Neural activity in the brain is correlated with the blood-oxygen level dependent (BOLD) contrast which can be measured non-invasively by functional magnetic resonance imaging (fMRI). Up to date, many fMRI analysis methods are based on simplifying assumptions about the BOLD signal. Two popular assumptions are spatial independence and homogeneity of the haemodynamic response function (HRF) across voxels. As single voxels usually are not independent and moreover also exhibit different haemodynamic response characteristics, these assumptions might lead astray interpretations of fMRI data. In this study we present an analysis framework that reveals the spatio-temporal correlation structure between simultaneously measured intracortical neurophysiological activity in primary visual cortex of the non-human primate and BOLD response. Given the spectrograms of neurophysiological activity and the simultaneously recorded BOLD data we compute a spatiotemporal convolution that links the activity measured at the electrode to the multivariate BOLD response. The convolution can be interpreted as the pattern in time-voxel space that reflects best the neural activity as it maximises the canonical correlation [1] between neural and haemodynamic activity. Inspection of the estimated time-voxel patterns yields new insights in the spatio-temporal dependency structure of neurovascular coupling mechanisms. This study thereby extends previous results reported in [2,3], where the convolution was a time-frequency convolution estimated for the neurophysiological activity. We show results from data collected during spontaneous activity and during visual stimulation. The analysis resulted in robust spatio-temporal coupling patterns across different experimental conditions. We compared the multivariate patterns with univariate coupling measures and spatial principal component analysis (SPCA), a popular method for connectivity analysis on fMRI data. Our findings suggest that neither univariate methods nor unimodal methods such as SPCA, which are based on autocorrelations of fMRI data only, can recover the multivariate spatio-temporal coupling structure in primary visual cortex. References 1. Hotelling H, Biometrika, 1936 2. Biessmann F, Meinecke FC, Gretton A, et al., Machine Learning Journal, 2009 3. Murayama Y, Biessmann F, Meinecke FC, et al., Magnetic Resonance Imaging, 2010