难度 运行 R 中的逻辑回归

Difficulty running a logistic regression in R

我在 运行 使用 glm 在 R 中进行逻辑回归时遇到了一些困难。有两种方法可以将二元响应变量传递给 glm 以执行逻辑回归。您可以将数据以串行数据格式传递给 glm(例如,每次观察一行,响应变量为 0 或 1,独立变量采用您拥有的任何值),或者您可以将其传递给作为一个table,至少有三列:第一列表示试验次数,第二列表示成功次数,第三列是自变量。

当我使用后一种数据格式(例如具有三列的数据框)使用 glm 时,我得到了预期的输出,但是当我使用前者(即串行数据格式)输入数据时,我没有得到预期的答案。

这是一个例子

prices <- c(89.99, 99.99, 149.99)
non_purchases <- c(11907, 2024, 5046)
purchases <- c(1369, 215, 31)
trials <- cbind(non_purchases, purchases)

model <- glm(trials ~ prices, family=binomial(link="logit"))

> summary(model)

Call:
glm(formula = trials ~ prices, family = binomial)

Deviance Residuals: 
     1       2       3  
 1.332  -4.440   1.553  

Coefficients:
             Estimate Std. Error z value Pr(>|z|)    
(Intercept) -1.923863   0.241677   -7.96 1.71e-15 ***
prices       0.044995   0.002593   17.35  < 2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

    Null deviance: 715.832  on 2  degrees of freedom
Residual deviance:  23.897  on 1  degrees of freedom
AIC: 49.228

Number of Fisher Scoring iterations: 4

在这种情况下,我得到了预期值,但是使用串行数据

> head(atable)
  ordered sale_price
1       0     149.99
2       0     149.99
3       0     149.99
4       0     149.99
5       0     149.99
6       0     149.99
> summary(atable)
    ordered          sale_price    
 Min.   :0.00000   Min.   : 89.99  
 1st Qu.:0.00000   1st Qu.: 89.99  
 Median :0.00000   Median : 89.99  
 Mean   :0.07843   Mean   :105.87  
 3rd Qu.:0.00000   3rd Qu.: 99.99  
 Max.   :1.00000   Max.   :149.99 

> conv_model <- glm(ordered ~ sale_price, family=binomial(link="logit"), data=atable)
> summary(conv_model)

Call:
glm(formula = ordered ~ sale_price, family = binomial(link = "logit"), 
    data = atable)

Deviance Residuals: 
    Min       1Q   Median       3Q      Max  
-0.4743  -0.4743  -0.4743  -0.1209   3.1376  

Coefficients:
             Estimate Std. Error z value Pr(>|z|)    
(Intercept)  0.549136   0.095341    5.76 8.43e-09 ***
sale_price  -0.019949   0.001002  -19.90  < 2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

    Null deviance: 11322  on 20591  degrees of freedom
Residual deviance: 10623  on 20590  degrees of freedom
AIC: 10627

Number of Fisher Scoring iterations: 7

并且只是为了表明它是相同的数据

> table(atable$ordered, atable$sale_price)

    89.99 99.99 149.99
  0 11907  2024   5046
  1  1369   215     31

我得到的输出完全不同,我完全糊涂了。谁能帮我吗?我假设我在做一些简单的事情

我认为您的问题是您正在切换 "success" 的定义。

来自 ?glm(强调我的)

For binomial and quasibinomial families the response can also be specified as... a two-column matrix with the columns giving the numbers of successes and failures.

所以第一列是 "successes"。在您的代码中,您使用 cbind(non_purchases, purchases),这使得 non_purchases 成为 "success" 列。但是在您的 table 中,非购买被编码为 0 表示失败。使用下面的代码,我们得到相同的结果:

prices <- c(89.99, 99.99, 149.99)
non_purchases <- c(11907, 2024, 5046)
purchases <- c(1369, 215, 31)
trials <- cbind(non_purchases, purchases)

dd = data.frame(
    price = c(rep(prices, non_purchases), rep(prices, purchases)),
    purchase = c(rep(0, sum(non_purchases)), rep(1, sum(purchases)))
)

coef(glm(purchase ~ price, data = dd, family = "binomial"))
# (Intercept)       price 
#  1.92386320 -0.04499477 

coef(glm(cbind(purchases, non_purchases) ~ prices, family = "binomial"))
# (Intercept)       price 
#  1.92386320 -0.04499477