仅使用模型对象重现二项式glm

时间:2014-10-24 10:39:22

标签: r statistics glm

我正在尝试重现R中二项式glm的结果。

考虑来自此处的数据http://www.ats.ucla.edu/stat/r/dae/logit.htm

mydata <- read.csv("http://www.ats.ucla.edu/stat/data/binary.csv")

我可以使用以下方式拟合模型:

model <- glm(admit ~ gre + gpa + rank, data = mydata, family = "binomial")

并且,仅使用对象重现模型:

model_r <- glm(as.numeric(model$y)~0+model.matrix(model), family = binomial)
cbind(coef(model), coef(model_r))
##                    [,1]        [,2]
## (Intercept) -3.44954840 -3.44954840
## gre          0.00229396  0.00229396
## gpa          0.77701357  0.77701357
## rank        -0.56003139 -0.56003139

现在假设专栏admit是该专栏中n篇论文中成功的数量:

mydata$n <- 1 + rbinom(n = 400, size = 2, prob = 0.5)

现在我必须使用以下方式拟合模型:

model <- glm(cbind(admit, n-admit) ~ gre + gpa + rank, data = mydata, 
                family = "binomial")

如何使用模型对象重现此模型?我问这个是因为R只保留model$y中的成功率。

1 个答案:

答案 0 :(得分:6)

您可以通过以下方式重现模型:

model_r <- glm(formula(model), model$data, family = family(model))

比较:

cbind(coef(model), coef(model_r))
#                     [,1]         [,2]
# (Intercept) -3.693688931 -3.693688931
# gre          0.001855502  0.001855502
# gpa          0.584915067  0.584915067
# rank        -0.450051862 -0.450051862

或者(类似于您的方法):

model_r2 <- glm(model.frame(model)[[1]] ~ 0 + model.matrix(model), 
                family = family(model))

cbind(coef(model), coef(model_r2))
#                     [,1]         [,2]
# (Intercept) -3.693688931 -3.693688931
# gre          0.001855502  0.001855502
# gpa          0.584915067  0.584915067
# rank        -0.450051862 -0.450051862