Why bother to calibrate? Model consistency and the value of prior information