Behavioral Data Analysis with R and Python

Chapter 12. Mediation and Instrumental Variables

In the last chapter, we saw that moderation allows us to open the black box of a causal relationship by revealing groups for which that relationship is stronger or weaker. Mediation refers to the presence of an intermediary variable between two variables in a chain; it offers a different way to probe that black box by understanding the causal mechanism at play—the “how” of the causal effect.

This has several benefits on both the causal and behavioral sides of our framework. From a causal perspective, mediation reduces the risk of false positives, and not accounting adequately for it can bias our analyses. From a behavioral perspective, mediation helps us better design and understand experiments. In a sense, mediation is nothing new, and most of the arguments in this chapter could have been summarized as “expand the chains in your CDs as much as you can, at least at the beginning.” But I believe such a simplification would have been a disservice to you, because the search for mediators is at the root of many scientific discoveries. “But why?” is one of the best follow-up questions after confirming the causal relationship between two variables. Customer satisfaction increases retention, but why? Is it because it reduces the probability of looking for alternatives or because it increases the customer’s opinion of the company?

Mediation also offers a nice stepping stone for the last tool we’ll see in this book, instrumental variables ...