For example, if we predict age based on location and the location variable is encoded as "south" or "north". Update make_mean_diffs() to work with that and return the discordant-style data.