In functional MRI (fMRI), analysis of multisubject data typically involves spatially normalizing (i.e. co-registering in a common standard space) all data sets and summarizing results in a single group activation map. This widely used approach does not explicitely account for between-subject anatomo-functional variability. Therefore, we propose a group effect analysis method which makes use of a multivariate model to select the main signal variations that are common to all subjects, while allowing final statistical inference on the individual scale. The normalization step is thus avoided and individual anatomo-functional features are preserved. The approach is evaluated by using simulated data and it is shown that sensitivity is drastically improved compared to more conventional individual analysis.