Thinking with Data

Chapter 5. Causality

Causal reasoning is an important and underappreciated part of data science, in part because causal methods are rarely taught in statistics or machine learning courses. Causal arguments are extremely important, because people will interpret claims we make as having a cause-and-effect nature whether we want them to or not. It behooves us to try our best to make our analyses causal. Making such arguments correctly gives far more insight into a problem than simple correlational observations.

Causal reasoning permeates data work, but there are no tools that, on their own, generate defensible causal knowledge. Causation is a perfect example of the necessity of arguments when working with data. Establishing causation stems from making a strong argument. No tool or technique alone is sufficient.

Cause and effect permeates how human think. Causation is powerful. Whenever we think about relationships of variables, we are tempted to think of them causally, even if we know better. All arguments about relationships, even about relationships that we know are only associative and not causal, are positioned in relation to the question of causation.

Noncausal relationships are still useful. Illustration sparks our imaginations, prediction allows us to plan, extrapolation can fill in gaps in our data, and so on. Sometimes the most important thing is just to understand how an event happened or what examples of something look like. In an ideal scenario, however, our knowledge eventually ...