A model independent safeguard against background mismodeling for statistical inference

We propose a safeguard procedure for statistical inference that provides universal protection against mismodeling of the background. The method quantifies and incorporates the signal-like residuals of the background model into the likelihood function, using information available in a calibration dataset. This prevents possible false discovery claims that may arise through unknown mismodeling, and corrects the bias in limit setting created by overestimated or underestimated background. We demonstrate how the method removes the bias created by an incomplete background model using three realistic case studies.