Confounding & Independence
Confounding and non-independence are the two most common sources of spurious effects in observational data and poorly controlled experiments. They cannot be corrected after the fact without the right model — and they cannot be addressed at all unless the relevant metadata (covariates, batch, subject IDs) are recorded and declared.
What is Confounding & Independence Guard?
The Confounding & Independence section surfaces situations where a third variable may be correlated with both the predictor and the outcome, where observations may not be truly independent given the design, and where the set of covariates included in the model may not be fully justified. Implementation depth varies across these three pages; see each page for current product behavior and known limitations.
Confounding is not bias: Confounding is a structural property of the data — a third variable that creates an apparent relationship between exposure and outcome. Detecting it before analysis allows appropriate adjustment. Failing to detect it produces estimates that appear causal but are not.