Portal de Periódicos da CAPES

A Paired Prentice-Wilcoxon Test for Censored Paired Data

1987; Oxford University Press; Volume: 43; Issue: 1 Linguagem: Inglês

10.2307/2531957

1541-0420

Peter C. O’Brien, Thomas R. Fleming,

Statistical Methods and Inference

Prentice (1978, Biometrika 65, 167-179) has provided a censored-data generalization of classical linear rank statistics. This paper investigates a method to generalize the Prentice statistics to the analysis of censored data. Particular emphasis is given to the paired-data version of the Prentice-Wilcoxon statistic, with discussion of the reasons for preferring Prentice over the Gehan scores. The Prentice-Wilcoxon (PPW) statistic makes use of interblock information and can be viewed as a censored-data generalization of the Conover-Iman (1981, The American Statistician 35, 124-129) paired t test on the ranks or of the Lam-Longnecker (1983, Biometrika 70, 510-513) modified Wilcoxon rank sum statistic. In simulations comparing the PPW, sign, and generalized signed rank (GSR) statistics, the PPW is most powerful against all but the exponential scale alternative. Furthermore, the small advantage of the GSR over the PPW in that alternative is lost if outlier pairs are introduced. When compared to the two-sample Prentice-Wilcoxon statistic, the PPW is nearly as powerful in uncorrelated data and its power becomes superior as correlation between members increases.