R Compute Statistics on Lagged Partitions - r

Question

I have a data.frame with one column containing categorical data, one column containing dates, and one column containing numeric values. For simplicity, see the sample below:
A B C
1 L 2015-12-01 5.7
2 M 2015-11-30 2.1
3 K 2015-11-01 3.2
4 L 2015-10-05 5.7
5 M 2015-12-05 1.2
6 L 2015-11-15 2.3
7 L 2015-12-03 4.4
I would like to, for each category in A, compute a lagging average (e.g. average of the previous 30 days' values in column C).
I cannot for the life of me figure this one out. I have tried using sapply and a custom function that subsets the data.frame on category and date (or a deep copy of it) and returns the statistic (think mean or sd) and that works fine for single values, but it returns all NA's from inside sapply.
Any help you can give is appreciated.

Answer 1

This could be done more compactly, but here I have drawn it out to make it easiest to understand. The core is the split, lapply/apply, and then putting it back together. It uses a date window rather than a solution based on sorting, so it is very general. I also put the object back to its original order to enable direct comparison.
# set up the data
set.seed(100)
# create a data.frame with about a two-month period for each category of A
df <- data.frame(A = rep(c("K", "L", "M"), each = 60),
B = rep(seq(as.Date("2015-01-01"), as.Date("2015-03-01"), by="days"), 3),
C = round(runif(180)*6, 1))
head(df)
## A B C
## 1 K 2015-01-01 1.8
## 2 K 2015-01-02 1.5
## 3 K 2015-01-03 3.3
## 4 K 2015-01-04 0.3
## 5 K 2015-01-05 2.8
## 6 K 2015-01-06 2.9
tail(df)
## A B C
## 175 M 2015-02-24 4.8
## 176 M 2015-02-25 2.0
## 177 M 2015-02-26 5.7
## 178 M 2015-02-27 3.9
## 179 M 2015-02-28 2.8
## 180 M 2015-03-01 3.6
# preserve original order
df$originalOrder <- 1:nrow(df)
# randomly shuffle the order
randomizedOrder <- order(runif(nrow(df)))
df <- df[order(runif(nrow(df))), ]
# split on A - your own data might need coercion of A to a factor
df.split <- split(df, df$A)
# set the window size
window <- 30
# compute the moving average
listD <- lapply(df.split, function(tmp) {
apply(tmp, 1, function(x) mean(tmp$C[tmp$B <= as.Date(x["B"]) & tmp$B (as.Date(x["B"]) - window)]))
})
# combine the result with the original data
result <- cbind(do.call(rbind, df.split), rollingMean = unlist(listD))
# and tidy up:
# return to original order
result <- result[order(result$originalOrder), ]
result$originalOrder <- NULL
# remove the row names
row.names(result) <- NULL
result[c(1:5, 59:65), ]
## A B C rollingMean
## 1 K 2015-01-01 1.8 1.800000
## 2 K 2015-01-02 1.5 1.650000
## 3 K 2015-01-03 3.3 2.200000
## 4 K 2015-01-04 0.3 1.725000
## 5 K 2015-01-05 2.8 1.940000
## 59 K 2015-02-28 3.6 3.080000
## 60 K 2015-03-01 1.3 3.066667
## 61 L 2015-01-01 2.8 2.800000
## 62 L 2015-01-02 3.9 3.350000
## 63 L 2015-01-03 5.8 4.166667
## 64 L 2015-01-04 4.1 4.150000
## 65 L 2015-01-05 2.7 3.860000