Fill in missing values by group in data.table

Question

If one wants to fill in missing values of a variable based on previous/posterior non NA observation within a group, the data.table command is

setkey(DT,id,date)
DT[, value_filled_in := DT[!is.na(value), list(id, date, value)][DT[, list(id, date)], value, roll = TRUE]]

which is quite complex. It's a shame since roll is a very fast and powerful option (esp compared with applying a function such as zoo::na.locf within each group)

I can write a convenience function to fill in missing values

   fill_na <-  function(x , by = NULL, roll =TRUE , rollends= if (roll=="nearest") c(TRUE,TRUE)
             else if (roll>=0) c(FALSE,TRUE)
             else c(TRUE,FALSE)){
    id <- seq_along(x)
    if (is.null(by)){
      DT <- data.table("x" = x, "id" = id, key = "id") 
      return(DT[!is.na(x)][DT[, list(id)], x, roll = roll, rollends = rollends, allow.cartesian = TRUE])

    } else{
      DT <- data.table("x" = x, "by" = by, "id" = id, key = c("by", "id")) 
      return(DT[!is.na(x)][DT[, list(by, id)], x, roll = roll, rollends = rollends, allow.cartesian = TRUE])
    }
  }

And then write

setkey(DT,id, date)
DT[, value_filled_in := fill_na(value, by = id)]

This is not really satisfying since one would like to write

setkey(DT,id, date)
DT[, value_filled_in := fill_na(value), by = id]

However, this takes a huge amount of time to run. And, for the end-user, it is cumbersome to learn that fill_na should be called with the by option, and should not be used with data.table by. Is there an elegant solution around this?

Some speed test

N <- 2e6
set.seed(1)
DT <- data.table(
         date = sample(10, N, TRUE),
           id = sample(1e5, N, TRUE),   
        value = sample(c(NA,1:5), N, TRUE),
       value2 = sample(c(NA,1:5), N, TRUE)                   
      )
setkey(DT,id,date)
DT<- unique(DT)

system.time(DT[, filled0 := DT[!is.na(value), list(id, date, value)][DT[, list(id, date)], value, roll = TRUE]])
#> user  system elapsed 
#>  0.086   0.006   0.105 
system.time(DT[, filled1 := zoo::na.locf.default(value, na.rm = FALSE), by = id])
#> user  system elapsed 
#> 5.235   0.016   5.274 
# (lower speed and no built in option like roll=integer or roll=nearest, rollend, etc)
system.time(DT[, filled2 := fill_na(value, by = id)])
#>   user  system elapsed 
#>  0.194   0.019   0.221 
system.time(DT[, filled3 := fill_na(value), by = id])
#>    user  system elapsed 
#> 237.256   0.913 238.405

Why don't I just use na.locf.default ? Even though the speed difference is not really important, the same issue arises for other kinds of data.table commands (those that rely on a merge by the variable in "by") - it's a shame to systematically ignore them in order to get an easier syntax. I also really like all the roll options.

How does the na.locf solution compare to this solution in terms of speed? — GSee
Is wrapping the entire thing (a la dplyr::mutate) not an option? — shadowtalker
It would be helpful if you provided code to create a sample data.table that we could use to check our results and to help with benchmarking. — GSee
If you get rid of the :: call in the zoo one, it's about 30% faster for me. i.e. call na.locf.default instead of zoo::na.locf.default — GSee
@ssdecontrol Because :: is a function and there is overhead associated with that extra function call. — GSee

eddi eddi · Accepted Answer · 2014-10-03T15:16:16

Here's a slightly faster and more compact way of doing it (version 1.9.3+):

DT[, filled4 := DT[!is.na(value)][DT, value, roll = T]]

Fill in missing values by group in data.table

2 Answers