S41.4: Comparison of nonparametric methods for the analysis of longitudinal data

✍ Scribed by Tania Schink; Klaus-Dieter Wernecke

• Publisher: John Wiley and Sons
• Year: 2004
• Tongue: English
• Weight: 72 KB
• Volume: 46
• Category: Article
• ISSN: 0323-3847
• DOI: 10.1002/bimj.200490145

✦ Synopsis

There are often longitudinal data in clinical research, where parametric methods cannot be used because of categorical response and/or small sample sizes. propose a generalization of the nonparametric log-rank and Gehan-Wilcoxon tests. Their tests are also asymptotically distribution-free and can handle missing values by assuming a random-censorship-model. Recently, marginal models, basing on only trivial assumptions, have been developed. They are valid for arbitrary, possibly noncontinuous distribution functions and can handle ties, missing values and singular covariance matrices. . These different nonparametric methods are compared in distinct situations. Simulated type-I errors and rejection probabilities are calculated for different covariance patterns (independence, compound symmetry, first order autoregressive process). Moreover the behaviour of the methods in critical situations (e.g. many time points, missing values and cells, ties) and in presence of very small sample sizes is investigated by simulations.