Evaluations of the Performance of the Nonparametric Tests for Comparing Accuracy Between Two Diagnostic Tests

• Main Authors: Tung, Tzu-Ping, 董子萍
• Other Authors: HWANG,YI-TING
• Format: Others
• Language: zh-TW
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/18630422940599588041

碩士 === 國立臺北大學 === 統計學系 === 101 === Many diagnostic tests have been developed as the government begins promoting many disease prevention programs.In general, the receiver operating characteristic (ROC) curve can be used to assess the performance ofthe diagnostic test when the test result is continuous. The area under the ROC curve (AUC) is the most common summary statistic to measure the accuracy of the diagnostic test. To evaluate the performance between two diagnostic tests, Hanley and McNeil (1983) suggested an area test comparing two AUCs. The concept of the area test is very intuitive and the asymptotic distribution of the test can be obtained. However, the power of this test is only good when two ROC curves have stochastic ordering. To have a better power for detecting the difference between two ROC curves that do not have the stochastic ordering property, Venkatraman and Begg (1996) and Venkatraman (2000) suggested a misclassification function to measure the difference between the two curves. Since the sampling distribution of this test is obtained by permuting the samples, this test is named as the permutation test. Since the area test derived by Hanley and McNeil (1983) needs the estimator of the covariance matrix for the AUC, this thesis provides the detailed derivations for obtaining the estimators. Furthermore, we suggest a modified permutation test that uses the smoothed misspecification function. Monte Carlo simulations are conducted to compare the performance of the modified tests. Finally, a real data is implemented to illustrate the feasibility of the modified test.The finding of this thesis can provide appropriate guidance for choosing a test.