R predict.glm when newdata 具有较少的级别

R predict.glm when newdata has fewer levels

我试图向自己证明,当 newdata 的标签和水平(因子水平的基础整数)与训练数据不匹配时,predict() 不会给出错误的预测.

我想我确实证明了这一点,我在下面分享了该代码,但我只想问一下 R 在预测 newdata 时到底在做什么。我知道它不是将 newdata 附加到训练数据,它是否在预测之前将 newdata 的因子标签转换为训练数据的相应表示?

options(stringsAsFactors = TRUE)
dat <- data.frame(x = rep(c("cat", "dog", "bird", "horse"), 100), y = rgamma(100, shape=3,  scale = 300))
model <- glm(y~., family = Gamma(link = "log"), data = dat)

coefficients(model)
# (Intercept)        xcat        xdog      xhorse 
#   6.5816536   0.2924488   0.3586094   0.2740487 

newdata1 <- data.frame(x = "cat")
newdata2 <- data.frame(x = "bird")
newdata3 <- data.frame(x = "dog")

predict.glm(object = model, newdata = newdata1, type = "response")
#       1 
# 966.907 
exp(6.5816536 + 0.2924488) #intercept + cat coef
# [1] 966.9071

predict.glm(object = model, newdata = newdata2, type = "response")
#        1 
# 721.7318 
exp(6.5816536)
# [1] 721.7318

predict.glm(object = model, newdata = newdata3, type = "response")
#        1 
# 1033.042 
exp(6.5816536 + 0.3586094)
# [1] 1033.042

unclass(dat$x)
#  [1] 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3
# [87] 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4
# [173] 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3
# [259] 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4
# [345] 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4 2 3 1 4
# attr(,"levels")
# [1] "bird"  "cat"   "dog"   "horse"

unclass(newdata1$x)
# [1] 1
# attr(,"levels")
# [1] "cat"

unclass(newdata2$x)
# [1] 1
# attr(,"levels")
# [1] "bird"

模型对象有一个 xlevels 记录因子水平用于模型估计。对于您的示例,我们有:

model$xlevels
#$x
#[1] "bird"  "cat"   "dog"   "horse"

当您的新数据出现在预测中时,将匹配因子水平。例如,您的 newdata1 将匹配到 "cat" 个级别,这是 xlevels 中的第二个级别。因此,predict 将毫不费力地找到该级别的正确系数。