6

I am trying to calculate a rolling mean of value grouped by multiple dimensions in R. Something I would do in SQL in the following way:

AVG(value) OVER 
   (PARTITION BY dim1, dim2 ORDER BY date 
       RANGE BETWEEN 5 PRECEDING AND CURRENT ROW)

The following seems to work if I select just a few dimensions:

s <- ave(df$value, 
     list(df$dim1, df$dim2), 
     FUN= function(x) rollapply(x, 5, mean, align='right'))

but gives the following error when I select full set of dimensions:

Error: k <= n is not TRUE

I get the same error when I run: 

rollapply(c(1:2), 3, mean, align='right')

so I guess the issue is that some combinations of dimensions do not have enough values to calculate mean.

How could I overcome it? I am fine with having a NA as a result for those combinations. Any help would be much appreciated..

Adam
  • 756
  • 4
  • 10
  • 23
  • 1
    Well, in your last example, you are trying to use the *last three values* on each iteration...but the iteration vector `c(1:2)` only has two! – Robert Krzyzanowski Apr 03 '14 at 18:41
  • 1
    Yes, I am aware what the issue is, second example is just to illustrate it. The question is, how can I overcome it and get NA if there are to few elements to calculate rollapply instead of getting an error message. – Adam Apr 04 '14 at 09:29

2 Answers2

3

roll_meanr from the RcppRoll package will do this by default:

library(RcppRoll)
> roll_meanr(c(1:2), 3)
# [1] NA NA
Jake Fisher
  • 3,220
  • 3
  • 26
  • 39
0

rollapply(c(1:10), 3, mean, align='right', fill=NA) should do the trick, provided your vector is long enough to produce any data.

note that rollapply(c(1:2), 3, mean, align='right', fill=NA) still returns the error for the reason stated by @robert-krzyzanowski

user2464194
  • 1