An analysis in which data capable of revealing the sought result are hidden while selection and inference procedures are chosen. It reduces result-dependent tuning. calibration samples and predefined validation samples remain necessary; blinding does not itself correct a defective background model.
