Statistical analysis of randomized trials in tobacco treatment: longitudinal designs with dichotomous outcome.
Statistical analysis of randomized trials in tobacco treatment: longitudinal designs with dichotomous outcome.
复制标题
DOI：
10.1080/14622200110050411
复制
发表时间：
2001-08-01
期刊：
影响因子：
--
通讯作者：
Niaura, R
中科院分区：
文献类型：
--
作者：
Hall, S M;Delucchi, K L;Niaura, R
本文考虑了烟草治疗临床试验数据统计处理中的两个重要问题：（1）纵向研究的数据分析策略和（2）缺失数据的处理。关于数据分析策略，方法被分类为“时间幼稚”或纵向。时间幼稚方法包括比例检验和逻辑回归。纵向方法包括广义估计方程和广义线性混合模型。它的结论是，尽管有一些优点累积到“时间天真”的方法，在大多数情况下，纵向的方法是优选的。纵向方法允许时间检验的直接影响和治疗与时间的相互作用，并允许基于所有可用数据的模型估计。缺失数据策略的讨论探讨了完整的案例分析，最后一次观察结转，平均替代方法，以及使用烟草编码缺失数据的参与者所产生的问题。审查了不同缺失数据情况之间的区别。得出的结论是，最佳缺失数据分析策略包括仔细描述数据缺失的原因，以及沿着使用模式混合或选择建模。提出了报告缺失数据的标准化方法。提供了数据分析策略和缺失数据处理的参考和软件程序。
This article considers two important issues in the statistical treatment of data from tobacco-treatment clinical trials: (1) data analysis strategies for longitudinal studies and (2) treatment of missing data. With respect to data analysis strategies, methods are classified as 'time-naive' or longitudinal. Time-naive methods include tests of proportions and logistic regression. Longitudinal methods include Generalized Estimating Equations and Generalized Linear Mixed Models. It is concluded that, despite some advantages accruing to 'time-naive' methods, in most situations, longitudinal methods are preferable. Longitudinal methods allow direct effects of the tests of time and the interaction of treatment with time, and allow model estimates based on all available data. The discussion of missing data strategies examines problems accruing to complete-case analysis, last observation carried forward, mean substitution approaches, and coding participants with missing data as using tobacco. Distinctions between different cases of missing data are reviewed. It is concluded that optimal missing data analysis strategies include a careful description of reasons for data being missing, along with use of either pattern mixture or selection modeling. A standardized method for reporting missing data is proposed. Reference and software programs for both data analysis strategies and handling of missing data are presented.