We estimate the speed of income convergence for a sample of 196 European NUTS 2 regions over the period 1985-1999. So far there is no direct estimator available for dynamic panels with strong spatial dependencies. We propose a two-step procedure, which involves first spatial filtering of the variables to remove the spatial correlation, and application of standard GMM estimators for dynamic panels in a second step. Our results show that ignorance of the spatial correlation leads to potentially misleading results. Applying a system GMM estimator on the filtered variables, we obtain a speed of convergence of 6.9 per cent and a capital elasticity of 0.43.