Comparing Methods of Uncertainty Estimation in Optimal Fingerprinting

Abstract

The performances of different uncertainty estimates in multipattern optimal fingerprinting are compared. The comparisons are performed in a perfect model context using the Canadian Earth System Model Large Ensemble to provide a clean comparison in the absence of model error or inconsistencies in internal variability. Performance is quantified in terms of the accuracy (correct coverage) of the confidence intervals for the fingerprint scaling factors. Sensitivity to temporal and spatial smoothing is tested. The most accurate method for interval estimation differs between variables, with a maximum likelihood method being most accurate for temperature and a bootstrap method being most accurate for precipitation. These differences can only partially be explained in terms of a relevant measure of signal‐to‐noise ratio.