Summary: | 碩士 === 國立成功大學 === 統計學系 === 89 === A statistical tool that is becoming popular for describing
diagnostic accuracy is the receiver operating characteristic
(ROC). Its purpose is to determine whether a test has the
capacity to discriminate between disease and nondisease
states. Traditionally, the area under the ROC curve (AUC)
has been used to analyze data by Hanley (1989). The estimator for this area is releated to the Mann-Whitney U statistic.
Anyway, most procedures are applicable for paired data, in which both diagnostic markers are performed on each subject. In this thesis, we compare multiple diagnostic markers at the same
time, and use four different statistical approaches to
discuss their advantage and disadvantage. We use bootsrap
procedure to construct the confidence interval of the area
difference for any two ROC curves, and use repeat measure to compare whether it is significant among different diagnostic markers. Then, use the method of DeLong, DeLong and
Clarke-Pearson(1988) through computing the estimator of
variance covariance matrix for AUC's. We also apply a nonparametric approach
to analysis AUC's. Furthermore, we use bootstrap procedure to compare the specificity at a fixed sensitivity.
Finally, these four methods are applied to a real data for
an example, and the Monte Carlo simulation results are provided.
|