## Causal learning

This book is by and large a book about statistical learning. Given data *X* and targets *Y*, we aim to estimate , the distribution of target values given certain data points. Statistical learning allows us to create a number of great models with useful applications, but it doesn't allow us to claim that *X* being *x* caused *Y* to be *y*.

This statement is critical if we intend to manipulate *X*. For instance, if we want to know whether giving insurance to someone leads to them behaving recklessly, we are not going to be satisfied with the statistical relationship that people with insurance behave more reckless than those without. For instance, there could be a self-selection bias present about the number of reckless people getting insurance, while those who are not marked as reckless don't.

Judea Pearl, a famous computer scientist, invented a notation for causal models called do-calculus; we are interested in , which is the probability of someone behaving recklessly after we manipulated...