Optimal risk sharing with different reference probabilities