Detection bias and the role of negative control outcomes.

Investigators, patients, or clinicians knowing which treatment is assigned in pragmatic randomised trials and observational analyses can lead to detection bias (ie, systematic differences in determining outcomes between groups). A structural definition of detection bias with directed acyclic graphs is provided, together with several published examples. Why negative control outcomes are best placed to assess detection bias is discussed, and how to correctly select a negative control outcome for this purpose is explained.