使用bootstrap进行分位数回归的置信区间

时间:2017-09-24 23:50:14

标签: r bootstrapping confidence-interval quantile

我试图获得线性和分位数回归的五种类型的自举间隔。我能够使用 car 中的Boot和 boot 中的boot.ci进行自举并找到5个boostrap间隔(分位数,正常,基本,学生化和BCa)进行线性回归。当我尝试使用来自 quantreg 的rq进行分位数回归时,它会引发错误。这是示例代码

创建模型

library(car)
library(quantreg)
library(boot)
newdata = Prestige[,c(1:4)]
education.c = scale(newdata$education, center=TRUE, scale=FALSE)
prestige.c = scale(newdata$prestige, center=TRUE, scale=FALSE)
women.c = scale(newdata$women, center=TRUE, scale=FALSE)
new.c.vars = cbind(education.c, prestige.c, women.c)
newdata = cbind(newdata, new.c.vars)
names(newdata)[5:7] = c("education.c", "prestige.c", "women.c" )
mod1 = lm(income ~ education.c + prestige.c + women.c, data=newdata)
mod2 = rq(income ~ education.c + prestige.c + women.c, data=newdata)

引导线性和分位数回归

mod1.boot <- Boot(mod1, R=999)
boot.ci(mod1.boot, level = .95, type = "all")
dat2 <- newdata[5:7]
mod2.boot <- boot.rq(cbind(1,dat2),newdata$income,tau=0.5, R=10000)
boot.ci(mod2.boot, level = .95, type = "all")
Error in if (ncol(boot.out$t) < max(index)) { : 
argument is of length zero

1)为什么boot.ci不适用于分位数回归

2)使用我从stackexchange获得的这个解决方案,我能够找到分位数CI。

rq

的分位数(百分位CI)的解决方案
t(apply(mod2.boot$B, 2, quantile, c(0.025,0.975)))

我如何获得其他CI for bootstrap(normal,basic,studentized,BCa)。

3)另外,我的用于线性回归的boot.ci命令会产生此警告

Warning message:
In sqrt(tv[, 2L]) : NaNs produced

这意味着什么?

2 个答案:

答案 0 :(得分:1)

使用summary.rq可以计算模型系数的boostrap标准误差。
有五种boostrap方法(bsmethods)可用(参见?boot.rq)。

summary(mod2, se = "boot", bsmethod= "xy")

# Call: rq(formula = income ~ education.c + prestige.c + women.c, data = newdata)
# 
# tau: [1] 0.5
#  
# Coefficients:
#             Value      Std. Error t value    Pr(>|t|)  
# (Intercept) 6542.83599  139.54002   46.88860    0.00000
# education.c  291.57468  117.03314    2.49139    0.01440
# prestige.c    89.68050   22.03406    4.07009    0.00010
# women.c      -48.94856    5.79470   -8.44712    0.00000

要计算瓶底置信区间,您可以使用以下技巧:

mod1.boot <- Boot(mod1, R=999)
set.seed(1234)
boot.ci(mod1.boot, level = .95, type = "all")

dat2 <- newdata[5:7]
set.seed(1234)
mod2.boot <- boot.rq(cbind(1,dat2),newdata$income,tau=0.5, R=10000)

# Create an object with the same structure of mod1.boot
# but with boostrap replicates given by boot.rq
mod3.boot <- mod1.boot
mod3.boot$R <- 10000
mod3.boot$t0 <- coef(mod2)
mod3.boot$t <- mod2.boot$B
boot.ci(mod3.boot, level = .95, type = "all")

# BOOTSTRAP CONFIDENCE INTERVAL CALCULATIONS
# Based on 10000 bootstrap replicates
# 
# CALL : 
# boot.ci(boot.out = mod3.boot, type = "all", level = 0.95)
# 
# Intervals : 
# Level      Normal              Basic             Studentized     
# 95%   (6293, 6838 )   (6313, 6827 )   (6289, 6941 )  
# 
# Level     Percentile            BCa          
# 95%   (6258, 6772 )   (6275, 6801 )  

答案 1 :(得分:0)

感谢所有帮助过的人。我自己能够找到解决方案。我运行了一个计算分位数回归系数的循环,然后分别使用boot和boot.ci。这是代码

仅引导命令,从问题

创建模型
mod3 <- formula(income ~ education.c + prestige.c + women.c)
coefsf <- function(data,ind){
rq(mod3, data=newdata[ind,])$coef
}
boot.mod <- boot(newdata,coefsf,R=10000)
myboot.ci <- list()
for (i in 1:ncol(boot.mod$t)){
myboot.ci[[i]] <- boot.ci(boot.mod, level = .95, type = 
c("norm","basic","perc", "bca"),index = i)
  }

我这样做是因为我想要所有变量上的CI而不仅仅是截距。