We provide a generalisation of activation pattern-based approaches to correct the interpretations of additive models, particularly in the presence of suppressors and correlated noise.