本帖最后由 乾乾 于 2011-9-15 09:57 编辑
对于 效应大小
我们知道,检验做的是相对比较,即处理效应的大小(M-u)相对于随机性差异(标准误)而言。如果标准误很小,即使有显著效应,处理效应可能也是很小的。所以存在显著效益不一定意味着存在本质上的处理效应。由此可知,假设检验的局限在于仅评估了处理效应的相对大小,而不是绝对大小,即统计上显著性不能提供关于处理效应的真实大小的信息。
这个时候,就引出了一个概念 —— 效应大小(effect size , ES),以它作为处理效应绝对大小的估计值,其大小表示两个分布的重合程度,ES越大,重叠越小,效应越大,反之,略。对于ES的计算,我们常使用科恩d值 科恩d值=平均数差/标准差
对于 统计效能 —— power
为什么要引入这个概念:
Rather than examining the potential for making an error, we examine the probability of reaching the correct decision. Remember that the researcher’s goal is to demonstrate that the experimental treatment actually does have an effect. This is the purpose of conducting the experiment in the first place. If the researcher is correct and the treatment really does have an effect, then what is the probability that the hypothesis test will correctly identify it? This is a question concerning the power of a statistical test.
定义:The power of a statistical test is the probability that the test will correctly reject .
也就是实验产生处理效应被统计检验侦测到的能力
As the definition implies, the more powerful a statistical test is, the more readily it will detect a treatment effect when one really exists (correctly rejecting H0).
注意:
It should be clear that the concepts of power and Type II error are closely related. When a treatment effect exists, the hypothesis test will have one of two results:
1. It can fail to discover the existing treatment effect (a Type II error).
2. It can correctly detect the presence of a treatment effect (rejecting a false null hypothesis).
简单来说,效应大小和统计效能是相关的。当处理效应增加时,处理后的分布将向右侧移动,会有更多的样本超过界限(z=1.96)。因此,当效应大小增加时,拒绝虚无假设的概率也增加了,这意味着统计效能增加了。因此,效应大小的测量,如科恩d值,与统计效能一样,都显示了处理效应的大小。
但统计效能“并不单纯”,也就是说,他并不单纯对应于效应大小的测量。
因为,除了效应大小外,它还被其他几个因素影响:样本大小、a水平、单尾或双尾。
|