Causal Inference in Python

Chapter 11. Noncompliance and Instruments

It’s not uncommon for companies to offer products or services to their existing customer base. For instance, a retailer can offer a subscription-based program where customers get free shipping. A streaming company can offer an ad-free version of its services for an additional fee. Or a bank can offer a prime credit card with lots of perks for customers who spend above a certain threshold.

In all of these examples, the customer must opt in for the additional service, which makes inferring its impact challenging. As the choice to participate lies with the customer, that choice often confounds the impact evaluation of the service; after all, customers who opt in and customers who don’t will likely have different $Y_0$. Even if the company randomizes the availability of the service or product, it can’t force customers to take it. This is called noncompliance, where not everyone that gets assigned to the treatment takes it. This chapter will walk you through how to think about this issue and what to consider when you want to design an experiment that suffers from noncompliance.

 

Noncompliance

Noncompliance comes from pharmaceutical science (though some of the tooling to deal with it comes from economics). Imagine you are conducting an experiment to test the effect of a new drug on an illness. Each subject gets assigned to a treatment: a drug or a placebo. But those subjects are imperfect human beings, who sometimes forget to take their medicine. ...