In this presentation, linear models for optimal detection performance in single-subject confirmatory fMRI analysis will be explored. In particular, the advantage of constraining the linear models will be discussed. Furthermore, conventional mass-univariate fMRI analysis methods like the GLM will be extended to a mass-multivariate analysis via the introduction of spatial basis functions. This allows us to model variations in spatial shape of active brain regions, resulting in an adaptive spatial filtering of the data. Canonical Correlation Analysis will be presented as one possible tool for such mass-multivariate fMRI analysis.