请教成分数据的问题

shiyiming

X和Y分别代表一组比重，每一个case代表一年的纪录。现在需要对它们建立回归模型。请问需要进行怎样的预处理，怎么判断数据服从何种分布。谢了

shiyiming

直接做回归不就行了吗？干吗还要预处理啊？？？

shiyiming

如果讲的是比重,也许问题不是那么简单.如果我理解的是对的,比重应当是一组百分比,而百分比分布在0-1之间,不是正态分布的.一般要采用对数变换.建议读一读ATCHINSON的一本书,Statistical analysis of the compositional data.

shiyiming

我看过埃克逊和张尧庭的著作，但是不太明白。主要问题是
1）如何判断资料是否服从加法logit分布还是狄氏分布
2）资料不满足回归模型的LINE中的I的前提怎么办?
即在case之间有可能不独立的情况下如何建立模型

shiyiming

Well, I think it is a topic beyond the discussion here, and it can not be explained in a short message.  In my understanding, the choice of L-dist and D-dist is not based on which distribution describe the data better, but what kind of statistics do you want to derive.  If data does not meet the basic requirements of the regression model, you have to go deeper in statistics to derive a more appropriate model.  You may think of mixed model or Baysian analysis, if they are appropriate for your purpose.

shiyiming

看来我还是书没有看透，需要再下功夫才行。