7 Interventions and causal effects
This chapter covers
- Case studies of interventions in machine learning engineering contexts
- How interventions relate to A/B tests and randomized experiments
- Implementing interventions on causal models with intervention operators
- Using a causal model to represent many interventional distributions
- Causal effects as natural extensions of an intervention distribution
An intervention is something an agent does to cause other things to happen. Interventions change the data generating process (DGP).
Interventions are the most fundamental concept in how we define causality. For example, the concept of intervention, written in terms of “manipulation” and “varying” a factor, is central to this definition from an influential 1979 textbook on experimental design:
The paradigmatic assertion in causal relationships is that manipulation of a cause will result in the manipulation of an effect . . . . Causation implies that by varying one factor I can make another vary. 1
Interventions are how we go from correlation to causality. Correlation is symmetric; the statements “Amazon’s laptop sales correlate with Amazon’s laptop bag sales” and “Amazon’s laptop bag sales correlate with Amazon’s laptop sales” are equivalent. But interventions make causality a one-way street: if Amazon recommends the sale of laptops, laptop bag sales will increase, but if Amazon promotes the sale of laptop bags, we wouldn’t expect people to respond by buying new laptops to fill them.