Collider bias from selecting disease samples distorts causal inferences.