Error when using broom (augment) and dplyr with lo

问题:

I am trying to use augment on a loess fit, but I receive the following error:

Error in data.frame(..., check.names = FALSE) : 
  arguments imply differing number of rows: 32, 11

In the error message, 11 happens to equal the number of observations in one segment and 32 is the total number of observations. The code is below.

require(broom)
require(dplyr)

# This example uses the lm method and it works
regressions <- mtcars %>% group_by(cyl) %>%  do(fit = lm(wt ~ mpg, .))
regressions %>% augment(fit)

# This example uses the loess method and it generates the error
regressions2 <- mtcars %>% group_by(cyl) %>%  do(fit = loess(wt ~ mpg, .))
regressions2 %>% augment(fit)

# The below code appropriately plots the loess fit using geom_smooth. 
# My current # workaround is to do a global definition as an aes object in geom_smooth`
cylc = unique(mtcars$cyl) %>% sort()
for (i in 1:length(cyl)){
  print(i)
  print(cyl[i])
  p<- ggplot(data=filter(mtcars,cyl==cylc[i]),aes(x=mpg,y=wt)) + geom_point() + geom_smooth(method="loess") + ggtitle(str_c("cyl = ",cyl[i]))
  print(p)
}

回答1:

This appears to be a problem related to the do() operator: when we check the model.frame() on one of the LOESS model objects, we get back all 32 rows rather than the subset corresponding to that model.

A workaround is to hold on to the data and not just the model, and pass that as the second argument to augment():

regressions2 <- mtcars %>%
  group_by(cyl) %>%
  do(fit = loess(wt ~ mpg, .),
     data = (.)) %>%
   augment(fit, data)

This is usually recommended with augment() anyway, since model.frame() doesn't get all the original columns.

---

Incidentally, I'm the maintainer of broom and I generally no longer recommend the do() approach (since dplyr has mostly been moving away from it).

Instead, I suggest using tidyr's nest() and purrr's map(), as described in this chapter of R4DS. This makes it a little bit easier to hold on to the data and incorporate into augment().

library(tidyr)
library(purrr)

mtcars %>%
  nest(-cyl) %>%
  mutate(fit = map(data, ~ loess(wt ~ mpg, .))) %>%
  unnest(map2(fit, data, augment))