Optimizing methods for linking cinematic features to fMRI data