Himes John H, Hoaglin David C
Division of Human Development and Nutrition, University of Minnesota, School of Public Health, Minneapolis, Minnesota 55455.
Department of Statistics, Harvard University, Boston, Massachusetts 02138.
Am J Hum Biol. 1989;1(2):165-173. doi: 10.1002/ajhb.1310010205.
Resistant delineation, a technique adapted from exploratory data analysis (Tukey, Exploratory Data Analysis, 1977), was applied to smooth age-specific percentiles for triceps skinfold thickness across ages from 1 to 20 years. Row percentiles were transformed to logarithms to promote symmetry and to render variability more nearly homogeneous across ages. The delineation involved smoothing resistantly the sequences of age-specific log medians and the sequence of age-specific positive differences between the "4253H, twice" (Velleman, J. Am. Statist. Assoc., 75:609-615, 1980). The delineation concluded by recombining these smoothed sequences to obtain smoothed percentiles in the log scale. Finally, the logarithmic transformation was reversed, yielding the smoothed age-specific percentiles. Comparisons of smoothed results from resistant, delineation with the original data indicated a satisfactory fit. Comparisons with published smoothed percentiles, obtained from the same data by a cubic-spline procedure, showed that the resistant delineation captured the structure of the raw data better than the cubic-spline procedure. The resistant delineation procedure makes few assumptions of the underlying data, it ensures a proper order relationship among the smoothed percentiles, it is relatively insensitive to isolated unusual data, and it is available in a common software package.
抗性描绘是一种从探索性数据分析（图基，《探索性数据分析》，1977年）改编而来的技术，用于平滑1至20岁各年龄段三头肌皮褶厚度的年龄特异性百分位数。行百分位数被转换为对数，以促进对称性并使各年龄段的变异性更接近均匀。描绘过程包括对抗性地平滑年龄特异性对数中位数序列以及“4253H，两次”（维勒曼，《美国统计协会杂志》，75:609 - 615，1980年）之间的年龄特异性正差异序列。描绘的最后一步是将这些平滑后的序列重新组合，以获得对数尺度下的平滑百分位数。最后，对数变换被反转，得到平滑后的年龄特异性百分位数。将抗性描绘的平滑结果与原始数据进行比较，显示拟合效果令人满意。与通过三次样条程序从相同数据获得的已发表平滑百分位数进行比较，结果表明抗性描绘比三次样条程序能更好地捕捉原始数据的结构。抗性描绘程序对基础数据的假设很少，它确保了平滑百分位数之间有恰当的顺序关系，对孤立的异常数据相对不敏感，并且在一个常见的软件包中即可使用。
