如何在 R 中为 lm() 保留 fit$model 中的变量，我*不*在 lm 调用本身中使用它？

Question

I want to be able to index my model after having fit the model. Say I have

df <- data.frame(a = c(1,2,3), 
                 b = c(2,3,1000), 
                 country = c("Malawi", "USA","UK"))

Then, I run:

fit<-lm(a~b,data=df)

My resulting fit$model no longer has the "country" variable, so it becomes hard to do things like

• run a regression and then remove certain countries as robustness tests.
• run a regression and then find out which countries were outliers.

I know there are 'hacks' around this like using row indices, but I frequently find myself further subsetting the original dataset, and I am afraid of keeping track of row indices.

eg From the example above, I see that UK is an outlier.

So, I have two options:

lm(a~b,data=fit$model[-3,])
lm(a~b,data=df[df$country!="UK",])

The second option is much clearer to me, but because summary statistics and tests in R (such as cook's distance) only give me the row index , I end up having to do the first option much more than I would like. This becomes especially tedious in large panel datasets where I'm trying to test robustness to outliers or leveraged data and would also like to know what countries (or other variables) those data are.

Ideally, I'd like an option to do something like

lm(a~b,data=fit$model[fit$model$country!="UK",])

Please help, and thank you so much!

Answer 1

I am assuming the problem is to identify the rows of the original data frame from an lm model that was run on a subset of those rows where the subset run was performed without using all columns.

Regarding the characterization of row names in the question, I would not regard their use as negative at all. The row names are an intrinsic part of every data frame and are intended to identify the rows. If you do identify the rows with case names these case names will be shown by many functions including case.names(fm), model.frame(fm), model.matrix(fm), cooks.distance(fm), hatvalues(fm), influence(fm), plot(fm), etc. so it is highly desirable that they be used. This is really the way this software was intended to work so it is highly advisable to go with the case names approach to simplify everything.

1) Thus if the country names are unique identifiers of the cases then they can be maintained as case names by simply assigning them to the row names. We omitted USA to make the example harder owing to the fact that it does not come at the end like UK does and if we used UK maybe it would just give the first two case names anyways.

df <- data.frame(a = c(1,2,3),  b = c(2,3,1000), country = c("Malawi", "USA","UK"))

rownames(df) <- df$country

fm <- lm(a ~ b, df)
fm2 <- update(fm, subset = country != "USA")  # omit USA
# or:  update(fm, subset = case.names(fm) != "USA")

case.names(fm2)
## [1] "Malawi" "UK"    

2) Although (1) seems preferable another possibility that works even if we don't assign the country column to the row names is to look up the case names in the original data frame:

df <- data.frame(a = c(1,2,3),  b = c(2,3,1000), country = c("Malawi", "USA","UK"))

fm <- lm(a ~ b, df)
fm2 <- update(fm, subset = country != "USA")  # omit USA

df[ case.names(fm2), ]
##   a    b country
## 1 1    2  Malawi
## 3 3 1000      UK

or as a function:

# first arg is lm object
# second arg is full data frame - data frame used in lm call if unspecified
# third arg is envir where full data frame stored - current envir if unspecified
extractData <- function(mod, data, envir = parent.frame()) {
  if (missing(data)) data <- eval(mod$call$data, envir)
  data[ case.names(mod), ]
}

# test

extractData(fm2)
##   a    b country
## 1 1    2  Malawi
## 3 3 1000      UK