Fisher确切检验

R 中的 A/B 测试

Lauryn Burleigh

Data Scientist

规范

分类数据的非参数方法
- 是否会再次食用干酪或意大利辣香肠披萨
样本 < 1000
至少20%单元频数 < 5
用Fisher确切检验求精确p值
2×2列联表

超几何分布

列联表的似然

table(Pizza$Topping, Pizza$EatAgain)

              No Yes
  Cheese     140 360
  Pepperoni   80 420

假设

原假设：再次食用与配料组无关

多只手拿起披萨切片的照片。

样本

library(pwr)
pwr.chisq.test(w = 0.1, df = 2, 
               power = 0.80, 
               sig.level = 0.05)

     Chi squared power calculation 
              w = 0.1
              N = 963.4689
             df = 2
      sig.level = 0.05
          power = 0.8
NOTE: N is the number of observations

table(Pizza$Topping, Pizza$EatAgain)

              No Yes
  Cheese     140 360
  Pepperoni   80 420

检验

freqtbl <- table(Pizza$Topping, 
                 Pizza$Time)
fisher.test(freqtbl)

    Fisher's Exact Test for Count Data
data:  freqtbl
p-value = 0.001809
alternative hypothesis: true odds ratio is not equal to 1
95 percent confidence interval:
 0.4569013 0.8417208
sample estimates:
odds ratio 
 0.6209987

效应量

freqtbl <- table(Pizza$Topping, 
                 Pizza$Time)
fisher.test(freqtbl)

小：1.5
中：2.5
大：4
意大利辣香肠披萨相对干酪披萨的比值比≈0.62

Fisher's Exact Test for Count Data
data:  freqtbl
p-value = 0.001809
alternative hypothesis: true odds 
ratio is not equal to 1
95 percent confidence interval:
 0.4569013 0.8417208
sample estimates:
odds ratio 
 0.6209987

检验功效

library(pwr)
pwr.2p.test(h = 0.62, n = 500, 
            sig.level = 0.002)

    Difference of proportion power 
 calculation for binomial distribution

              h = 0.62
              n = 500
      sig.level = 0.002
          power = 1
    alternative = two.sided

NOTE: same sample sizes

Passons à la pratique !

R 中的 A/B 测试