Self S G, Etzioni R
Fred Hutchinson Cancer Research Center, Seattle, Washington 98104, USA.
Biometrics. 1995 Mar;51(1):44-50.
In randomized cancer screening trials, mortality rates for the screened group relative to those of the control group are not likely to be constant as a function of years from randomization due to the inherent lag between initiation of screening and any putative effects of screening on mortality. In this situation, a log rank test for differences in mortality between the randomization groups will not be optimal. Although optimality could potentially be recovered by use of a weighted log rank statistic, the optimal weights are difficult to specify a priori and the potential loss of power by use of poorly specified weights is great. We describe a likelihood ratio test with two degrees of freedom for use in this situation which is based on a fit of a weakly structured full model. Computation of an approximate significance level for this test is described and a large sample justification for this approximation is given. Size and power properties of the proposed statistic are compared to that of several other statistics in a small simulation study and the statistic is applied to data from the HIP Breast Cancer Screening Trial.
在随机癌症筛查试验中，由于筛查开始与筛查对死亡率的任何假定影响之间存在固有的时间滞后，筛查组相对于对照组的死亡率不太可能作为随机分组后年份的函数保持恒定。在这种情况下，用于检验随机分组组之间死亡率差异的对数秩检验并非最优。尽管通过使用加权对数秩统计量有可能恢复最优性，但最优权重难以事先确定，并且使用指定不当的权重可能导致的功效损失很大。我们描述了一种在此情况下使用的具有两个自由度的似然比检验，该检验基于一个结构较弱的全模型拟合。文中描述了此检验近似显著性水平的计算方法，并给出了这种近似的大样本合理性依据。在一项小型模拟研究中，将所提出统计量的大小和功效特性与其他几种统计量进行了比较，并将该统计量应用于HIP乳腺癌筛查试验的数据。
