A First Course in Causal Inference by Peng Ding

Peng Ding, A First Course in Causal Inference, a freely available textbook posted on arXiv.

Blog Written by Mingzhao Hu

I should disclose at the outset that I have known Peng Ding since I was a PhD student, back when this book was still something he mentioned he was working on. So I came to A First Course in Causal Inference already rooting for it. But the reason I first considered teaching from it was very practical: the whole book is available free on arXiv (https://arxiv.org/abs/2305.18793). For a reading seminar with students and trainees who do not all share the same library access or book budget, that mattered. Everyone could open the same PDF on day one, with no waiting on interlibrary loan or navigating licenses. If you take only one thing from this review, it can be that link. But free access is not, by itself, a reason to teach from a book. The reason I keep recommending Ding’s text is the way it sets up the subject.

Many of us first encounter causal inference as a collection of methods: propensity scores here, instrumental variables there, inverse-probability weighting when a reviewer asks for it. Ding starts somewhere quieter and more fundamental. He frames causal inference as a missing-data problem. Each unit has two potential outcomes, only one of them is observed, and the rest of the book is about what assumptions are needed to fill in the missing piece. In our seminar, the early chapters drew a clear line between the missing-data language students already knew (for example, MCAR, MAR, MNAR) and the causal versions of the same idea. A randomized trial is the clean case where treatment assignment does not depend on anything; unconfoundedness is the observational-study cousin, where treatment can depend on covariates but not on the potential outcomes themselves. Watching that connection click for a room of health-sciences students was the moment I knew the book was doing real work.

The book builds from there. Randomized experiments come first, where balance is created by design and students can see exactly what randomization buys. Then comes covariate adjustment, used to improve efficiency and address chance imbalance. Only after that does Ding move into observational studies, where nothing is guaranteed and inference moves forward by considering the covariates using approaches like propensity scores, matching, weighting, doubly robust estimators, sensitivity analysis, and finally instrumental variables. By the time students reach more sophisticated methods, they begin grasping the intuition underlying the concepts. That is very different from memorizing an estimator and hoping it applies.

For a health-sciences audience, this sequencing is close to ideal. The arc from randomized trials to observational data to sensitivity analysis is the arc many of our trainees will encounter in their own work, whether they are analyzing data from a registry, defending a comparative-effectiveness study, or responding to a reviewer’s concern about unmeasured confounding. The R code described throughout the book allows students to run the methods, not just read about them, and the design-based flavor keeps the focus on what the study design actually supports rather than on which package to call.

There are a few points to note. The book is rigorous. The potential-outcomes machinery and finite-population, design-based arguments ask more of students than a softer applied text would. A first-year master’s group may want a companion source, additional examples, or a supportive instructor who would patiently explain some of the proofs. The book is also lean on worked clinical narratives, so instructors in health sciences will likely want to bring their own examples from clinical trials, EHR studies, registries, or comparative-effectiveness research. Our group considered all of these as important features.

My recommendation is straightforward: A First Course in Causal Inference is an excellent resource for a graduate reading group, doctoral seminar, or methods-oriented health-sciences audience that wants to understand causal inference from first principles rather than as a menu of techniques. It is not the lightest introduction, and it is not a clinical casebook. But it is clear, principled, freely accessible, and highly effective at teaching students about causal methods.