Causal Inference in Python

Chapter 6. Effect Heterogeneity

This chapter introduces what is perhaps the most interesting development of causal inference applied to the industry: effect heterogeneity. Up until this point, you understood the general impact of a treatment. Now, you’ll focus on finding how it can affect people differently. The idea that the treatment effect is not constant is simple, yet incredibly powerful. Knowing which units respond better to a certain treatment is key in deciding who gets it. Effect heterogeneity offers a causal inference approach to the cherished idea of personalization.

You’ll start by understanding effect heterogeneity on a theoretical level, what the challenges are in estimating effect heterogeneity, and how you can expand what you already learned to work around those challenges. Next, you’ll see that estimating heterogeneous effects is closely related to predictive problems, which are already very familiar to data scientists. Consequently, you’ll see how the idea of cross-validation and model selection still applies in treatment heterogeneity models. However, validating your effect estimate is much more challenging than evaluating a simple predictive model, which is why you’ll see some novel ideas on how to do it.

The chapter closes with some guidelines and examples on how to use effect heterogeneity to guide decision making. Although not exhaustive, I hope those examples will inform you on how to use these ideas on your own business problems.

From ATE to CATE

So far, ...