|
本帖最后由 Menuett 于 2013-12-22 15:59 编辑 : Y# p- E* ^! X! ?
煮酒正熟 发表于 2013-12-20 12:05
# I" t/ c3 b& y基本可以说是显著的。总的来说,在商界做统计学分析,95%信心水平是用得最多的,当95%上不显著时,都会去 ...
! b. s8 L4 ]$ g1 U9 C* k: g1 Q; w7 G
这个其实是一种binomial response,应该用Contigency Table或者Logisitic Regression(In case there are cofactors)来做。只记比率丢弃了Number of trial的信息(6841和1217个客户)。 8 ?. Y# c" v* h/ F
8 J! m" C# C3 y" [, Y8 b- Q
结果p=0.5731。 远远不显著。要在alpha level 0.05的水平上检验出76.42%和75.62%的区别,即使实验组和对照组各自样本大小相同,各自尚需44735个样本(At power level 80%)。see: Statistical Methods for Rates and Proportions by Joseph L. Fleiss (1981)
3 \' Z! } {9 s- H3 c
( R% R) \5 O k' ~7 M: B. N' OR example:3 m+ a5 b6 `' @) M4 l& W K
' ]- H) K3 C, p# W6 ?> M<-as.table(rbind(c(1668,5173),c(287,930))); ]: x3 Y2 H$ u+ s; ^# f/ W
> chisq.test(M)
0 J- X' K, D. V% O- l7 E8 [
3 _6 z& p( @' } Pearson's Chi-squared test with Yates' continuity correction2 _* s* W$ ^6 R$ J- v' `
; B0 u, T8 P- J1 p! e4 ^
data: M3 V1 Z8 K) y& f' k+ {8 R
X-squared = 0.3175, df = 1, p-value = 0.57310 K5 _/ t5 q* N' x8 p# L# G
4 ~6 E: x+ B4 a/ z; H/ KPython example:
( ?" G) e- S+ y) u7 }2 O6 Q
" H8 Z" U$ k) B6 n>>> from scipy import stats6 b2 L; k- h" s7 ^& X: d
>>> stats.chi2_contingency([[6841-5173,5173],[1217-930,930]])$ c7 s& R& s1 I
(0.31748297614660292, 0.57312422493552839, 1, array([[ 1659.73628692, 5181.26371308],) J, b& B" x6 P
[ 295.26371308, 921.73628692]])) |
|