Parametric Exploration of the Mann-Whitney U Test
Abraham Nunes MD PhD MBA
Dalhousie University, Halifax, Nova Scotia, Canada
The Binormal ROC Model and the U-Statistic
Recall that the CDF of a normal distribution is
(1.1) |
(1.2) |
Assuming that and that we can compute the ROC curve for two normal distributions as follows:
(1.3) |
The area under the ROC curve (AUC) can be computed as follows:
(1.4) |
The U-statistic for the binormal model with sample sizes and can be computed from the AUC as follows:
(1.5) |
Computing P-Values from the U-Statistic
At large samples, the U-statistic is normally distributed with mean
(2.1) |
and the standard deviation is
(2.2) |
We can therefore compute a Z-score:
(2.3) |
A two-tailed p-value is thus
(2.4) |