 |
 |
The theory of the M-W test requires the two samples being compared to be drawn from continuous distributions so that there are no ties (with probability one). In practice, texts suggest that a moderate number of ties may be allowed, with tied values being given equal average rank, and some sort of correction possibly being applied. What about the case of large samples of discrete "count" data where there are almost necessarily large numbers of ties? It seems intuitively clear that the test must be weakened to the extent that it is virtually useless as well as flying in the face of the underlying assumption of continuous distributions. Are there any definitive results on the performance of M-W in such circumstances, or even any sorts of rules of thumb as to when one might get away with it? Any guidance, positive or negative would, be much appreciated. Quentin Burrell