The Starter Guide to Causal Inference

Four chapters, running from probability basics through graphical models, interventions and counterfactuals. Every chapter ends in study questions.

Those study questions are the entire reason to buy it.

011 tells you the ladder exists. 012 proves the machinery is complete. Neither makes you compute anything. This one hands you a graph, a table of numbers, and asks for the back-door adjustment. You find out quickly whether you understood d-separation or merely nodded at it.

The highlight for me is colliders, and it completes a pattern across all three books.

A confounder is a variable you must adjust for. A collider is a variable that, if you adjust for it, manufactures a correlation that was never there. Same data, same regression code, opposite error. Nothing in the data tells you which one you’re looking at, because the difference lives in the arrows, not the numbers.

That’s the same lesson as Simpson’s paradox in 011 and the front-door criterion in 012, arriving a third time from a third direction. The graph decides. The data does not.

Where I’d push back.

“Primer” oversells it. One reviewer put the problem well: the presentation is informal and assumes little beyond basic probability, which helps as an introduction and frustrates anyone who wants rigor. It still takes real work. Don’t hand this to someone who last saw conditional probability at school.

Download the errata before you open the book. Corrections went to Wiley in 2016, 2018, 2019 and 2020. Depending on your printing, you can lose an evening debugging the book instead of the idea.

Causal discovery gets almost nothing. Again. Three books, three times, and it’s the part I keep needing.

And there’s no code. Fair enough in 2016, less so now. Pair it with DoWhy or something similar, or you’ll finish with the theory and none of the practice.

Verdict: 5 out of 5 as the place to start.

A read order, if you want one. This book to learn it. The Book of Why to see why anyone cared. Causality when you hit something neither can answer. Then Hernan and Robins, or Imbens and Rubin, so you get the potential-outcomes side from people who actually believe in it.