We examine the similarity between patterns of changes in decadally-averaged surface air temperature simulated by ensemble experiments with the HadCM2 coupled climate model and observed patterns over the period 1860--1999, using a spatial pattern correlation method. The analysis is conducted for two anthropogenic (greenhouse gases with/without sulphate aerosols) and two natural (solar and volcanic) modelled forcing histories. For each scenario an ensemble of at least four model runs is used. We compare transient model signal and observed patterns at corresponding times, and establish significance using a 1000 year model control run to construct sampling distributions for random noise patterns. We find evidence of an anthropogenic signal in observed surface temperature patterns emerging over recent decades, particularly in northern winter and spring seasons. Signals from greenhouse gas forcing alone or in combination with sulphate aerosols are both plausible in our analysis, but the highest significance is achieved using seasonally-defined greenhouse gas plus aerosol signals. There is little evidence of any detectable solar signal, but we do find two periods (late 19th century; 1960s--90s) during which a weak volcanic signal is detectable. Using ensemble simulations makes the conclusions quite robust, but the somewhat different interpretations from this analysis compared to optimal fingerprint analysis of the same simulations reveals some sensitivity to the choice of methodology. ¿ 2001 American Geophysical Union