The trouble with likelihood ratios