Estimating signal amplitudes in optimal fingerprinting. Part II: application to general circulation models

We show that there is a significant low bias in standard estimates of the amplitudes of climate change signals estimated by small ensembles of coupled ocean atmosphere general circulation models. This bias can be eliminated either by making larger ensembles of at least eight members or by employing total least squares regression (TLS) to take account of sampling uncertainty in model-simulated signals. Results using TLS agree with previous work using ordinary least squares regression (OLS) in showing that recent interdecadal warming trends in near-surface temperature are largely anthropogenic in origin. Consistent with previous results, we detect evidence of solar influence on surface temperature changes in the first half of the twentieth century. However the amplitudes of model-predicted signals in the observed record were previously underestimated by ordinary least squares regression. This implies that over the last 30 years the observations are consistent with a greater degree of greenhouse warming and sulfate cooling than previously thought and the early century warming is consistent with a greatly enhanced model response to solar changes with very little contribution from anthropogenic causes. The model-simulated response to solar forcing is, however, relatively weak and subject to large uncertainties. Contributions of both anthropogenic and natural forcings to the early century warming are therefore very poorly constrained.