Identifying Causal Effects for Data Science Causal Inference
What you'll learn
- Measure the causal impact of treatments for product analytics and policy evaluation
- Think through causal inference problems from first principles
- Map information you know about the world and the industry you work in into measures of causal impacts
- Distinguish between what the data tells you about causal effects and what comes via your assumptions
- Translate knowledge about how an industry or product works into bounds on treatment effects
- Think carefully and deductively about making inference on treatment effects
- Use instrumental variables with both heterogenous treatment effects and homogenous effects
- Use the Conditional Independence Assumption to Identify Average Treatment Effects and the risks and potential bias from using the assumption
- Use the parallel trends assumption to motivate difference-in-difference estimation, the downsides of the assumption, and a more robust alternative
Requirements
- Basic statistics and probability. CDFs. Conditional Expectations.
- Useful to have some background on linear regression.
- Not a math-heavy course. Some algebra and basic properties of probability and expectations.
Description
The most common question you’ll be asked in your career as a data scientist is: What was/is/will be the effect of X? In many roles, it’s the only question you’ll be asked. So it makes sense to learn how to answer it well.
This course teaches you how to identify these “treatment effects” or "causal effects". It teaches you how to think about identifying causal relationships from first principles. You'll learn to ask:
What does the data say by itself?
What do I know about the world that the data doesn’t know?
What happens when I combine that knowledge with the data?
This course teaches you how to approach these three questions, starting with a blank page. It teaches you to combine your knowledge of how the world works with data to find novel solutions to thorny data analysis problems.
This course doesn't teach a "cookbook" of methods or some fixed procedure. It teaches you to think through identification problems step-by-step from first principles.
As for specifics:
This course takes you through various weak assumptions that bound the treatment effect—oftentimes, the relevant question is just: “Is the treatment effect positive?”—and stronger assumptions that pin the treatment effect down to a single value. We learn what the data alone—without any assumptions—tells us about treatment effects, and what we can learn from common assumptions, like:
Random treatment assignment (Experimentation)
Conditional independence assumptions (Inverse propensity weighting or regression analysis)
Exclusion restrictions (Instrumental variable assumptions)
Repeated Measurement assumptions
Parallel Trends (Difference-in-difference)
And many assumptions you will probably not see in other courses, like:
Monotone instrumental variables
Monotone confounding
Monotone treatment selection
Monotone treatment response
Monotone trends
(Why do they all include “monotone” in the name? The answer to that question is beyond the scope of this course.)
Lectures include lecture notes, which make it easy to review the math step by step. The course also includes quizzes and assignments to practice using and applying the material.
My background: I have a PhD in Economics from the University of Wisconsin — Madison and have worked primarily in the tech industry. I’m currently a Principal Data Scientist, working mainly on demand modeling and experimentation analysis problems—both examples of treatment effect estimation! I am from sunny San Diego, California, USA.
I hope you’ll try the Preview courses and enroll in the full course! I’m always available for Q/A.
-Zach
Who this course is for:
- Data Scientists looking to expand their repertoire of methods for estimating causal effects
- Data Scientists looking to improve their skills to derive novel approaches for measuring causal effects from first principles
- People who know statistics but want to learn how to think carefully about identifying causal treatment effects
- Anyone who wants to better understand how assumptions and data work together to identify treatment effects
Instructor
- 4.8 Instructor Rating
- 6 Reviews
- 108 Students
- 1 Course
Hi all, I'm Zach, an Economist and Data Scientist from sunny San Diego, California, USA. I have a PhD in Economics from the University of Wisconsin Madison, where I focused on Econometrics (using statistics for economic problems) and Industrial Organization (how companies compete). I work in the tech industry as a Principal Data Scientist, studying all sorts of problems, including how to analyze experiments, price products, and use data to inform product design.