The holy grail of causal inference is the individual-level treatment effect: how would a particular patient respond to a drug? Which users will respond most to a targeted ad? Would a given student be helped or harmed by a classroom intervention? This session introduces machine learning tools for estimating heterogeneous treatment effects like random causal forests. The course goes over the theory and concepts as well as the nitty-gritty of coding the methods up in python, R, and Stata using real-world examples. This course can be taken as a follow-up to the Machine Learning and Causal Inference mixtape session, or as a stand-alone course.
-
Review of Machine Learning and Causal Infernece Course
- Potential outcomes and treatment effects
- Basic causal inference summary
- Prediction Target
- Prediction Methods
- Prediction mechanics
- Decision Trees
- Forest for the Trees
-
Combining causal effects and ML: predicting heterogeneous treatment effects
- Traditional heterogeneity analysis: Interacted regression
- Challenges with traditional heterogeneity analysis
- Predicting outcomes vs. treatment effects
- Adapting ML to predict treatment effects
The following is a set of readings for analyzing heterogeneous effects with machine learning methods:
Athey and Imbens (2016): Introduction to using trees to estimate heterogeneous treatment effects.
Athey, Tibshirani, and Wager (2019): Warning! Very technical material. But contains the theory for using machine learning to estimate heterogeneous effects in a wide class of settings
Athey and Wager (2019): A more accessible application of the methods we will be using
Wager and Athey (2018): Technical/theoretical background for random causal forests