使用data.table避免按组内的最后n个观察值的滚动总和中的NA

Question

According to this threat I learned, rolling sums for variable b in the following data.table can be achieved as follows:

data creation + computing rolling sums:

    x <- data.table(a = sample(letters[1:3], 100, replace = TRUE), b = runif(100))
    setorder(x, a)

    # alternative 1
    x[, .(b, Reduce(`+`, shift(b, 0:2))), by = a]
    # alternative 2
    x[, .(b, stats::filter(b, rep(1, 3), sides = 1)), by = a]

Current + desired output:

     a           b        V2   V2_desired
  1: a 0.457665568        NA   0.457665568
  2: a 0.752555834        NA   1.210221
  3: a 0.864672124 2.0748935   2.0748935
  4: a 0.542168656 2.1593966   2.1593966
  5: a 0.197962875 1.6048037   1.6048037

Now there are NAs generated for the first two obs. in each by group. I need to adjust one of the alternatives to sum only the current obs. (last two obs.) in cases where the group index starts (is at position 2). This should be generalizable such that I could consider windows of last n values and the exceptions are handled.

Any idea?

Answer 1

I'm not 100% sure I'm getting what you need, but the shift function leaves behind NA values by default. You can change that behaviour by passing a fill argument. In your case, since you're summing the data, you might want to try it with fill=0 :

set.seed( 123 )
x[, .(b, Reduce(`+`, shift(b, 0:2, fill=0))), by = a]

head returns:

   a         b       V2
1: a 0.5999890 0.599989
2: a 0.8903502 1.490339
3: a 0.7205963 2.210935
4: a 0.5492847 2.160231
5: a 0.9540912 2.223972
6: a 0.5854834 2.088859