Fast matrix subset computation

Question

I have a dataset of a about a million rows ordered by id, start (some ids have multiple starting points) and year, and would like to calculate 5-year averages (start-5 to start) of the two variables (var1 and var2) within each id.

For example, the 5-year averages in var1 would be 243.2=(47+99+1000+60+10)/5 and 46=(133+13+88-50)/4 (4-year average due to data range limitation) for id==1 and id==2, respectively. Is there a fast alternative to the code below?

Sample data:

id  start  year  var1  var2
1   2005   2000  500   333
1   2005   2001  10    444
1   2005   2002  60    555
1   2005   2003  1000  99
1   2005   2004  99    15
1   2005   2005  47    0
1   2005   2006  180   NA
2   2003   2000  -50   NA
2   2003   2001  88    17
2   2003   2002  13    77
2   2003   2003  133   55
2   2003   2004  86    30
2   2003   2005  10    100

Code:

  # Find startpoint per id
  idx <- which(year==start)

  # Compute 
  sapply(idx, function(x){
    with( dat, c(id[x],
                  start[x],
                  mean( var1[id==id[x] & (year>=max(2000,year[x]-4) & year<=year[x])], na.rm=T ),
                  mean( var2[id==id[x] & (year>=max(2000,year[x]-4) & year<=year[x])], na.rm=T )) )
  })

Tweaked version based on accepted solution below:

data <- setDT(data)[, .(var1_avg5 = mean(var1[year > start-5 & year <= start], na.rm = T),
                        var2_avg5 = mean(var2[year > start-5 & year <= start], na.rm = T),
                        start, 
                        year), 
               by=id]

Answer 1

Is this what you want?

library(data.table)

# data simulation
n = 7e6
data = data.table(
  id = sample(seq(1,n / 7), n, replace = TRUE),
  year = sample(seq(2000, 2010), n, replace = TRUE),
  var1 = rnorm(n),
  var2 = rexp(n)
)
data[, start := max(year) - sample(c(1,2), 1), id]


# calculation
t1 = Sys.time()
data = data[year > start - 5 & year <= start]
data[, .(var1 = mean(var1, na.rm = T),
         var2 = mean(var2, na.rm = T)), id]
t2 = Sys.time()
print(t2 - t1)

Time difference of 0.511766 secs

Answer 2

Consider rollmean in zoo (well-known time series package) passed via tapply :

library(zoo)
...
df$var1_five_yr_avg <- with(df, unlist(tapply(var1, id, function(x) rollmeanr(x, k=5, fill=NA))))

df$var2_five_yr_avg <- with(df, unlist(tapply(var2, id, function(x) rollmeanr(x, k=5, fill=NA))))

df
#    id start year var1 var2 var1_five_yr_avg var2_five_yr_avg
# 1   1  2005 2000  500  333               NA               NA
# 2   1  2005 2001   10  444               NA               NA
# 3   1  2005 2002   60  555               NA               NA
# 4   1  2005 2003 1000   99               NA               NA
# 5   1  2005 2004   99   15            333.8            289.2
# 6   1  2005 2005   47    0            243.2            222.6
# 7   1  2005 2006  180   NA            277.2               NA
# 8   2  2003 2000  -50   NA               NA               NA
# 9   2  2003 2001   88   17               NA               NA
# 10  2  2003 2002   13   77               NA               NA
# 11  2  2003 2003  133   55               NA               NA
# 12  2  2003 2004   86   30             54.0               NA
# 13  2  2003 2005   10  100             66.0             55.8

However, you indicate a more dynamic need to run multiple rolling means depending on data availability. Hence, consider running multiple rolling means with ifelse logic.

proc_rollmeans <- function(x) {

  five_yr <- rollmeanr(x, k=5, fill=NA)
  four_yr <- rollmeanr(x, k=4, fill=NA)
  three_yr <- rollmeanr(x, k=3, fill=NA)
  two_yr <- rollmeanr(x, k=2, fill=NA)
  one_yr <- x

  ifelse(!is.na(five_yr), five_yr,
         ifelse(!is.na(four_yr), four_yr,
                ifelse(!is.na(three_yr), three_yr,
                       ifelse(!is.na(two_yr), two_yr, one_yr)
                      )
               )
         )

}

df$var1_five_yr_avg <- with(df, unlist(tapply(var1, id, proc_rollmeans)))

df$var2_five_yr_avg <- with(df, unlist(tapply(var2, id, proc_rollmeans)))

df

#    id start year var1 var2 var1_five_yr_avg var2_five_yr_avg
# 1   1  2005 2000  500  333            500.0        333.00000
# 2   1  2005 2001   10  444            255.0        388.50000
# 3   1  2005 2002   60  555            190.0        444.00000
# 4   1  2005 2003 1000   99            392.5        357.75000
# 5   1  2005 2004   99   15            333.8        289.20000
# 6   1  2005 2005   47    0            243.2        222.60000
# 7   1  2005 2006  180   NA            277.2               NA
# 8   2  2003 2000  -50   NA            -50.0               NA
# 9   2  2003 2001   88   17             19.0         17.00000
# 10  2  2003 2002   13   77             17.0         47.00000
# 11  2  2003 2003  133   55             46.0         49.66667
# 12  2  2003 2004   86   30             54.0         44.75000
# 13  2  2003 2005   10  100             66.0         55.80000