I wanted to complement the previous answers by mentioning two ways in which I approach this problem in the important case when you are not guaranteed that each group has data for every time period. That is, you still have a regularly spaced time series, but there might be missings here and there. I will focus on two ways to improve the dplyr solution.
We start with the same data that you used...
library(dplyr)
library(tidyr)
set.seed(1)
data_df = data.frame(time = c(1:3, 1:4),
groups = c(rep(c("b", "a"), c(3, 4))),
value = rnorm(7))
data_df
#> time groups value
#> 1 1 b -0.6264538
#> 2 2 b 0.1836433
#> 3 3 b -0.8356286
#> 4 1 a 1.5952808
#> 5 2 a 0.3295078
#> 6 3 a -0.8204684
#> 7 4 a 0.4874291
... but now we delete a couple of rows
data_df = data_df[-c(2, 6), ]
data_df
#> time groups value
#> 1 1 b -0.6264538
#> 3 3 b -0.8356286
#> 4 1 a 1.5952808
#> 5 2 a 0.3295078
#> 7 4 a 0.4874291
Simple dplyr solution no longer works
data_df %>%
arrange(groups, time) %>%
group_by(groups) %>%
mutate(lag.value = lag(value)) %>%
ungroup()
#> # A tibble: 5 x 4
#> time groups value lag.value
#> <int> <fct> <dbl> <dbl>
#> 1 1 a 1.60 NA
#> 2 2 a 0.330 1.60
#> 3 4 a 0.487 0.330
#> 4 1 b -0.626 NA
#> 5 3 b -0.836 -0.626
You see that, although we don't have the value for the case (group = 'a', time = '3'), the above still shows a value for the lag in the case of (group = 'a', time = '4'), which is actually the value at time = 2.
Correct dplyr solution
The idea is that we add the missing (group, time) combinations. This is VERY memory-inefficient when you have lots of possible (groups, time) combinations, but the values are sparsely captured.
dplyr_correct_df = expand.grid(
groups = sort(unique(data_df$groups)),
time = seq(from = min(data_df$time), to = max(data_df$time))
) %>%
left_join(data_df, by = c("groups", "time")) %>%
arrange(groups, time) %>%
group_by(groups) %>%
mutate(lag.value = lag(value)) %>%
ungroup()
dplyr_correct_df
#> # A tibble: 8 x 4
#> groups time value lag.value
#> <fct> <int> <dbl> <dbl>
#> 1 a 1 1.60 NA
#> 2 a 2 0.330 1.60
#> 3 a 3 NA 0.330
#> 4 a 4 0.487 NA
#> 5 b 1 -0.626 NA
#> 6 b 2 NA -0.626
#> 7 b 3 -0.836 NA
#> 8 b 4 NA -0.836
Notice that we now have a NA at (group = 'a', time = '4'), which should be the expected behaviour. Same with (group = 'b', time = '3').
Tedious but also correct solution using the class zoo::zooreg
This solution should work better in terms of memory when the amount of cases is very large, because instead of filling the missing cases with NA's, it uses indices.
library(zoo)
zooreg_correct_df = data_df %>%
as_tibble() %>%
# nest the data for each group
# should work for multiple groups variables
nest(-groups, .key = "zoo_ob") %>%
mutate(zoo_ob = lapply(zoo_ob, function(d) {
# create zooreg objects from the individual data.frames created by nest
z = zoo::zooreg(
data = select(d,-time),
order.by = d$time,
frequency = 1
) %>%
# calculate lags
# we also ask for the 0'th order lag so that we keep the original value
zoo:::lag.zooreg(k = (-1):0) # note the sign convention is different
# recover df's from zooreg objects
cbind(
time = as.integer(zoo::index(z)),
zoo:::as.data.frame.zoo(z)
)
})) %>%
unnest() %>%
# format values
select(groups, time, value = value.lag0, lag.value = `value.lag-1`) %>%
arrange(groups, time) %>%
# eliminate additional periods created by lag
filter(time <= max(data_df$time))
zooreg_correct_df
#> # A tibble: 8 x 4
#> groups time value lag.value
#> <fct> <int> <dbl> <dbl>
#> 1 a 1 1.60 NA
#> 2 a 2 0.330 1.60
#> 3 a 3 NA 0.330
#> 4 a 4 0.487 NA
#> 5 b 1 -0.626 NA
#> 6 b 2 NA -0.626
#> 7 b 3 -0.836 NA
#> 8 b 4 NA -0.836
Finally, lets check that both correct solutions are actually equal:
all.equal(dplyr_correct_df, zooreg_correct_df)
#> [1] TRUE
