What is Confounding?
Confounding is distortion of a treatment-outcome association caused by differences in common causes of treatment assignment and the outcome, so the observed comparison does not represent the intended causal effect.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
A confounder is a causal role, not a table column type
The same variable can be a confounder for one treatment-outcome pair, a mediator for another, or irrelevant under a different time boundary. Subject-matter knowledge and temporal order must determine the role. The AHRQ causal-DAG guide explains how open backdoor paths represent confounding and why adjusting for a collider can open a path that was previously blocked.
Adjustment targets comparability, not predictive accuracy
Restriction, randomization, stratification, matching, standardization, weighting, and outcome regression can address measured confounding under their assumptions. Select pre-treatment covariates based on the causal structure and estimand, then check overlap and balance. A treatment model with excellent discrimination can worsen overlap, and a highly predictive outcome model can still omit the variable needed for causal identification.
Residual and unmeasured confounding require honest limits
The Cochrane ROBINS-I guidance distinguishes confounding from selection and information bias and emphasizes pre-specifying confounding domains. Measurement error, coarse proxies, wrong functional forms, and missing common causes leave residual bias. Sensitivity analyses, negative controls, alternate adjustment sets, and target-trial checks can quantify fragility but cannot certify no hidden confounding.
Key Characteristics
- Makes an observed association differ from the intended causal effect
- Depends on the treatment, outcome, time order, and causal structure
- Typically involves common causes of assignment and outcome
- Is not identified by correlation, significance, or feature importance alone
- Can remain after adjustment through omission or measurement error
- Differs from selection bias, mediation, effect modification, and collider bias
Common Use Cases
- Explaining why self-selected feature users have higher retention
- Choosing baseline covariates for an observational treatment study
- Detecting Simpson's paradox across risk strata
- Auditing post-treatment variables before model fitting
- Designing sensitivity analyses for unmeasured common causes
Example
Loading code...Frequently Asked Questions
Is every variable associated with treatment and outcome a confounder?
No. Statistical association is only a screening clue. A confounder is defined by its causal role and timing. Mediators, colliders, instruments, and proxies can also be associated with treatment and outcome, but adjusting for them may change the estimand or introduce bias.
Can machine learning eliminate confounding?
Machine learning can estimate flexible propensity or outcome functions for measured covariates. It cannot guarantee that all common causes were measured, correct post-treatment adjustment, create overlap, or determine the causal graph from predictive accuracy.
What is residual confounding?
Residual confounding is bias that remains after adjustment because a confounding domain was omitted, measured with error, represented too coarsely, or modeled inadequately. A larger dataset reduces random error but does not automatically remove this systematic bias.
Why can adjusting for more variables increase bias?
Conditioning on a collider can open a noncausal path, while conditioning on a mediator can block part of the total effect. Post-treatment variables may encode both mechanisms. Adjustment sets should follow the causal question and assumed graph, not an all-features rule.
How should unmeasured confounding be handled?
State it as an identification risk, use design knowledge to bound plausible common causes, and perform quantitative sensitivity analyses, negative controls, or alternate designs where possible. No ordinary balance metric or regression fit proves that unmeasured confounding is absent.