摘要
目的对完全随机设计时两样本比较的Wilcoxon、Kruskal-Wallis及Median三种方法的检验功效进行比较。方法用SAS9.13软件编程,采用Monte Carlo方法,设置数据不同的分布类型、样本量、样本量比率及方差齐性与否条件,比较三种方法的检验功效。结果正态分布,方差不等较之方差相等三法的功效高。方差相等时,偏态分布较之正态分布功效高。当分布为偏态,方差相等较之方差不等功效高。Kruskal-Wallis法在小样本时(n≤30)功效高于Wilcoxon法,n大于30时两者近似相等;中位数法功效低于其他两法,但在样本量n=100,且效应量较大(ES=0.8)时,其功效接近其他两法。结论样本量较小时(n≤30)建议采用Kruskal-Wallis法,样本量n>30时,Kruskal-Wallis和Wilcoxon法均可,样本量大于n>100时,三种方法均可采用。
Objective To compare the powers among the Wilcoxon test, Kruskal-Wallis test and Median test on completely randomized design of independent samples. Methods Monte Carlo Simulating method was used to evaluate the comparative powers of the three nonpara- metric tests with SAS program. Results The power of Kruskal-Wanis test was higher than that of Wilcoxon test when the sample size was small（ n≤ 30）. Otherwise, they are similar. The power of Median test was lower than that the others when it came to small sample size, but it was almost equal when the sample size was large enough（ n 〉 100） and effect size is large（ES = 0.8）. Conclusion When sample size is no more than 30, the Kruskal- Wallis test was recommended. When sample size is no more than 100, Wilcoxon test and Kruskal-Wanis test were recommended. When sample size is large enough（ 〉 100） ） and effect size is large （ES = 0.8）, the powers of the three tests were nearly equal.
出处
《中国卫生统计》
CSCD
北大核心
2008年第3期230-232,235,共4页
Chinese Journal of Health Statistics
