R：使用miceadds lm.cluster的聚类稳健标准错误 - 子集和权重的错误

Question

我正在尝试使用miceadds 包中的lm.cluster 函数来为多重插补数据集获取稳健的聚类标准误差。

我能够运行它的标准版本，但是当我尝试添加子集或权重时出现以下错误：

Error in eval(substitute(subset), data, env) : 
..1 used in an incorrect context, no ... to look in

没有子集或权重的示例：

require("mice")
require("miceadds")
data(data.ma01)
# imputation of the dataset: use six imputations
dat <- data.ma01[ , - c(1:2) ]
imp <- mice::mice( dat , maxit=3 , m=6 )
datlist <- miceadds::mids2datlist( imp )
# linear regression with cluster robust standard errors
mod <- lapply(datlist, FUN = function(data){miceadds::lm.cluster( data=data ,         
formula=read ~ paredu+ female ,  cluster = data.ma01$idschool )}  )

# extract parameters and covariance matrix
betas <- lapply( mod , FUN = function(rr){ coef(rr) } )
vars <- lapply( mod , FUN = function(rr){ vcov(rr) } )
# conduct statistical inference
summary(pool_mi( qhat = betas, u = vars ))

与子集中断的示例：

mod <- lapply(datlist, FUN = function(data){miceadds::lm.cluster( data=data ,         
formula=read ~ paredu+ female ,  cluster = data.ma01$idschool, subset=
(data.ma01$urban==1))}  )

Error during wrapup: ..1 used in an incorrect context, no ... to look in

用权重打断的例子：

mod <- lapply(datlist, FUN = function(data){miceadds::lm.cluster( data=data ,         
formula=read ~ paredu+ female ,  cluster = data.ma01$idschool,
weights=data.ma01$studwgt)}  )

Error during wrapup: ..1 used in an incorrect context, no ... to look in

通过搜索，我认为我在通过 lm 或 glm 包装器传递这些命令时遇到了与其他人类似的问题（例如：Passing Argument to lm in R within Function 或 R : Pass argument to glm inside an R function 或 Passing the weights argument to a regression function inside an R function）

但是，我不确定如何解决估算数据集和现有 lm.cluster 命令的问题。

谢谢

Answer 1

这适用于 CRAN 上的 estimatr 包和 estimatr::lm_robust() 函数。两个注意事项：（1）您可以使用 se_type = 更改标准错误的类型，（2）我将 idschool 保留在数据中，因为我们希望集群在相同的 data.frame 中，因为我们适合模型。

library(mice)
library(miceadds)
library(estimatr)

# imputation of the dataset: use six imputations
data(data.ma01)
dat <- data.ma01[, -c(1)] # note I keep idschool in data
imp <- mice::mice( dat , maxit = 3, m = 6)
datlist <- miceadds::mids2datlist(imp)

# linear regression with cluster robust standard errors
mod <- lapply(
  datlist, 
  function (dat) {
    estimatr::lm_robust(read ~ paredu + female, dat, clusters = idschool)
  }
)

# subset
mod <- lapply(
  datlist, 
  function (dat) {
    estimatr::lm_robust(read ~ paredu + female, dat, clusters = idschool, subset = urban == 1)
  }
)

# weights
mod <- lapply(
  datlist, 
  function (dat) {
    estimatr::lm_robust(read ~ paredu + female, dat, clusters = idschool, weights = studwgt)
  }
)

# note that you can use the `se_type` argument of lm_robust() 
# to change the vcov estimation

# extract parameters and covariance matrix
betas <- lapply(mod, coef)
vars <- lapply(mod, vcov)
# conduct statistical inference
summary(pool_mi( qhat = betas, u = vars ))

Answer 2

我不是专家，但是将权重传递给lm() 存在问题。我知道这不是一个理想的情况，但我设法通过修改 lm.cluster() 函数来硬编码权重传递，然后使用我自己的方法来让它工作。

lm.cluster <- function (data, formula, cluster, wgts=NULL, ...) 
{
  TAM::require_namespace_msg("multiwayvcov")
  if(is.null(wgts)) {
    mod <- stats::lm(data = data, formula = formula)
  } else {
    data$.weights <- wgts
    mod <- stats::lm(data = data, formula = formula, weights=data$.weights)
  }
  if (length(cluster) > 1) {
    v1 <- cluster
  }
  else {
    v1 <- data[, cluster]
  }
  dfr <- data.frame(cluster = v1)
  vcov2 <- multiwayvcov::cluster.vcov(model = mod, cluster = dfr)
  res <- list(lm_res = mod, vcov = vcov2)
  class(res) <- "lm.cluster"
  return(res)
}