How can I use acf() for mutliple subsets in R? - r

Question

I am trying write code that will do autocorrelation for multiple subsets. For example. I have health data for multiple countries over time. I want to get each country's autocorrelation for each variable. Any help would be great!
Here are some things I have tried, unsuccessfully:
require(plyr)
POP_ACF=acf(PhD_data_list_view$POP, lag.max=NULL, type=c("correlation"),
plot=TRUE, na.action=na.pass, demean=TRUE)
dlply(PhD_data_list_view, .(Country), function(x) POP_ACF %+% x)
POP_ACF=function(PhD_data_list_view$POP) c(acf(PhD_data_list_view$POP, plot=TRUE)$acf)

Answer 1

acf is a function takes a vector and returns a list. That makes it a natural fit for the purrr package, which maps functions over lists, but it can also be done using base R.
I'll use the beaver1 dataset from the datasets package since you didn't provide yours. I'll use different days of observations as the analogue to your different countries, and temperature for your POP variable.
Base R:
split turns the vector beaver1$temp into a list of vectors along the second argument, beaver1$day.
Then mapply runs the function acf on each element of that list.
Since we're using mapply instead of lapply, we can also provide another list of arguments, here the titles for each plot, main = unique(beaver1$day).
The last argument, SIMPLIFY = F, tells it to return the default output, not attempt to coerce the list into anything else.
par(mfrow = c(1,2))
mapply(acf,
split(beaver1$temp, beaver1$day),
main = unique(beaver1$day),
SIMPLIFY = F)
# $`346`
#
# Autocorrelations of series ‘dots[[1L]][[1L]]’, by lag
#
# 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
# 1.000 0.838 0.698 0.593 0.468 0.355 0.265 0.167 0.113 0.069 0.028 0.037 0.087 0.108 0.145 0.177 0.151 0.125 0.123 0.106
# $`347`
#
# Autocorrelations of series ‘dots[[1L]][[2L]]’, by lag
#
# 0 1 2 3 4 5 6 7 8 9 10 11 12 13
# 1.000 0.546 0.335 0.130 0.080 0.024 -0.025 -0.103 -0.090 -0.032 0.168 0.036 -0.089 -0.306
purrr and the tidy way:
This way is a bit more flexible depending what you want to do with the output. We can use purrr::map as a direct drop-in for mapply:
library(purrr)
beaver1 %>%
split(.$day) %>%
map(~acf(.$temp, main = unique(.$day)))
Which returns the exact same output. But we can also go fully tidy and return the data from acf as a dataframe so that we can explore it further with ggplot2.
The first map is returning a list of outputs, each of which is a list containing, among other things, variables lag, acf, and n.used.
The map_dfr is running the function data.frame, assigning each of those variables to a new column.
We also make a column to calculate the CIs. Refer to: How is the confidence interval calculated for the ACF function?
Then we can use ggplot to make any kind of plot we want, and we still have the data for any other analysis you want to do.
library(ggplot2)
beaver_acf <-
beaver1 %>%
split(.$day) %>%
map(~acf(.$temp, plot = F)) %>%
map_dfr(
~data.frame(lag = .$lag,
acf = .$acf,
ci = qnorm(0.975)/sqrt(.$n.used)
), .id = "day")
head(beaver_acf)
# day lag acf ci
# 1 346 0 1.0000000 0.2054601
# 2 346 1 0.8378889 0.2054601
# 3 346 2 0.6983476 0.2054601
# 4 346 3 0.5928198 0.2054601
# 5 346 4 0.4680912 0.2054601
# 6 346 5 0.3554939 0.2054601
ggplot(beaver_acf, aes(lag, acf)) +
geom_segment(aes(xend = lag, yend = 0)) +
geom_hline(aes(yintercept = ci), linetype = "dashed", color = "blue") +
geom_hline(aes(yintercept = -ci), linetype = "dashed", color = "blue") +
facet_wrap(~variable)