The weak magnetic fields generated by neural brain activity can be measured outside the head using the magnetoencephalography (MEG) technique. MEG-scans are usually recorded in order to determine the location of certain functional brain areas, which are typically activated by the presentation of an external stimulus, e.g., a picture, a sound etc. Based on repeatedly measured responses, the location and the strength of the sources that generate the neural response, the source parameters, are estimated. The responses, however, are contaminated with noise, consisting mainly of brain background activity, that is correlated both in time and in space. Therefore an estimated covariance matrix has to be incorporated in the estimation of the source parameters. Because of the huge dimension of this matrix, parameterization of the covariance is needed. We discuss three different models for this noise covariance, which are all based on the general decomposition of a matrix into a sum of Kronecker products.