R: rolling sums for groups with uneven time gaps

You can use rollapply from zoo once you fill out the missing dates first:

library(dplyr)
library(zoo)

set.seed(3737)
DF2 = data.frame(user_id = c(rep(27, 7), rep(11, 7)),
             date = as.Date(rep(c('2016-01-01', '2016-01-03', '2016-01-05', '2016-01-07', '2016-01-10', '2016-01-14', '2016-01-16'), 2)),
             value = round(rnorm(14, 15, 5), 1))

all_combinations <- expand.grid(user_id=unique(DF2$user_id), 
                            date=seq(min(DF2$date), max(DF2$date), by="day"))

res <- DF2 %>% 
    merge(all_combinations, by=c('user_id','date'), all=TRUE) %>%
    group_by(user_id) %>% 
    arrange(date) %>% 
    mutate(v_minus7=rollapply(value, width=8, FUN=function(x) sum(x, na.rm=TRUE), partial=TRUE, align='right'),
           v_minus14=rollapply(value, width=15, FUN=function(x) sum(x, na.rm=TRUE), partial=TRUE, align='right')) %>%
    filter(!is.na(value))

Here is another idea with findInterval to minimize comparisons and operations. First define a function to accomodate the basic part ignoring the grouping. The following function computes the cumulative sum, and subtracts the cumulative sum at each position from the one at its respective past date:

ff = function(date, value, minus)
{
    cs = cumsum(value)  
    i = findInterval(date - minus, date, left.open = TRUE) 
    w = which(as.logical(i))
    i[w] = cs[i[w]]
    cs - i
}

And apply it by group:

do.call(rbind, 
        lapply(split(DF2, DF2$user_id), 
               function(x) data.frame(x, 
                         minus7 = ff(x$date, x$value, 7), 
                         minus14 = ff(x$date, x$value, 14))))
#      user_id       date value minus7 minus14
#11.8       11 2016-01-01   6.8    6.8     6.8
#11.9       11 2016-01-03  21.3   28.1    28.1
#11.10      11 2016-01-05  19.8   47.9    47.9
#11.11      11 2016-01-07  22.0   69.9    69.9
#11.12      11 2016-01-10  19.4   82.5    89.3
#11.13      11 2016-01-14  17.5   58.9   106.8
#11.14      11 2016-01-16  19.3   56.2   119.3
#27.1       27 2016-01-01  15.0   15.0    15.0
#27.2       27 2016-01-03  22.4   37.4    37.4
#27.3       27 2016-01-05  13.3   50.7    50.7
#27.4       27 2016-01-07  21.9   72.6    72.6
#27.5       27 2016-01-10  20.6   78.2    93.2
#27.6       27 2016-01-14  18.6   61.1   111.8
#27.7       27 2016-01-16  16.4   55.6   113.2

The above apply-by-group operation can, of course, be replaced by any method prefereable.

Here is a zoo solution.

Define a function sum_last that given a zoo object takes the sum of the values whose times are within k days of the last day in the series and define a roll function which applies it to an entire series. Then use ave to apply roll to each user_id once for k=7 and once for k=14.

Note that this makes use of the coredata argument to rollapply that was introduced in the most recent version of zoo so be sure you don't have an earlier version.

library(zoo)

# compute sum of values within k time units of last time point
sum_last <- function(z, k) {
  tt <- time(z)
  sum(z[tt > tail(tt, 1) - k])
}

# given indexes ix run rollapplyr on read.zoo(DF2[ix, -1])
roll <- function(ix, k) {
 rollapplyr(read.zoo(DF2[ix, -1]), k, sum_last, coredata = FALSE, partial = TRUE, k = k)
}

nr <- nrow(DF2)
transform(DF2, 
  v_minus7 = ave(1:nr, user_id, FUN = function(x) roll(x, 7)),
  v_minus14 = ave(1:nr, user_id, FUN = function(x) roll(x, 14)))