On Some Difficulties in a Frequency Theory of Inference