Date and POSIXct coerced to numeric when calculating median by group

[Henrik-P]

I have data with class Date ('date') and POSIXct ('time'), and a grouping variable 'g'

d <- data.table(
   date = as.Date(c("2018-01-01", "2018-01-03", "2018-01-08",
                 "2018-01-10", "2018-01-25", "2018-01-30")),
   g = rep(letters[1:2], each = 3))
d[ , time := as.POSIXct(date)]

When calculating median of 'date' and 'time' by group, the result is coerced to numeric:

d[ , median(date), by = g]
#    g    V1
# 1: a 17534
# 2: b 17556

d[ , median(time), by = g]
#    g         V1
# 1: a 1514937600
# 2: b 1516838400

---

However, 'date' and 'time' is not coerced when calculating median without grouping:

d[ , median(date)]
# [1] "2018-01-09"

d[ , median(time)]
# [1] "2018-01-09 01:00:00 CET"

---

Other things I've tried which don't coerce:

Mean 'date' and 'time' by group:

d[ , mean(date), by = g]
#    g         V1
# 1: a 2018-01-04
# 2: b 2018-01-21

d[ , mean(time), by = g]
#    g                  V1
# 1: a 2018-01-04 01:00:00
# 2: b 2018-01-21 17:00:00

Median 'date' and 'time' by group using aggregate:

aggregate(date ~ g, data = d, median)
#   g       date
# 1 a 2018-01-03
# 2 b 2018-01-25

aggregate(time ~ g, data = d, median)
#   g                time
# 1 a 2018-01-03 01:00:00
# 2 b 2018-01-25 01:00:00

---

R version 3.5.1 (2018-07-02)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows Server >= 2012 x64 (build 9200)
data.table_1.11.6

[MichaelChirico]

re-run with verbose = TRUE

…

On Fri, Sep 28, 2018, 12:48 AM Henrik-P ***@***.***> wrote: I have data with class Date ('date') and POSIXct ('time'), and a grouping variable 'g' d <- data.table( date = as.Date(c("2018-01-01", "2018-01-03", "2018-01-08", "2018-01-10", "2018-01-25", "2018-01-30")), g = rep(letters[1:2], each = 3)) d[ , time := as.POSIXct(date)] When calculating median of 'date' and 'time' by group, the result is coerced to numeric: d[ , median(date), by = g] # g V1 # 1: a 17534 # 2: b 17556 d[ , median(time), by = g] # g V1 # 1: a 1514937600 # 2: b 1516838400 ------------------------------ However, 'date' and 'time' is *not* coerced when calculating median *without* grouping: d[ , median(date)] # [1] "2018-01-09" d[ , median(time)] # [1] "2018-01-09 01:00:00 CET" ------------------------------ Other things I've tried which don't coerce: Mean 'date' and 'time' by group: d[ , mean(date), by = g] # g V1 # 1: a 2018-01-04 # 2: b 2018-01-21 d[ , mean(time), by = g] # g V1 # 1: a 2018-01-04 01:00:00 # 2: b 2018-01-21 17:00:00 Median 'date' and 'time' by group using aggregate: aggregate(date ~ g, data = d, median) # g date # 1 a 2018-01-03 # 2 b 2018-01-25 aggregate(time ~ g, data = d, median) # g time # 1 a 2018-01-03 01:00:00 # 2 b 2018-01-25 01:00:00 ------------------------------ R version 3.5.1 (2018-07-02) Platform: x86_64-w64-mingw32/x64 (64-bit) Running under: Windows Server >= 2012 x64 (build 9200) data.table_1.11.6 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#3079>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHQQdQkQW1TkuszM-3S7lwHoviAOkinEks5ufQFNgaJpZM4W9Dmz> .

[Henrik-P]

Thanks Michael. Here we go:

d[ , median(date), by = g, verbose = TRUE]

Detected that j uses these columns: date
Finding groups using forderv ... 0.000sec
Finding group sizes from the positions (can be avoided to save RAM) ... 0.000sec
lapply optimization is on, j unchanged as 'median(date)'
GForce optimized j to 'gmedian(date)'
Making each group and running j (GForce TRUE) ... 0.020sec
g V1
1: a 17534
2: b 17556

---

d[ , median(time), by = g, verbose = TRUE]

Detected that j uses these columns: time
Finding groups using forderv ... 0.000sec
Finding group sizes from the positions (can be avoided to save RAM) ... 0.000sec
lapply optimization is on, j unchanged as 'median(time)'
GForce optimized j to 'gmedian(time)'
Making each group and running j (GForce TRUE) ... 0.000sec
g V1
1: a 1514937600
2: b 1516838400

[MichaelChirico]

Thanks... I suspected GForce was at root here but am still surprised that mean works (verbose = TRUE for those shows it's still activated)....

For now, if you're just interested in moving on, you can temporarily disable GForce:

old = getOption('datatable.optimize')
options(datatable.optimize = 1L)
d[ , median(time), by = g]
#    g                  V1
# 1: a 2018-01-03 08:00:00
# 2: b 2018-01-25 08:00:00
options(datatable.optimize = old)

[Henrik-P]

Thanks a lot for your rapid response. I think this may be a regression - as far as I recall this used to work in earlier versions.

[MichaelChirico]

It certainly sounds familiar... there may be an outstanding issue...

[MichaelChirico]

Very similar to #1876

[franknarf1]

Another bug, maybe related -- gmedian is coercing integers to reals, in contrast with base median:

> d[, idate := as.IDate(date)]
> d[, dput(median(idate)), by=g]
structure(17534L, class = c("IDate", "Date"))
structure(17556L, class = c("IDate", "Date"))
   g         V1
1: a 2018-01-03
2: b 2018-01-25
> d[, median(idate), by=g][, dput(V1)]
c(17534, 17556)
[1] 17534 17556

I guess one could argue that the base median behavior is wrong (since return type is unpredictable)

[jangorecki]

Very rarely I'm against consistency with base R but in this particular case of gmedian I prefer to have double returned always.