I have a simulation package that allows for the simulation of regression models including nested data structures. You can see the package on github here: simReg. Over the weekend I updated the package to allow for the simulation of unbalanced designs. I’m hoping to put together a new vigenette soon highlighting the functionality.
I am working on a simulation that uses the unbalanced functionality and while simulating longitudinal data I’ve found the function is much slower than the cross sectional counterparts (and balanced designs). I’ve ran some additional testing and I believe I have the speed issues narrowed down to the fact that I am generating a time variable. Essentially, I have a vector of number of observations per cluster. The function then turns this vector of lengths into a time variable starting at 0 up to the maximum number of observations minus 1 by 1. As an example:
x <- round(runif(5, min = 3, max = 10), 0)
unlist(lapply(1:length(x), function(xx) (1:x[xx]) - 1))## [1] 0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 6 7 8 9 0 1 2 3 0 1 2 3 4 5 6 7From the code above, you can see that there the number of observations is generated using runif which is saved to the object x. Then I use a combination of lapply, unlist, and the ‘:’ operator to generate the sequence. This is the same code used in my package above to generate the time variable.
As such, I was interested in testing various ways to generate the sequence and do a performance comparison. I compared the following ways, the ':' operator, seq.int, seq, do.call with mapply, and rep.int for the balanced case as a comparison to how it was done before. This was all done with the great microbenchmark package.
Here are the results from the 7 comparisons:
library(microbenchmark)
x <- round(runif(100, min = 3, max = 15), 0)
microbenchmark(
colon = unlist(lapply(1:length(x), function(xx) (1:x[xx]) - 1)),
seq.int = unlist(lapply(1:length(x), function(xx) seq.int(0, x[xx] - 1, 1))),
seq = unlist(lapply(1:length(x), function(xx) seq(0, x[xx] - 1, 1))),
seq.int_mapply = do.call(c, mapply(seq.int, 0, x - 1)),
seq_mapply = do.call(c, mapply(seq, 0, x - 1)),
colon_mapply = do.call(c, mapply(':', 0, x - 1)),
rep.int = rep.int(1:8 - 1, times = 100), # balanced case for reference.
times = 1000L
)## Unit: microseconds
## expr min lq mean median uq max neval
## colon 56.280 60.5265 73.208671 62.3660 65.2010 1501.166 1000
## seq.int 64.806 70.0705 82.608393 74.5145 78.5070 1605.745 1000
## seq 891.870 922.5175 1029.244121 944.3845 985.0470 6452.576 1000
## seq.int_mapply 78.060 84.3060 105.103943 88.4605 93.4495 5220.393 1000
## seq_mapply 352.497 378.5025 432.245410 394.0690 411.3875 1970.160 1000
## colon_mapply 69.371 74.0085 86.326447 76.7165 81.2535 1517.687 1000
## rep.int 1.728 2.2615 3.039856 2.5075 3.2385 30.710 1000The results (in microseconds) show that this is where the significant slowdown is coming in my package implementing the unbalanced cases, although it appears that the ‘:’ operator is the second best alternative. For those that have not seen the significant speed bump of the seq.int and rep.int over the seq and rep alternatives should also pay close attention (compare lines 2 and 3 above).
I’d be interested in alternative procedures that I am not aware of as well. Although not a big deal when running the package once, doing it 50,000 times does add up.
Lastly, for those that are interested, we can show they are all equivalent methods (except for the rep.int case).
identical(
unlist(lapply(1:length(x), function(xx) (1:x[xx]) - 1)),
unlist(lapply(1:length(x), function(xx) seq.int(0, x[xx] - 1, 1))),
unlist(lapply(1:length(x), function(xx) seq(0, x[xx] - 1, 1))),
do.call(c, mapply(seq.int, 0, x - 1)),
do.call(c, mapply(seq, 0, x - 1)),
do.call(c, mapply(':', 0, x - 1))
)## [1] TRUE