Skip to contents

Simulate from an estimated Nonmem model

Source: R/NMsim.R
NMsim.Rd

Supply a data set and an estimation input control stream, and NMsim can create neccesary files (control stream, data files), run the simulation and read the results. It has additional methods for other simulation types available, can do multiple simulations at once and more. Please see vignettes for an introduction to how to get the most out of this.

Usage

NMsim(
  file.mod,
  data,
  dir.sims,
  name.sim,
  order.columns = TRUE,
  script = NULL,
  subproblems = NULL,
  reuse.results = FALSE,
  seed.R = NULL,
  seed.nm = NULL,
  args.psn.execute,
  table.vars,
  table.options,
  text.sim = "",
  method.sim = NMsim_default,
  execute = TRUE,
  sge = FALSE,
  nc = 1,
  transform = NULL,
  method.execute,
  method.update.inits,
  create.dirs = TRUE,
  dir.psn,
  list.sections,
  sim.dir.from.scratch = TRUE,
  col.row,
  args.NMscanData,
  path.nonmem = NULL,
  nmquiet = FALSE,
  as.fun,
  suffix.sim,
  text.table,
  system.type = NULL,
  dir.res,
  file.res,
  wait,
  quiet = FALSE,
  check.mod = TRUE,
  seed,
  ...
)

Arguments

file.mod

Path(s) to the input control stream(s) to run the simulation on. The outpult control stream is for now assumed to be stored next to the input control stream and ending in .lst instead of .mod. The .ext file must also be present. If simulating known subjects, the .phi is necessary too.

data

The simulation data as a data.frame.

dir.sims

The directory in which NMsim will store all generated files. Default is to create a folder called `NMsim` next to `file.mod`.

name.sim

Give all filenames related to the simulation a suffix. A short string describing the sim is recommended like "ph3_regimens".

order.columns

reorder columns by calling NMdata::NMorderColumns before saving dataset and running simulations? Default is TRUE.

script

The path to the script where this is run. For stamping of dataset so results can be traced back to code.

subproblems

Number of subproblems to use as SUBPROBLEMS in $SIMULATION block in Nonmem. The default is subproblem=0 which means not to use SUBPROBLEMS.

reuse.results

If simulation results found on file, should they be used? If TRUE and reading the results fail, the simulations will still be rerun.

seed.R

A value passed to set.seed(). It may be better use seed.R rather than calling set.seed() manually because the seed can then be captured and stored by NMsim() for reproducibility. See seed.nm for finer control of the seeds that are used in the Nonmem control streams.

seed.nm

Control Nonmem seeds. If a numeric, a vector or a `data.frame`, these are used as the the seed values (a single value or vector will be recycled so make sure the dimesnsions are right, the number of columns in a data.frame will dictate the number of seeds in each Nonmem control stream. Use a list with elements `values`, and `dist` and others for detailed control of the random sources. See ?NMseed for details on what arguments can be passed this way.

Default is to draw seeds betwen 0 and 2147483647 (the values supported by Nonmem) for each simulation. You can pass a function that will be evaluated (say to choose a different pool of seeds to draw from).

In case method.sim=NMsim_known, seed is not used and will be set to 1.

args.psn.execute

A charachter string that will be passed as arguments PSN's `execute`.

table.vars

Variables to be printed in output table as a character vector or a space-separated string of variable names. The default is to export the same tables as listed in the input control stream. If table.vars is provided, all output tables in estimation control streams are dropped and replaced by a new one with just the provided variables. If many variables are exported, and much fewer are used, it can speed up NMsim significantly to only export what is needed (sometimes this is as little as "PRED IPRED"). Nonmem writes data slowly so reducing output data can make a big difference in execution time. See table.options too.

table.options

A character vector or a string of space-separated options. Only used if table.vars is provided. If constructing a new output table with table.vars the default is to add two options, NOAPPEND and NOPRINT. You can modeify that with table.options. Do not try to modify output filename - NMsim takes care of that.

text.sim

A character string to be pasted into $SIMULATION. This must not contain seed or SUBPROBLEM which are handled separately. Default is to include "ONLYSIM". To avoid that, use text.sim="".

method.sim

A function (not quoted) that creates the simulation control stream and other necessary files for a simulation based on the estimation control stream, the data, etc. The default is called NMsim_default which will replace any estimation and covariance step by a simulation step. See details section on oter methods, and see examples and especially vignettes on how to use the different provided methods.

execute

Execute the simulation or only prepare it? `execute=FALSE` can be useful if you want to do additional tweaks or simulate using other parameter estimates.

sge

Submit to cluster? Default is not to, but this is very useful if creating a large number of simulations, e.g. simulate with all parameter estimates from a bootstrap result.

nc

Number of cores used in parallelization. This is so far only supported with method.execute="psn".

transform

A list defining transformations to be applied after the Nonmem simulations and before plotting. For each list element, its name refers to the name of the column to transform, the contents must be the function to apply.

method.execute

Specify how to call Nonmem. Options are "psn" (PSN's execute), "nmsim" (an internal method similar to PSN's execute), and "direct" (just run Nonmem directly and dump all the temporary files). "nmsim" has advantages over "psn" that makes it the only supported method when type.sim="NMsim_known". "psn" has the simple advantage that the path to nonmem does not have to be specified if "execute" is in the system search path. So as long as you know where your Nonmem executable is, "nmsim" is recommended. The default is "nmsim" if path.nonmem is specified, and "psn" if not.

method.update.inits

The initial values of all parameters are by updated from the estimated model before running the simulation. NMsim can do this with a native function or use PSN to do it - or the step can be skipped to not update the values. The possible values are

  • "psn" uses PSN's "update_inits". Requires a functioning PSN installation and possibly that dir.psn is correctly set. The advantages of this method are that it keeps comments in the control stream and that it is a method known to many.

