calculating sum of previous 3 rows in R data.table (by grid-square)

Question

I would like to calculate the rainfall that has fallen over the last three days for each grid square, and add this as a new column in my data.table. To be clear, I want to sum up the current and PREVIOUS two (2) days of rainfall, for each meterological grid square

library ( zoo )
library (data.table)


# making the data.table
rain           <- c(NA, NA, NA, 0, 0, 5, 1, 0, 3, 10)  # rainfall values to work with
square         <- c(1,1,1,1,1,1,1,1,1,2)               # the geographic grid square for the rainfall measurement
desired_result <- c(NA, NA, NA, NA, NA, 5, 6, 6, 4, NA )  # this is the result I'm looking for (the last NA as we are now on to the first day of the second grid square)
weather <- data.table(rain, square, desired_result)  # making the data.table

My attempt to answer: this line used to work, but no longer does

weather[, rain_3 := filter(rain, rep(1, 2), sides = 1), by = list(square)]  

So here I am trying another method:

# this next line gets the numbers right, but sums the following values, not the preceeding ones. 
weather$rain_3 <- rollapply(zoo(weather$rain), list(seq(-2,0)), sum)

# here I add in the by weather$ square, but still no success
weather$rain_3 <- rollapply(zoo(weather$rain), list(seq(-2,0)), sum, by= list(weather$square))

I would greatly appreciate any insights or suggestions you may have.

Many thanks!

Answer 1

Here's a quick and efficient solution using the latest data.table version (v 1.9.6+)

weather[, rain_3 := Reduce(`+`, shift(rain, 0:2)), by = square]
weather
#     rain square desired_result rain_3
#  1:   NA      1             NA     NA
#  2:   NA      1             NA     NA
#  3:   NA      1             NA     NA
#  4:    0      1             NA     NA
#  5:    0      1             NA     NA
#  6:    5      1              5      5
#  7:    1      1              6      6
#  8:    0      1              6      6
#  9:    3      1              4      4
# 10:   10      2             NA     NA

The basic idea here is to shift the rain column twice and then sum up the rows.

Answer 2

The rollapply solution would be done like this:

weather[, rain_3 := rollapplyr(rain, 3, sum, fill = NA_real_), by = square]

giving:

    rain square desired_result rain_3
 1:   NA      1             NA     NA
 2:   NA      1             NA     NA
 3:   NA      1             NA     NA
 4:    0      1             NA     NA
 5:    0      1             NA     NA
 6:    5      1              5      5
 7:    1      1              6      6
 8:    0      1              6      6
 9:    3      1              4      4
10:   10      2             NA     NA

Update

Have simplified based on version of zoo that came out since this question was originally asked.

Answer 3

weather[, rain_3 := filter(rain, rep(1, 3), sides = 1), by = list(square)]  
#Error in filter(rain, rep(1, 3), sides = 1) : 
#  'filter' is longer than time series
weather[, rain_3 := if(.N > 2) filter(rain, rep(1, 3), sides = 1) else NA_real_, 
        by = square] 
#    rain square desired_result rain_3
# 1:   NA      1             NA     NA
# 2:   NA      1             NA     NA
# 3:   NA      1             NA     NA
# 4:    0      1             NA     NA
# 5:    0      1             NA     NA
# 6:    5      1              5      5
# 7:    1      1              6      6
# 8:    0      1              6      6
# 9:    3      1              4      4
#10:   10      2             NA     NA

Take care that dplyr is not loaded because it masks filter. If you need dplyr, you can call stats::filter explicitly.