Causal Inference in Python

Chapter 10. Geo and Switchback Experiments

In Part IV of this book, you learned how to use repeated observations over time to aid in the process of causal inference. Now, in this chapter, you will approach the same problem from a different angle. What if, instead of having to use panel data to identify a treatment effect, you had to design an experiment to gather that data? Part V of this book is dedicated to alternative experimental design when simple A/B testing won’t work.

For example, let’s consider the marketing problem from the previous chapter. Remember that inferring the impact of marketing is challenging because you cannot randomize people who are not yet your customers. Online marketing provides you with attribution tools, but attribution is not the same as incrementality. In this case, a promising alternative is to conduct a geo-experiment: treat entire markets, such as a city or a state, while leaving others as control. This approach would provide you with panel data to which you could apply the techniques learned in Part IV. However, in Part IV, you took the panel data as given and did not learn how to best select the treated and control markets in such an experiment. In this chapter, you will cover that gap. The first part of this chapter will teach you how to select geographical treatment units to get an effect estimate that approximates the effect you would have if the entire market (country, state) were treated.

The idea is to zoom out the unit of analysis from ...