变量变换及其在营养调查资料分析中的应用

目的：通过对营养调查资料分析中的有关统计方法的比较和对变量变换方法的应用探讨说明在营养调查资料分析中如何正确地应用有关统计方法。方法：利用一项在上海地区开展的大规模女性肺癌病例对照研究资料，先采用直方图、Ｐ－Ｐ图及偏度、峰度检验等方法来说明数据是否符合正态分布。再对参数与非参数统计学检验的方法进行比较。重点讨论了变量变换及其应用，文中采用的变量变换方法为：自然对数变换、平方根变换、Ｂｏｘ－Ｃｏｘ幂指数转换。结果：无论是病例，还是对照人群在膳食营养素的摄入量方面往往都是非正态分布的。而且同一样本数据参数与非参数统计学检验两者的结果并不一致。变量变换后几乎所有的膳食营养素的分布情况都有了不同程度的改善。但转换后真正满足正态分布的数据病例组仅约三分之一，而对照组更低。其余数据要么偏度失正态、要么峰度失正态，或两者同时失正态。在各种变换方法中似以Ｂｏｘ－Ｃｏｘ幂指数转换法较好。结论：应从样本数据本身的性质或特点出发，决定在营养调查资料分析中采用的统计方法；也可考虑利用一些变量变换法对数据进行预处理，然后再进行统计学检验

DATA TRANSFORMATION AND ITS APPLICATION IN DATA ANALYSIS IN NUTRITIONAL SURVEY DATA

Objective: In order to demonstrate how to use statistical methods in data analysis in nutritional sunvey accurately, we compared the performance of parametric and non parametric statistical methods which were frequently used in data analysis. Further, we explored the methods on data transformation and its application in detail. Method: The data generated from a large scale population based case control stduy on female lung cancer in Shanghai urban were used in our study. First, the techniques we used in our paper for testing the normality were the histogram, P P plot, Skewness and Kurtosis tests. Second, we compared the statistical tests used in data analysis, then the methods on data transformation and its application were discussed in detail. The methods of natural logarithm, square root and Box Cox power transformation were compared to those before transfromation. Result: Our data analysis indicated that the distributions of nutrient intake data in case and control were not normality, and there were some differences in the results from parametric and non parametric tests. After transformation, there were about 30% of nutrient intake data in the case distributing normally, and less than 30% in the control. However, the others were not normal distributions after transformation by Skewness or/and Kurtosis tests. The Box Cox transformation seemed to be better than other transfromation methods. The comparisons of results of the tests before and after transformation were also performed. Conclusion: The choice of statistical tests should be based on the characteristics of the sample data. Some transformation methods are recommended to be used in data analysis before statistical tests.