Sun Yanqing, Lee Jimin
Department of Mathematics and Statistics, University of North Carolina at Charlotte, Charlotte, NC 28223.
Stat Sin. 2011 Jul 1;21(3):1315-1339. doi: 10.5705/ss.2009.251.
A common problem associated with longitudinal studies is the dropouts of subjects or censoring before the end of follow-up. In most existing methods, it is assumed that censoring is noninformative about missed responses. This assumption is crucial to the validity of many statistical procedures. We develop some nonparametric hypothesis testing procedures to test for independent censoring in the absence/presence of covariates. The test statistics are constructed by contrasting two estimators of the conditional mean of cumulative responses for each stratum of covariate space from sample subsets with different patterns of censoring. Our method does not require the modelling of longitudinal response processes, therefore is robust to model misspecifications. A diagnostic plot procedure is also developed that can be used to identify dependent censoring to certain covariate strata. The finite sample performances of the tests are investigated through extensive simulation studies. The potential of our methods is demonstrated through the application of the tests to a chronic granulomatous disease study.
纵向研究中一个常见的问题是在随访结束前受试者退出或失访。在大多数现有方法中，假定失访对于缺失的反应没有信息性。这一假定对于许多统计程序的有效性至关重要。我们开发了一些非参数假设检验程序，以在存在/不存在协变量的情况下检验独立截尾。检验统计量是通过对比来自具有不同截尾模式的样本子集的协变量空间各层累积反应条件均值的两个估计量来构建的。我们的方法不需要对纵向反应过程进行建模，因此对模型误设具有稳健性。还开发了一种诊断图程序，可用于识别对某些协变量层的相依截尾。通过广泛的模拟研究考察了检验的有限样本性能。通过将这些检验应用于一项慢性肉芽肿病研究，展示了我们方法的潜力。
