Back-door criterion
A graphical rule for which variables to adjust for to get a clean causal estimate.
What it means
The back-door criterion is Judea Pearl's graphical test for deciding which variables to adjust for when estimating the effect of a treatment on an outcome. A back-door path is any path from treatment to outcome that starts with an arrow into the treatment; when left unblocked, such paths transmit spurious association. If a set of variables contains no descendant of the treatment and blocks every back-door path — without opening new ones by conditioning on colliders — then adjusting for it identifies the causal effect, provided the assumed graph is correct. The no-descendant clause is what keeps mediators and other downstream variables out of the adjustment set. The test reads confounding off the structure of the diagram rather than off correlations in the data, which is why it can reject a variable that looks confounding and demand one that does not.
How the rule works
Pearl states it in two parts. A set Z satisfies the back-door criterion for the effect of X on Y if no member of Z is a descendant of X, and Z blocks every path from X to Y that begins with an arrow into X. Blocking follows the d-separation rules: a chain or a fork is blocked by conditioning on its middle variable, while a collider — two arrowheads meeting head-to-head — is blocked already and gets opened by conditioning on it or on anything downstream of it. That asymmetry is the whole trick. If Z passes, the effect follows from the adjustment formula: estimate the outcome within each stratum of Z, then average across strata weighting by the population's own distribution of Z.
What it replaced
Before the criterion, confounder selection ran on habit: adjust for every pretreatment variable available, or keep the ones that shift the estimate or reach significance. The back-door rule shows why that cannot work in general. Whether a variable belongs in the model is a property of where it sits in the causal structure, not of how strongly it correlates with treatment or outcome. Two variables can carry identical associations with both and still demand opposite handling: one blocks a spurious path, the other is a collider whose inclusion creates one. Nothing in the correlation matrix separates them. This is also why adjusting for more is not the cautious choice — an added variable can push the estimate further from the truth, and the fit statistics will not report it.
Where it breaks down
The criterion is a theorem, so it does not fail on its own terms; the graph you feed it does. It certifies a set as sufficient given the assumed structure, and one missing arrow or unmeasured common cause voids the certificate without leaving a trace in the output. The rule is also sufficient rather than necessary. Perkovic and colleagues give an adjustment criterion that is sound and complete for DAGs and for the equivalence classes you get when a graph is learned rather than assumed, and it accepts valid sets that back-door reasoning alone rejects. Validity is not optimality either: among qualifying sets the asymptotic variance differs, and Rotnitzky and Smucler derive graphical rules for choosing the efficient one. Blocking every back-door path is a floor, not a finish.
Using it in practice
Draw the graph before you see the estimate and write it down; the criterion's value is that it makes an assumption auditable, and that is lost if a DAG is reverse-engineered from a result you liked. Tools such as dagitty enumerate the minimal sufficient sets. When the structure is genuinely unknown, VanderWeele and Shpitser's disjunctive cause criterion is a usable fallback: adjust for any pretreatment covariate that causes the treatment, the outcome, or both, which suffices whenever any observed subset does. Two much-cited hazards deserve calibration rather than panic, though how much calibration is itself contested. M-bias, from conditioning on a collider of two latent causes, is real, and Ding and Miratrix find its magnitude typically small beside the confounding bias adjustment removes — a reading Pearl rejects in a published comment on their paper, arguing that under an exact M-structure the bias runs to the same order of magnitude as the bias adjustment is meant to remove. Adjusting for a near-instrument amplifies bias, though Myers and colleagues bound the inflation near twenty percent in plausible ranges.
Examples
To estimate a drug's effect, the back-door criterion might tell you to adjust for age and disease severity but not for a downstream lab value.
To estimate whether a training course raises wages, block the back-door path through prior ability by adjusting for education and experience, but not for the job someone took after training, which the course caused.
Comparing two hospitals' survival rates, adjust for how ill patients were on arrival, since that shapes both which hospital they reach and whether they live; adjusting for length of stay would not help.
Estimating whether a paid ad drove a purchase, adjust for the browsing history that led the platform to show the ad; conditioning on clicks, which the ad causes, would absorb the effect you want.
To estimate whether a credit-limit increase raises default, adjust for the income and score the bank used to grant it; but restricting the sample to accounts the bank later flagged for review conditions on a collider — flags are triggered both by the increase and by early distress — and manufactures an association between them.
First described in Judea Pearl (1993, 1995).
Key references
- Rotnitzky, A., & Smucler, E. (2020). Efficient adjustment sets for population average causal treatment effect estimation in graphical models. Journal of Machine Learning Research, 21(188), 1-86. www.jmlr.org/papers/v21/19-1026.html
- Perkovic, E., Textor, J., Kalisch, M., & Maathuis, M. H. (2018). Complete graphical characterization and construction of adjustment sets in Markov equivalence classes of ancestral graphs. Journal of Machine Learning Research, 18(220), 1-62. jmlr.org/papers/v18/16-319.html
- Ding, P., & Miratrix, L. W. (2015). To adjust or not to adjust? Sensitivity analysis of M-bias and butterfly-bias. Journal of Causal Inference, 3(1), 41-57. doi.org/10.1515/jci-2013-0021
- VanderWeele, T. J., & Shpitser, I. (2011). A new criterion for confounder selection. Biometrics, 67(4), 1406-1413. doi.org/10.1111/j.1541-0420.2011.01619.x
- Pearl, J. (2015). Comment on Ding and Miratrix: "To adjust or not to adjust?". Journal of Causal Inference, 3(1), 59-60. doi.org/10.1515/jci-2015-0004
- Myers, J. A., Rassen, J. A., Gagne, J. J., Huybrechts, K. F., Schneeweiss, S., Rothman, K. J., Joffe, M. M., & Glynn, R. J. (2011). Effects of adjusting for instrumental variables on bias and precision of effect estimates. American Journal of Epidemiology, 174(11), 1213-1222. doi.org/10.1093/aje/kwr364
- Pearl, J. (1995). Causal diagrams for empirical research. Biometrika, 82(4), 669-688. doi.org/10.1093/biomet/82.4.669