evaluation and model selection of CATE methods; for Trials; for Observational Data

Further Topics (Subject to change)

Lecture 17: Censoring

Lecture 18: Dynamic regime; Optimal regime; off-policy RL; Surrogates

Lecture 19: More structural approaches to un-observed confounding: diff-in-diff; synthetic controls; regression discontinuity (soft RD)

Lecture 20: TBA