take mean of previous 7 values with rollapply

library(data.table)
library(pipeR)
library(zoo)
library(plyr)

# generate data
numDays <- 20
numID <- 3000
numRecords <- numDays * numID * 0.5
set.seed(5)
DT <- data.table(matrix(rnorm(numDays * numID), numID)) %>>% `[`(j = ID := 1:numID) %>>%
  melt.data.table(id.vars = numDays + 1, variable.name = "Day", value.name = "X") %>>%
  `[`(j = Day := as.integer(mapvalues(Day, paste0("V", 1:numDays), 1:numDays))) %>>%
  `[`(i = sample(nrow(.), numRecords)) %>>% setorder(ID, Day)

DT[ , Day2 := factor(Day, levels = seq(min(DT$Day), max(DT$Day)))]
setkey(DT, ID, Day)

# check the number of observation is greater than 1
# DT[ , .N, by = .(ID)] %>>% `[[`("N") %>>% `!=`(1) %>>% all # TRUE

mean2 <- function(x) {
  if (length(x) == 1 || all(is.na(x)))
    return(NA)
  mean(head(x, length(x) - 1), na.rm = TRUE)
}

f1 <- function(DT) {
  DT2 <- dcast.data.table(DT, ID ~ Day2, sum, fill = NA, drop = FALSE, value.var = "X")
  meanDT <- DT2[ , 2:ncol(DT2)] %>>% as.matrix %>>% t %>>%
    rollapply(8, mean2, partial = TRUE, align = "right") %>>% data.table %>>%
    melt.data.table(measure.var = 1:ncol(.), variable.name = "ID", value.name = "x_mean",
                    variable.factor = FALSE) %>>%
    `[`(j = Day := 1:(ncol(DT2)-1), by = .(ID)) %>>%
    `[`(j = ID := as.integer(mapvalues(ID, paste0("V", 1:nrow(DT2)), DT2$ID)))
  return(merge(DT, meanDT, by = c("ID", "Day")))
}

f2 <- function(DT) {
  DT[CJ(unique(ID), seq(min(Day), max(Day)))] %>>%
    `[`(j = .(Day = rollapply(Day, 8, max, partial = TRUE, align = "right"),
              x_mean = rollapply(X, 8, mean2, partial = TRUE, align = "right")),
        by = .(ID)) %>>%
    merge(DT, by = c("ID", "Day"), all.y = TRUE)
}

mvAvg_f <- function(val, day, span = 7) {
  sapply(day, function(y){
    tmp <- day - y
    mean(val[which(tmp < 0 & tmp >= -span)])
  })
}
f3 <- function(DT) {
  DT[ , x_mean_2 := mvAvg_f(X, Day), by = .(ID)]
}

f4 <- function(DT) {
  DT[ , x_mean_3 := sapply(Day, function(s) mean(X[between(Day, s-7, s-1)])), by = .(ID)]
}

f5 <- function(DT) {
  DT[ , x_mean_4 := sapply(Day, function(s) mean(X[Day >= s-7 & Day < s])), by = .(ID)]
}

resDT1 <- f1(DT)
resDT2 <- f2(DT)
all.equal(resDT1, resDT2, ignore.col.order = TRUE) # TRUE
f3(DT)
f4(DT)
f5(DT)
all.equal(resDT1$x_mean, DT$x_mean_2, ignore.col.order = TRUE) # TRUE
all.equal(resDT1$x_mean, DT$x_mean_3, ignore.col.order = TRUE) # TRUE
all.equal(resDT1$x_mean, DT$x_mean_4, ignore.col.order = TRUE) # TRUE

library(microbenchmark)
microbenchmark(f1(DT), f2(DT), f3(DT), f4(DT), f5(DT), times = 10L)
# Unit: milliseconds
#    expr       min        lq      mean    median        uq       max neval
#  f1(DT) 4879.9619 4909.4480 4961.4962 4938.1009 4997.4011 5152.2432    10
#  f2(DT) 4051.7383 4095.4616 4133.2843 4133.9591 4162.8146 4245.0343    10
#  f3(DT)  321.6416  322.3145  324.9348  323.4118  325.8534  331.5509    10
#  f4(DT) 1942.0352 1956.5268 1992.7289 1979.8526 2014.8336 2096.0038    10
#  f5(DT)  149.6435  151.4261  155.0831  152.9530  161.5350  162.9220    10