Question

I want to “loop through” the rows of a data.table and calculate an average for each row. The average should be calculated based on the following mechanism:

1. Look up the identifier ID in row i (ID(i))
2. Look up the value of T2 in row i (T2(i))
3. Calculate the average over the Data1 values in all rows j, which meet these two criteria: ID(j) = ID(i) and T1(j) = T2(i)

4. Enter the calculated average in the column Data2 of row i

    DF = data.frame(ID=rep(c("a","b"),each=6), 
                T1=rep(1:2,each=3), T2=c(1,2,3), Data1=c(1:12))
    DT = data.table(DF)
    DT[ , Data2:=NA_real_]
        ID T1 T2  Data1 Data2
   [1,]  a  1  1     1    NA
   [2,]  a  1  2     2    NA
   [3,]  a  1  3     3    NA
   [4,]  a  2  1     4    NA
   [5,]  a  2  2     5    NA
   [6,]  a  2  3     6    NA
   [7,]  b  1  1     7    NA
   [8,]  b  1  2     8    NA
   [9,]  b  1  3     9    NA
   [10,] b  2  1    10    NA
   [11,] b  2  2    11    NA
   [12,] b  2  3    12    NA
   

For this simple example the result should look like this:

      ID T1 T2  Data1 Data2
[1,]  a  1  1     1    2
[2,]  a  1  2     2    5
[3,]  a  1  3     3    NA
[4,]  a  2  1     4    2
[5,]  a  2  2     5    5
[6,]  a  2  3     6    NA
[7,]  b  1  1     7    8
[8,]  b  1  2     8    11
[9,]  b  1  3     9    NA
[10,] b  2  1    10    8
[11,] b  2  2    11    11
[12,] b  2  3    12    NA

I think one way of doing this would be to loop through the rows, but I think that is inefficient. I’ve had a look at the apply() function, but I’m sure if it would solve my problem. I could also use data.frame instead of data.table if this would make it much more efficient or much easier. The real dataset contains approximately 1 million rows.

Answer 1

The rule of thumb is to aggregate first, and then join to that.

agg = DT[,mean(Data1),by=list(ID,T1)]
setkey(agg,ID,T1)
DT[,Data2:={JT=J(ID,T2);agg[JT,V1][[3]]}]
      ID T1 T2 Data1 Data2
 [1,]  a  1  1     1     2
 [2,]  a  1  2     2     5
 [3,]  a  1  3     3    NA
 [4,]  a  2  1     4     2
 [5,]  a  2  2     5     5
 [6,]  a  2  3     6    NA
 [7,]  b  1  1     7     8
 [8,]  b  1  2     8    11
 [9,]  b  1  3     9    NA
[10,]  b  2  1    10     8
[11,]  b  2  2    11    11
[12,]  b  2  3    12    NA

As you can see it’s a bit ugly in this case (but will be fast). It’s planned to add drop which will avoid the [[3]] bit, and maybe we could provide a way to tell [.data.table to evaluate i in calling scope (i.e. no self join) which would avoid the JT= bit which is needed here because ID is in both agg and DT.

keyby has been added to v1.8.0 on R-Forge so that avoids the need for the setkey, too.