Checking for model consistency in optimal fingerprinting

Current approaches to the detection and attribution of an anthropogenic influence on climate involve quantifying the level of agreement between model-predicted patterns of externally forced change and observed changes in the recent climate record. Analyses of uncertainty rely on simulated variability from a climate model. Any numerical representation of the climate is likely to display too little variance on small spatial scales, leading to a risk of spurious detection results. The risk is particularly severe if the detection strategy involves optimisation of signal-to-noise because unrealistic aspects of model variability may automatically be given high weight through the optimisation. The solution is to confine attention to aspects of the model and of the real climate system in which the model simulation of internal climate variability is adequate, or, more accurately, cannot be shown to be deficient. We propose a simple consistency check based on standard linear regression which can be applied to both the space-time and frequency domain approaches to optimal detection and demonstrate the application of this check to the problem of detection and attribution of anthropogenic signals in the radiosonde-based record of recent trends in atmospheric vertical temperature structure. The influence of anthropogenic greenhouse gases can be detected at a high confidence level in this diagnostic, while the combined influence of anthropogenic sulphates and stratospheric ozone depletion is less clearly evident. Assuming the time-scales of the model response are correct, and neglecting the possibility of non-linear feedbacks, the amplitude of the observed signal suggests a climate sensitivity range of 1.2-3.4 K, although the upper end of this range may be underestimated by up to 25% due to uncertainty in model-predicted response patterns.