Causal Inference in Python

Chapter 7. Metalearners

Just to recap, in Part III you’re focusing on treatment effect heterogeneity, that is, identifying how units respond differently to the treatment. In this framework, you want to estimate:

$${\tau }_i\left(x\right)=E[Y_i\left(1\right)-Y_i\left(0\right)|X]=E\left[{\tau }_i\right|X]$$

or, $E\left[\delta Y_i\left(t\right)\right|X]$ in the continuous case. In other words, you want to know how sensitive the units are to the treatment. This is super useful in the case where you can’t treat everyone and need to do some prioritization of the treatment; for example, when you want to give discounts but have a limited budget. Or when the treatment effect is positive for some units but negative for others.

Previously, you saw how you could use regression with interaction terms to get conditional average treatment effect (CATE) estimates. Now, it’s time to throw some machine learning algorithms into the mix.

Metalearners are an effortless way to leverage off-the-shelf predictive machine learning algorithms for approximating treatment effects. They can be used to estimate the ATE, but, in general, they are mostly used for CATE estimation, since they can deal pretty well with high-dimensional data. Metalearners serve to recycle predictive models for causal inference. All predictive models, such as linear regression, boosted decision trees, neural networks, or Gaussian processes, can be repurposed for causal inference using the approaches described in this chapter. Therefore, the success of the metalearner ...