On the interpretation of differences between groups for compositional data