In electroencephalography (EEG) brain imaging neural sources are estimated with the help of measured electric potentials around the head and prior information. The problem is highly ill-posed and prone to modelling errors. The most typical modelling errors are the geometry of the head and the electrical properties of the different tissues. In this talk, I will show our recent results regarding the use of the Bayesian approximation error (BAE) approach which can take variantions in these features into account. In BAE, first a probabilistic model is postulated for the uncertain parameters and subsequently an approximate marginalization is carried out. The results show that BAE can give feasible result without the precise knowledge of the uncertain head parameters.