Six problems for causal inference from fMRI