Backdoor Adjustment

This concept gives you a formal rule for answering a question that comes up constantly in real-world AI work: "If I have data on many variables, which ones should I control for to find out the true effect of one thing on another?" Without this rule, you guess — and guessing produces misleading answers that can get a loan denied, a medical diagnosis wrong, or a recommendation system recommending the wrong thing. Mastering backdoor adjustment unlocks the rest of causal inference: you will be able to understand and apply deconfounding, design better experiments (or know why you can't), and build AI systems that don't just predict but actually explain why something happened. This connects directly to the other topics in your library: Causal Deconfounding, Fairness and Bias, and Explainable AI. With backdoor adjustment, you will be able to look at a causal graph and say with confidence which variables to condition on — and which to avoid — to get an unbiased estimate of a causal effect.

Imagine you are trying to figure out whether a new medicine truly helps people recover from a disease. You collect data on thousands of patients: who took the medicine, who recovered, and a bunch of other things about them — age, how sick they were at the start, whether they also took a supplement, and so on. If you just look at the raw numbers, you might find that people who took the medicine recovered more often. But that could be misleading: maybe older, sicker people were less likely to be given the medicine in the first place, and their age and initial sickness are the real reasons they recovered less. The medicine might actually help, but the data hides it. To find the true effect of the medicine, you need to 'adjust' for those other variables — you need to compare people who are similar in age and initial sickness, but differ only in whether they took the medicine. The question is: which variables do you need to adjust for, and which should you NOT adjust for (because doing so would actually cause more harm)? The backdoor adjustment is a formal rule that tells you exactly that. It works by looking at the causal graph — a diagram showing which variables cause which other variables. The rule identifies all the 'backdoor paths' — paths that go from the treatment to the outcome but not through the treatment directly — and says: you must condition on a set of variables that blocks all of these paths. If you condition on the right set, the remaining association between treatment and outcome reflects only the direct causal path, not any spurious correlation. This is the core idea: correlation is not causation, but with the right adjustment, you can turn correlation into a causal estimate.

Think of a detective trying to solve a murder. There are two ways the suspect could be linked to the victim: directly (they were seen stabbing the victim) and indirectly (they had a motive, an opportunity, and a history of anger — but no one saw them at the scene). The detective wants to know if the suspect actually did it, not just if they had a motive. The 'backdoor' in this analogy is the motive. The suspect's motive (e.g., a large inheritance) could explain why they were near the victim, even if they never touched them. To figure out if the suspect actually did the killing, the detective must 'control for' the motive — that is, they must ask: 'Even among people with the same motive, did the suspect kill the victim more often than others?' If the answer is yes, that gives evidence of a direct causal link. If the answer is no, then the motive explains the entire association, and the suspect might be innocent. In causal graphs, the 'motive' is a confounder — a variable that causes both the treatment (being near the victim) and the outcome (the murder). The backdoor path is the path that goes from the suspect to the motive to the victim. To block it, you condition on the motive. But here is where the analogy shows its limits: in a real causal graph there can be many variables, and some of them are colliders — variables that are caused by two other variables. If you condition on a collider, you actually open a path that was previously closed. Imagine the detective conditions on 'the victim's phone records' — but that phone record is caused by both the victim's life and the suspect's calls. By conditioning on it, the detective accidentally creates a false correlation between the suspect and the victim, even if there was none before. So the detective must be careful: conditioning on some variables closes backdoor paths, but conditioning on a collider opens a new one. This is the subtlety that the backdoor adjustment rule makes precise.

Backdoor adjustment is a formal criterion that, given a causal graph, identifies which variables to condition on (the 'adjustment set') so that all backdoor paths between a treatment and an outcome are blocked, thereby allowing an unbiased estimate of the causal effect.

You have not yet studied any other topics in causal inference, so this page builds from scratch. But your library notes mention that backdoor adjustment is the first step in Causal Deconfounding — removing common causes to turn an association into a causal effect. Once you master this, you will be able to understand why randomised experiments are the gold standard (because by design they block all backdoor paths), and why observational data requires this kind of adjustment. This concept also connects to Fairness and Bias in AI: when a model uses a protected attribute (like gender or race) to make decisions, backdoor adjustment helps estimate the true effect of that attribute on the outcome, separate from confounding factors. Your library's book notes say: 'Causal graphs and backdoor adjustment let you estimate a protected attribute's true effect rather than its correlation.' That is exactly what you will be able to do after this page.