Could it be possible that utility/importance of levels across variables varies with counts, what could be the possible reasons...
Only thing I can think of is we estimate utilities on overall basis and not on segment wise/variable wise basis so the variation could be lost when we look at it that ways