Learn how to utilize DML for estimating individual level treatment effects to enable data-driven targeting
This article is the 2nd in a 2 part series on simplifying and democratizing Double Machine Learning. In the 1st part, we covered the fundamentals of Double Machine Learning, along with two basic causal inference applications. Now, in pt. 2, we will extend this knowledge to turn our causal inference problem into a prediction task, wherein we predict individual level treatment effects to aid in decision making and data-driven targeting.
Double Machine Learning, as we learned in part 1 of this series, is a highly flexible partially-linear causal inference method for estimating the average treatment effect (ATE) of a treatment. Specifically, it can be utilized to model highly non-linear confounding relationships in observational data and/or to reduce the variation in our key outcome in experimental settings. Estimating the ATE is particularly useful in understanding the average impact of a specific treatment, which can be extremely useful for future decision making. However, extrapolating this treatment effect assumes a degree homogeneity in the effect; that is, regardless of the population we roll treatment out to, we anticipate the effect to be similar to the ATE. What if we are limited in the number of individuals who we can target for future rollout and thus want to understand among which subpopulations the treatment was most effective to drive highly effective rollout?
This issue described above concerns estimating treatment effect heterogeneity. That is, how does our treatment effect impact different subsets of the population? Luckily for us, DML provides a powerful framework to do exactly this. Specifically, we can make use of DML to estimate the Conditional Average Treatment Effect (CATE). First, let’s revisit our definition of the ATE:
Now with the CATE, we estimate the ATE conditional on a set of values for our covariates, X: