r - replace NA in a dplyr chain

Question

Question has been edited from the original.

After reading this interesting discussion I was wondering how to replace NAs in a column using dplyr in, for example, the Lahman batting data:

Source: local data frame [96,600 x 3]
Groups: teamID

   yearID teamID G_batting
1    2004    SFN        11
2    2006    CHN        43
3    2007    CHA         2
4    2008    BOS         5
5    2009    SEA         3
6    2010    SEA         4
7    2012    NYA        NA

The following does not work as I expected

library(dplyr)
library(Lahman)

df <- Batting[ c("yearID", "teamID", "G_batting") ]
df <- group_by(df, teamID )
df$G_batting[is.na(df$G_batting)] <- mean(df$G_batting, na.rm = TRUE)

Source: local data frame [20 x 3] Groups: yearID, teamID

   yearID teamID G_batting
1    2004    SFN  11.00000
2    2006    CHN  43.00000
3    2007    CHA   2.00000
4    2008    BOS   5.00000
5    2009    SEA   3.00000
6    2010    SEA   4.00000
7    2012    NYA  **49.07894**

> mean(Batting$G_battin, na.rm = TRUE)
[1] **49.07894**

In fact it imputed the overall mean and not the group mean. How would you do this in a dplyr chain? Using transform from base R also does not work as it imputed the overall mean and not the group mean. Also this approach converts the data to a regular dat. a frame. Is there a better way to do this?

df %.% 
  group_by( yearID ) %.%
  transform(G_batting = ifelse(is.na(G_batting), 
    mean(G_batting, na.rm = TRUE), 
    G_batting)
  )

Edit: Replacing transform with mutate gives the following error

Error in mutate_impl(.data, named_dots(...), environment()) : 
  INTEGER() can only be applied to a 'integer', not a 'double'

Edit: Adding as.integer seems to resolve the error and does produce the expected result. See also @eddi's answer.

df %.% 
  group_by( teamID ) %.%
  mutate(G_batting = ifelse(is.na(G_batting), as.integer(mean(G_batting, na.rm = TRUE)), G_batting))

Source: local data frame [96,600 x 3]
Groups: teamID

   yearID teamID G_batting
1    2004    SFN        11
2    2006    CHN        43
3    2007    CHA         2
4    2008    BOS         5
5    2009    SEA         3
6    2010    SEA         4
7    2012    NYA        47

> mean_NYA <- mean(filter(df, teamID == "NYA")$G_batting, na.rm = TRUE)
> as.integer(mean_NYA)
[1] 47

Edit: Following up on @Romain's comment I installed dplyr from github:

> head(df,10)
   yearID teamID G_batting
1    2004    SFN        11
2    2006    CHN        43
3    2007    CHA         2
4    2008    BOS         5
5    2009    SEA         3
6    2010    SEA         4
7    2012    NYA        NA
8    1954    ML1       122
9    1955    ML1       153
10   1956    ML1       153

> df %.% 
+   group_by(teamID)  %.%
+   mutate(G_batting = ifelse(is.na(G_batting), mean(G_batting, na.rm = TRUE), G_batting))
Source: local data frame [96,600 x 3]
Groups: teamID

   yearID teamID  G_batting
1    2004    SFN          0
2    2006    CHN          0
3    2007    CHA          0
4    2008    BOS          0
5    2009    SEA          0
6    2010    SEA 1074266112
7    2012    NYA   90693125
8    1954    ML1        122
9    1955    ML1        153
10   1956    ML1        153
..    ...    ...        ...

So I didn't get the error (good) but I got a (seemingly) strange result.

question from:https://stackoverflow.com/questions/21714867/replace-na-in-a-dplyr-chain

Answer 1

The main issue you're having is that mean returns a double while the G_batting column is an integer. So wrapping the mean in as.integer would work, or you'd need to convert the entire column to numeric I guess.

That said, here are a couple of data.table alternatives - I didn't check which one is faster.

library(data.table)

# using ifelse
dt = data.table(a = 1:2, b = c(1,2,NA,NA,3,4,5,6,7,8))
dt[, b := ifelse(is.na(b), mean(b, na.rm = T), b), by = a]

# using a temporary column
dt = data.table(a = 1:2, b = c(1,2,NA,NA,3,4,5,6,7,8))
dt[, b.mean := mean(b, na.rm = T), by = a][is.na(b), b := b.mean][, b.mean := NULL]

And this is what I'd want to do ideally (there is an FR about this):

# again, atm this is pure fantasy and will not work
dt[, b[is.na(b)] := mean(b, na.rm = T), by = a]

---

The dplyr version of the ifelse is (as in OP):

dt %>% group_by(a) %>% mutate(b = ifelse(is.na(b), mean(b, na.rm = T), b))

I'm not sure how to implement the second data.table idea in a single line in dplyr. I'm also not sure how you can stop dplyr from scrambling/ordering the data (aside from creating an index column).