Efficiently removing missing values from the start and end of multiple time series in 1 data frame

Question

Using R, I'm trying to trim NA values from the start and end of a data frame that contains multiple time series. I have achieved my goal using a for loop and the zoo package, but as expected it is extremely inefficient on large data frames.

My data frame look like this and contains 3 columns with each time series identified by it's unique id. In this case AAA, B and CCC.

id   date          value
AAA  2010/01/01    NA
AAA  2010/02/01    34
AAA  2010/03/01    35
AAA  2010/04/01    30
AAA  2010/05/01    NA
AAA  2010/06/01    28
B    2010/01/01    NA
B    2010/02/01    0
B    2010/03/01    1
B    2010/04/01    2
B    2010/05/01    3
B    2010/06/01    NA
B    2010/07/01    NA
B    2010/07/01    NA
CCC  2010/01/01    0
CCC  2010/02/01    400
CCC  2010/03/01    300
CCC  2010/04/01    200
CCC  2010/05/01    NA

I would like to know, how can I efficiently remove the NA values from the start and end of each time series, in this case AAA, B and CCC. So it should look like this.

id   date          value
AAA  2010/02/01    34
AAA  2010/03/01    35
AAA  2010/04/01    30
AAA  2010/05/01    NA
AAA  2010/06/01    28
B    2010/02/01    0
B    2010/03/01    1
B    2010/04/01    2
B    2010/05/01    3
CCC  2010/01/01    0
CCC  2010/02/01    400
CCC  2010/03/01    300
CCC  2010/04/01    200

Answer 1

I would do it like this, which should be very fast :

require(data.table)
DT = as.data.table(your data)   # please provide something pastable

DT2 = DT[!is.na(value)]
setkey(DT,id,date)
setkey(DT2,id,date)
tokeep = DT2[DT,!is.na(value),rolltolast=TRUE,mult="last"]
DT = DT[tokeep]

This works by rolling forward the prevailing non-NA, but not past the last one, within each group.

The mult="last" is optional. It should speed it up if v1.8.0 (on CRAN) is used. Interested in timings with and without it. By default data.table joins to groups (mult="all"), but in this case we're joining to all columns of the key, and, we know the key is unique; i.e., no dups in key. In v1.8.1 (in dev) there isn't a need to know about this and it looks after you more.