学了一点儿统计，觉得ROBINSNOW要介绍一下这个STUDY的DESIGN先。 在我看来, you have an outcome variable which is continuous and two covariates (group factor & subgroup factor) which are categorical. What is the group factor, treatment or block? What is the subgroup factor? Is subgroup factor independent of group factor or nested within it? If it is the former, then it seems you have a 2x2 factorial design. If it is the latter, then it is a block design. For either case, you should use ANOVA to evaluate the differences between groups or subgroups. However, ANOVA is just a general term, there are different methodologies for different designs. Personally, I don't think two-sample t test or Wilcoxon rank sum test will do the work in this case. Just my two cents.

In addition, you don't need to calculate the test result yourself. Besides, sometimes the computation is complicated and to do it by hand would be very hard, especially when you have to estimate variances or compare multiple pairs using Tukey or Bonferroni. I suggest you use a statistical software such as SAS.

以下是引用robinsnow在2004-5-18 14:44:06的发言：
1. 两个subgroup的data point之间，并没有一一对应的关系，所以这个difference, 只有统记意义上的阿
2. 我以为这个normal distribution是说两个subgroup的data, 是不是各自成正态分布
因为只有他们自己成正态分布的条件下，才能够用variance去normalize两个mean之间的difference

那么这个正态分布的假设，究竟是哪个variable的分布呢？

KAKA，好像大家都越说越乱。ROBINSNOW，要不你把RAW DATA贴一下，如果不涉及机密的话，呵呵。要是DATA比较大，取一小段举例说明也成。我还是觉得要知道每个VARIABLE是什么(可以用假名)，还有到底what the question of interest is.