The Backdoor Criterion

The one idea: There's an exact rule for which variables you must control for and which you must leave alone. Get the set right and an observational comparison becomes a defensible causal estimate. Get it wrong — even with fancier methods and more variables — and you get a precise, confident, wrong number.

⚠️ The trap: "If the confounder isn't in my data, a fancier method or more rows will make up for it." That feels right and it's wrong. No amount of matching, propensity scoring, or extra data fixes a confounder you never measured — if something unseen drives both the treatment and the outcome, you need a different research design, not a bigger model.

The last two lessons left you with a tension. Confounders demand controls; colliders and mediators punish them. So which set of variables is enough — and which is too much? That question has an exact answer, and it's the closest thing causal inference has to a checklist you can run in a meeting: the backdoor criterion, from Judea Pearl's work and the workhorse of both Cunningham's Mixtape and Facure's Brave and True.

The stakes are concrete. A control set that satisfies the criterion turns your observational comparison into a defensible effect estimate. A control set that doesn't — even one with more variables, fancier matching, and a propensity score — produces a precise, confident, wrong number. The difference between the two isn't visible in the regression output. It's only visible in the graph.

The criterion in plain English

You want the causal effect of treatment D on outcome Y. A set of control variables Z is good…