根据带有data.table解决方案的id将观察分组到指定数量的组中

Question

I have the following data.table:

dt <- data.table(id = rep(1:5, 5), obs = rnorm(1, n = 25))[order(id)]
dt 

   id      obs
1:  1  0.1470735
2:  1  1.6954685
3:  1  2.3947260
4:  1  2.1782338
5:  1  0.5168873
6:  2 -0.8879545
7:  2  1.9320034
8:  2  2.6269272
9:  2  1.5212627
10: 2 -0.1581711

Which has a total of 5 distinct ids (numbers 1 through 5) and 5 observations (obs) for each id. I want to group the ids together randomly in groups of X ids according to id and create a new column with the grouping. For this example, let's say I want to end up with a data.table like this:

   id      obs      group
1:  1  0.1470735      A
2:  1  1.6954685      A
3:  1  2.3947260      A
4:  1  2.1782338      A
5:  1  0.5168873      A
6:  2 -0.8879545      A
7:  2  1.9320034      A
8:  2  2.6269272      A
9:  2  1.5212627      A
10: 2 -0.1581711      A

Where ids 1 and 2 are assigned to group A, ids 3 and 4 are assigned to group B, and id 5 is assigned to group C.

My actual dataset is much larger and will not necessarily group evenly, but I do not need the groups to contain the same number of ids. I do need to control the general size of the group (for example I want to be able to say 5 ids per group and if the last group has only 3 ids that's fine).

Could someone please help me with an elegant data.table way to accomplish this?

Answer 1

This is the same as @Shree's answer, just using length.out in rep and no dplyr.

I do need to control the general size of the group (for example I want to be able to say 5 ids per group and if the last group has only 3 ids that's fine).

You can make an id table; assign groups there; and if necessary merge back:

# bigger, reproducible example
library(data.table)
max_per_group = 5
n_ids = 1e5+1
DT = data.table(id = rep(1:nid, each = max_per_group), obs = 1)

# make an id table
idDT = unique(DT[, "id"])

# randomly assign groups
idDT[, g := sample(rep(.I, each = 5, length.out = .N))]

# merge back if needed
DT[idDT, on=.(id), g := i.g]

You refer to "my actual dataset" -- but R allows you to juggle multiple tables. Trying to do everything in one is almost always counterproductive.

Answer 2

EDIT: Didn't notice that you needed this with data.table . I'll leave this out here as an alternative.

I am creating a dataframe with id and randomly assigned group. This will be joined with your data to get groups for each record by id -

library(dplyr)
library(data.table)

dt <- data.table(id = rep(1:5, 5), obs = rnorm(1, n = 25))[order(id)]

max_per_group <- 5
n_ids <- length(unique(dt$id))

data.frame(id = unique(dt$id), grp = sample(rep(LETTERS, max_per_group), n_ids)) %>%
  left_join(dt, ., by = "id")

   id         obs grp
1   1  1.28879713   S
2   1  1.04471197   S
3   1  0.36470847   S
4   1  0.46741567   S
5   1  1.07749891   S
6   2  1.73640785   K
7   2  1.61144042   K
8   2  2.85196859   K
9   2  1.84848117   K
10  2  2.11395863   K
11  3  0.88623462   S
12  3  2.11706351   S
13  3  1.29225433   S
14  3  0.30458037   S
15  3 -1.72070005   S
16  4  2.24593162   U
17  4  2.10346287   U
18  4  2.28724412   U
19  4  0.02978044   U
20  4  0.56234660   U
21  5  2.92050008   F
22  5  1.08048974   F
23  5  0.58885261   F
24  5  1.53299092   F
25  5  1.47271123   F