Simulating new subjects is a very common task, and this is even the
default simulation method in NMsim (because Nonmem does
that with in a $SIMULATION step). However, sometimes one
may want to perform multiple simulations using the same
simulated subjects. So we want to simulate new subjects, but
only once, and then we want to be able to reuse those subjects. This can
be wanted for at least these reasons:
Compare simulations of different scenarios with multiple subjects where ensuring that the difference in results isn’t caused by differences in simulated subjects (less important the more subjects that are used in the simulations)
Ensure reproducibility of results. If the simulation input data is modified by say adding a simulation scenario in between two existing ones - or even just a sample time, the random number generator will be in different state when sampling the subjects the next time it runs. Suddenly, an estimated prediction interval comes out a little different.
NMsim has a way to do this. It includes a function,
simPopEtas(), to generate a simulated .phi
file which is a Nonmem-native format for storing emperical Bayes
estimates (essentially individual parameters). It does so by reading and
simulating from the estimated OMEGA matrix (random-effect
variance-covariance matrix - a covariance step is not needed to get
this). The simulation method called NMsim_known() can then
be used to simulate subjects as stored in this file. See NMsim-known.html
for more information on NMsim_known.
What is a subject?
What is refered to here by “subject” is really a combination of
ETAs. Covariates must be handled by the user in the
simulation input dataset. This is also discussed in NMsim-known.html.
Example
file.project <- function(...)file.path(system.file("examples",package="NMsim"),...)
file.mod <- file.project("nonmem/xgxr032.mod")Let’s simulate 10,000 ETA combinations and store them in a file
called xgxr032_simEtas.phi.
NMsim:::simPopEtas(file=file.mod,
N=1e4,
seed=238861,
file.phi="xgxr032_simEtas.phi"
)We’ll just use the simulation data set created in NMsim-DataCreate.html
dat.sim <- read_fst(path="simulate-results/dat_sim.fst",as.data.table=TRUE)And now we can use NMsim_known().
dat.sim.multiple <- egdt(dat.sim[,!("ID")],data.table(ID=c(1,4,89)))
#> data nrows ncols
#> <char> <int> <int>
#> 1: dt1 171 10
#> 2: dt2 3 1
#> 3: result 513 11
setorder(dat.sim.multiple,ID,TIME,EVID)
simres <- NMsim(
file.mod=file.mod,
data=dat.sim.multiple,
method.sim=NMsim_known,
file.phi="xgxr032_simEtas.phi",
name.sim="reuseSubjs",
table.vars="PRED IPRED",
path.nonmem="/opt/NONMEM/nm75/run/nmfe75",
dir.sims="simulate-tmp",
dir.res="simulate-results"
)