Rolling regression with dplyr and lsfit

Question

I'm trying to run a rolling regression with dplyr . I'm using rollapplyr from package zoo and lsfit as I'm only interested in the beta of the regression. Here's what I've tried:

library(dplyr); library(zoo)

df1 = expand.grid(site = seq(10),
                    year = 2000:2004,
                    day = 1:50)

df1 %>%
group_by(year) %>%
mutate(beta1 = rollapplyr(data = site,
                            width = 5,
                            FUN = lsfit,
                            x=day))

I'm getting this error: Error: not all arguments have the same length

I think rollapplyr accepts non-zoo objects but I may be wrong. It could also be that the piping ( %>% ) does not play well with rollapplyr as it requires a data object in the function.

Any idea?

EDIT My question is different from: rolling regression with dplyr I want to use pipes in order to use group_by

Answer 1

The function will not cycle through multiple vectors. The sliced site vector is being compared to the full vector day . We can write our own rolling apply function with Map to go through groups of our vector:

rollapplydf <- function(xx, width) {
  l <- length(xx)
  sq <- Map(':', 1:(l-width+1), width:l)
  lst <- lapply(sq, function(i) lm(xx[i] ~ seq(length(xx[i])))$coeff[2] )
  do.call('rbind', c(rep(NA, width-1L), lst))
}

So we can add this to the pipe:

library(dplyr)
df1 %>% 
  group_by(year) %>% 
  mutate(beta1 = rollapplydf(xx = site, width = 5) )

# Source: local data frame [2,500 x 4]
# Groups: year [5]
# 
#     site  year   day beta1
#    (int) (int) (int) (dbl)
# 1      1  2000     1    NA
# 2      2  2000     1    NA
# 3      3  2000     1    NA
# 4      4  2000     1    NA
# 5      5  2000     1     1
# 6      6  2000     1     1
# 7      7  2000     1     1
# 8      8  2000     1     1
# 9      9  2000     1     1
# 10    10  2000     1     1
# ..   ...   ...   ...   ...