groupby summarise outside of groupby dplyr - r

Question

I'm trying to group ids with date in this dataset, but I want to summarise based on one of the features outside of the group.
library(dplyr)
library(lubridate)
set.seed(100)
df <- data.frame(ids = sample(c('436247', '2465347', '346654645'), 10000, replace=TRUE),
date = sample(seq.Date(ymd('2018-03-01'), ymd('2018-05-01'), by=1), 10000, replace=TRUE))
new_df <- df %>%
group_by(ids, date) %>%
summarise(events = length(ids[date >= date - 30 & date <= date]))
I'm trying to take this dataframe and answer the question - "for each of the ids, and each date, how many other records within that id, are within the past 30 days of that date". Unfortunately, when I group_by both the ids and date, it only looks within the grouped date. I've created the solution below, but not sure if there is a better one with dplyr?
groupby_function <- function(df, spec_date){
result <- df %>%
group_by(ids) %>%
summarise(events = length(ids[date >= spec_date - 30 & date <= spec_date])) %>%
mutate(date = spec_date)
return(result)
}
date_vector <- seq.Date(ymd('2018-03-01'), ymd('2018-05-01'), by=1)
list_results <- lapply(date_vector, groupby_function, df=df)
x <- do.call(rbind, list_results)

Answer 1

"for each of the ids, and each date, how many other records within that id, are within the past 30 days of that date"
For that, a "join by" condition makes sense, but isn't yet included in dplyr. Until it is, you could use data.table inside your dplyr chain:
# enumerate id-date combos of interest
grid_df = expand.grid(
id = unique(df$ids),
d = seq(min(df$date), max(df$date), by="day")
)
# helper function
library(data.table)
count_matches = function(DF, targetDF, ...){
onexpr = substitute(list(...))
data.table(targetDF)[DF, on=eval(onexpr), .N, by=.EACHI]$N
}
# use a non-equi join to count matching rows
res = grid_df %>%
mutate(d_dn = d - 30) %>%
mutate(n = count_matches(., df, ids = id, date >= d_dn, date <= d)) %>%
as.tibble
# A tibble: 186 x 4
id d d_dn n
<fctr> <date> <date> <int>
1 436247 2018-03-01 2018-01-30 72
2 2465347 2018-03-01 2018-01-30 69
3 346654645 2018-03-01 2018-01-30 51
4 436247 2018-03-02 2018-01-31 123
5 2465347 2018-03-02 2018-01-31 120
6 346654645 2018-03-02 2018-01-31 100
7 436247 2018-03-03 2018-02-01 170
8 2465347 2018-03-03 2018-02-01 166
9 346654645 2018-03-03 2018-02-01 154
10 436247 2018-03-04 2018-02-02 228
# ... with 176 more rows
It should work fine for equality conditions to write either ids = id or ids == id, I think.
If you're interested, the syntax is x[i, on=, j, by=.EACHI] where x and i are tables. For each row of i, we look up rows of x based on the on= criteria (left-hand side refers to columns in x; right-hand to columns in i); then we do j for each ("by each row of i" so by=.EACHI). In this case, j = .N means that we count matched rows of x, returned as a column of counts N.

Answer 2

You can look at the "ungrouped" data by just going back to the original data frame(calling df$date or df$ids). So I think what you are after is
test_df <- df %>%
group_by(ids, date) %>%
summarise(events = length(df$ids[df$date >= date[1] - 30 & df$date <= date[1] & df$ids == ids[1]]))
Also, I ran your proposed function, but I did not see any difference in the result from your original group_by solution, so I don't think that is what you want.

Answer 3

If a 'non dplyr' solution is acceptable, this gives you what you want.
df$diff <- as.vector(
sapply(unique(df$ids), function(x)
sapply(df$date[df$ids == x], function(y)
sum(abs(y - df$date[df$ids == x]) >= 30)
)
)
)
Alternatively, in dplyr, you can get a result like the above using:
f <- function(x) {
sapply(x, function(y) sum(abs(y - x) >= 30))
}
df$diff <- unlist(
df %>%
group_by(ids) %>%
do(diff = f(.$date)) %>%
.$diff
)

Answer 4

Here's an answer. But it assumes there's a continuous sequence of dates in each id.
df %>%
group_by(ids, date) %>%
count() %>%
arrange(ids, date) %>%
group_by(ids) %>%
mutate(
events = cumsum(n) - cumsum(lag(n, 30, 0))
)