我不知道解析解决方案,所以我会在 R 中使用包中的块引导boot
程序boot
。
这是我将作为基准的 Stata 代码。
cd "C:\Users\Richard\Desktop\"
use "http://www.ats.ucla.edu/stat/stata/dae/binary.dta", clear
generate group = int((_n - 1) / 20) + 1
probit admit gpa gre, vce(cluster group)
outsheet using "binary.txt", replace
这是 R 中的等价物。第二个块提供了块 bootstrap on group
,这是我在 Stata 中创建的随机聚类变量。
setwd("C:/Users/Richard/Desktop/")
df <- read.delim("binary.txt")
# homoskedastic
probit <- glm(admit ~ gpa + gre, data=df, family=binomial(link=probit))
# with block bootstrap using `boot` package
library(boot)
myProbit <- function(x, y) {
myDf <- do.call("rbind", lapply(y, function(n) subset(df, group == x[n])))
myModel <- glm(admit ~ gpa + gre, data=myDf, family=binomial(link=probit))
coefficients(myModel)
}
groups <- unique(df$group)
probitBS <- boot(groups, myProbit, 500)
# comparison
summary(probit)
probitBS
它们非常接近(Stata 结果,然后是 R 块引导结果)。
Probit regression Number of obs = 400
Wald chi2(2) = 24.03
Prob > chi2 = 0.0000
Log pseudolikelihood = -240.094 Pseudo R2 = 0.0396
(Std. Err. adjusted for 20 clusters in group)
------------------------------------------------------------------------------
| Robust
admit | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
gpa | .454575 .1531717 2.97 0.003 .1543641 .7547859
gre | .0016425 .0006404 2.56 0.010 .0003873 .0028977
_cons | -3.003536 .520864 -5.77 0.000 -4.024411 -1.982662
------------------------------------------------------------------------------
> probitBS
ORDINARY NONPARAMETRIC BOOTSTRAP
Call:
boot(data = groups, statistic = myProbit, R = 500)
Bootstrap Statistics :
original bias std. error
t1* -3.003535745 -3.976856e-02 0.5420935780
t2* 0.454574799 3.781773e-03 0.1530609943
t3* 0.001642537 4.200797e-05 0.0006210689