Machine learning estimators of causal effects

Machine learning methods can be applied just like parametric estimators

• define a causal question where potential outcomes are missing
• assume a Directed Acyclic Graph to identify causal effects
• predict potential outcomes as a (machine learning) function of confounders and treatment
• aggregate across units to any quantity of interest

The strategy above can be improved by additionally estimating a model to predict the probability of treatment and combining the two in a particular way. We will not go into the mathematical details in this course—our goal is to see how machine learning could plug in to approaches you’ve already learned.