Watch a short overview of the course and what you’ll learn.
This 10-hour course is the first part of a series on causal inference with linear regression. It is designed for people who want to understand precisely how linear regression can be used as an estimation tool in causal analysis, and under which conditions regression coefficients can be interpreted as causal effects.
The course starts with a concise introduction to the core concepts from causal inference that are needed for the remainder of the course. This includes causal quantities, conditional exchangeability, and the adjustment formula.
From there, the course turns to the mechanics of linear regression, with an emphasis on how regression coefficients are computed and how linear regression can be understood as an approximation to the conditional expectation function.
Building on this, the course develops the connection between causal quantities and linear regression as a statistical tool. Linear structural causal models are introduced to define the causal parameters of interest, and their relationship to regression equations is made explicit. This allows us to study the exact conditions under which regression coefficients recover these causal parameters, and when they do not.
Finally, the course covers robustness tests and sensitivity analysis for unobserved confounding. The focus here is on how violations of assumptions such as unconfoundedness affect OLS-based causal estimates, how this sensitivity can be quantified, and how these methods can be implemented in Python.
The course makes explicit how linear regression relates to causal inference. Rather than relying on abstract assumptions such as exogeneity or ambiguous arguments about unbiasedness for “true parameters,” the connection between regression coefficients and causal quantities is developed in a transparent and structured way.
A central part of the course is dealing with violations of the unconfoundedness assumption. In particular, the course covers robustness tests and the sensitivity analysis framework of Cinelli & Hazlett (2020), which allows you to quantify how strong unobserved confounding would need to be to change your conclusions. These tools are essential in practice, but rarely covered in detail.
The material is based on high-quality sources in causal inference, including work by Judea Pearl, Joshua Angrist & Jörn-Steffen Pischke, Carlos Cinelli, and Chad Hazlett.
The course makes heavy use of coding demonstrations and exercises. The goal is to make sure that everything we discuss conceptually also becomes tangible and applicable in practice.
The course develops a structured understanding of how linear regression can be used for causal inference.
Individual, conditional, and average treatment effects (ITE, CATE, ATE) for both categorical and continuous treatments
Conditional exchangeability (unconfoundedness)
The adjustment formula
Directed acyclic graphs (DAGs)
The role of bidirected (double-headed) arrows
The backdoor criterion
How OLS determines regression coefficients
The Frisch–Waugh–Lovell theorem
The relationship between linear regression and the conditional expectation function
Linear structural causal models
The meaning of structural causal parameters
The relationship between structural causal parameters and regression coefficients
The conditions under which OLS coefficients recover causal parameters
The consequences of violations of unconfoundedness for OLS-based causal estimates
When and how robustness tests can be used to stress-test causal assumptions
Sensitivity analysis for unobserved confounding based on the framework of Cinelli & Hazlett (2020), including interpretation of sensitivity measures and practical implementation in Python
This course is intended for data scientists, analysts, and other quantitative practitioners who want a deep understanding of how and when linear regression can be used for causal inference.
A basic understanding of probability and statistics is assumed. Some familiarity with Python is helpful, as coding examples and exercises will be in python.
Upon completion of the course, you will receive a certificate of completion from Causal Academy.
This course develops the foundation for using linear regression in causal analysis by making the connection between regression and causal inference explicit.
In Part II of the course series, this foundation is extended to more flexible linear modelling approaches, including feature engineering and double machine learning based on LASSO and OLS, and embedded in a complete end-to-end applied causal analysis workflow.
Managing Director @ Accenture
"A very good choice. Step by step it builds confidence and helps differentiate the subtleties of the argument."
Data Expert @ Athentic Consulting
“High quality (both presentation slides and narration). This is the course I am looking for to learn causal inference.”
Based on 30+ reviews from previous versions of this course on Udemy.