Evaluation of climate model simulations by means of statistical methods

Sammanfattning: Evaluation of climate model simulations is a key issue within climate research. The statistical framework proposed by Sundberg et al., 2012, provides a theoretical underpinning of methods for evaluation of climate models by use of climateproxy data from the last millennium. In the present work, the statistical framework above is used to suggest several latent factor models of different complexity that can be used for estimating the amplitude of a forcing effect in aclimate model by comparison with the observed/reconstructed climate. The performance of the models is evaluated and compared in a pseudo-proxy experiment, in which the true unobservable temperature series is replaced by selected realizations of a climate simulation model. For different levels of added noise, different conclusions can be drawn. However, for realistic noise levels, we find that the simplest model, the just-identified two-indicator one-factor model, denoted j.i.FA(2,1), is a competitive alternative to models with more complicated structure. Moreover, we discover that the Fieller method of constructing confidence regions, associated with the j.i.FA(2,1)-model, outperforms the Wald confidence interval, which in most cases fails to provide sensible and interpretable conclusions about the climate model under consideration. Last but not least, the results indicate a good performance of the j.i.FA(2,1)-model even in the presence of heteroscedasticity.