Evaluation of statistical treatments of left-censored environmental data using coincident uncensored data sets: I. Summary statistics
Evaluation of statistical treatments of left-censored environmental data using coincident uncensored data sets: I. Summary statistics
复制标题
DOI：
10.1021/es071301c
复制
发表时间：
2008-05-15
影响因子：
11.4
通讯作者：
Taylor, Howard E.
中科院分区：
文献类型：
--
作者：
Antweiler, Ronald C.;Taylor, Howard E.
用于确定文献中使用的基本统计量的低于检测限（左删失）环境数据的统计处理的主要类别是替代方法，最大似然法，顺序统计回归（ROS）和非参数技术。通过检查已知删失数据真实值的数据集，对这些处理以及使用所有仪器生成的数据（甚至低于检测值的数据）进行了评价，沿着。研究发现，对于删失数据小于70%的数据集，确定汇总统计量的总体最佳技术是非参数Kaplan-Meier技术。ROS和将检测限值的一半分配给删失数据或将零和检测限之间的随机数分配给删失数据的两种替代方法是适当的替代方法。然而，使用这两种替代方法需要彻底了解实验室如何审查数据。采用所有仪器基因额定数据的技术-包括低于检测限的数字-被认为是不够的，比上述技术。在高度删失（大于70%的删失数据）时，没有技术提供汇总统计量的良好估计。最大似然技术被认为是远远劣于所有其他治疗，除了替代零或检测限值删失数据。
The main classes of statistical treatment of below-detection limit (left-censored) environmental data for the determination of basic statistics that have been used in the literature are substitution methods, maximum likelihood, regression on order statistics (ROS), and nonparametric techniques. These treatments, along with using all instrument-generated data (even those below detection), were evaluated by examining data sets in which the true values of the censored data were known. It was found that for data sets with less than 70% censored data, the best technique overall for determination of summary statistics was the nonparametric Kaplan-Meier technique. ROS and the two substitution methods of assigning one half the detection limit value to censored data or assigning a random number between zero and the detection limit to censored data were adequate alternatives. The use of these two substitution methods, however, requires a thorough understanding of how the laboratory censored the data. The technique of employing all instrument-gene rated data-including numbers below the detection limit-was found to be less adequate than the above techniques. At high degrees of censoring (greater than 70% censored data), no technique provided good estimates of summary statistics. Maximum likelihood techniques were found to be far inferior to all other treatments except substituting zero or the detection limit value to censored data.