  • "nmsim" Uses a simple internal method to update the parameter values based on the ext file. The advantages of "nmsim" are it does not require PSN, and that it is very robust. "nmsim" fixes the whole OMEGA and SIGMA matrices as single blocks making the $OMEGA and $SIGMA sections of the control streams less easy to read. On the other hand, this method is robust because it avoids any interpretation of BLOCK structure or other code in the control streams.

  • "none" Do nothing. This is useful if the model to simulate has not been estimated but parameter values have been manually put into the respective sections in the control stream.

create.dirs

If the directories specified in dir.sims and dir.res do not exists, should it be created? Default is TRUE.

dir.psn

The directory in which to find PSN's executables ('execute' and 'update_inits'). The default is to rely on the system's search path. So if you can run 'execute' and 'update_inits' by just typing that in a terminal, you don't need to specify this unless you want to explicitly use a specific installation of PSN on your system.

list.sections

Named list of additional control stream section edits. Note, these can be functions that define how to edit sections. This is an advanced feature which is not needed to run most simulations. It is however powerful for some types of analyses, like modifying parameter values. See vignettes for further information. Documentation still under development.

sim.dir.from.scratch

If TRUE (default) this will wipe the simulation directory before running new simulations. The directory that will be emptied is _not_ dir.sims where you may keep many or all your simulations. It is the subdirectory named based on the run name and name.sim. The reason it is advised to wipe this directory is that if you in a previous simulation created simulation runs that are now obsolete, you could end up reading those too when collecting the results. NMsim will delete previously generated simulation control streams with the same name, but this option goes further. An example where it is important is if you first ran 1000 replications, fixed something and now rand 500. If you choose FALSE here, you can end up with the results of 500 new and 500 old simulations.

col.row

Only used if data is not supplied (which is most likely for simulations for VPCs) A column name to use for a row identifier. If none is supplied, NMdataConf()[['col.row']] will be used. If the column already exists in the data set, it will be used as is, if not it will be added.

args.NMscanData

If execute=TRUE&sge=FALSE, NMsim will normally read the results using NMreadSim. Use this argument to pass additional arguments (in a list) to that function if you want the results to be read in a specific way. This can be if the model for some reason drops rows, and you need to merge by a row identifier. You would do `args.NMscanData=list(col.row="ROW")` to merge by a column called `ROW`. This is only used in rare cases.

path.nonmem

The path to the Nonmem executable to use. The could be something like "/usr/local/NONMEM/run/nmfe75" (which is a made up example). No default is available. You should be able to figure this out through how you normally execute Nonmem, or ask a colleague.

nmquiet

Silent messages from Nonmem.

as.fun

The default is to return data as a data.frame. Pass a function (say `tibble::as_tibble`) in as.fun to convert to something else. If data.tables are wanted, use as.fun="data.table". The default can be configured using NMdataConf.

suffix.sim

Deprecated. Use name.sim instead.

text.table

A character string including the variables to export from Nonmem.

system.type

A charachter string, either \"windows\" or \"linux\" - case insensitive. Windows is only experimentally supported. Default is to use Sys.info()[["sysname"]].

dir.res

Provide a path to a directory in which to save rds files with paths to results. Default is to use dir.sims. After running `NMreadSim()` on these files, the original simulation files can be deleted. Hence, providing both `dir.sims` and `dir.res` provides a structure that is simple to clean. `dir.sims` can be purged when `NMreadSim` has been run and only small `rds` and `fst` files will be kept in `dir.res`. Notice, in case multiple models are simulated, multiple `rds` (to be read with `NMreadSim()`) files will be created by default. In cases where multiple models are simulated, see `file.res` to get just one file refering to all simulation results.

file.res

Path to an rds file that will contain a table of the simulated models. This is useful for subsequently retrieving all the results using `NMreadSim()`. The default is to create a file called `NMsim_paths.rds` under the model simulation directory. However, if multiple models are simulated, this will result in multiple rds files. Specifying a path ensures that one rds file containing information about all simulated models will be created.

wait

Wait for simulations to finish? Default is to do so if simulations are run locally but not to if they are sent to the cluster. Waiting for them means that the results will be read when simulations are done. If not waiting, path(s) to `rds` files to read will be returned. Pass them through `NMreadSim()` (which also supports waiting for the simulations to finish).

quiet

If TRUE, messages from what is going on will be suppressed to the extend implemented.

check.mod

Check the provided control streams for contents that may cause issues for simulation. Default is `TRUE`, and it is only recommended to disable this if you are fully aware of such a feature of your control stream, you know how it impacts simulation, and you want to get rid of warnings.

seed

Deprecated. See seed.R and seed.nm.

...

Additional arguments passed to method.sim.

Value

A data.frame with simulation results (same number of rows as input data). If `sge=TRUE` a character vector with paths to simulation control streams.

Details

Loosely speaking, the argument method.sim defines _what_ NMsim will do, method.execute define _how_ it does it. method.sim takes a function that converts an estimation control stream into whatever should be run. Features like replacing `$INPUT`, `$DATA`, `$TABLE`, and handling seeds are NMsim features that are done in addition to the method.sim. Also the list.sections argument is handled in addition to the method.sim. The subproblems and seed arguments are available to all methods creating a $SIMULATION section.

Notice, the following functions are internally available to `NMsim` so you can run them by say method.sim=NMsim_known without quotes. To see the code of that method, type NMsim_known.

  • NMsim_default The default behaviour. Replaces any $ESTIMATION and $COVARIANCE sections by a $SIMULATION section.

  • NMsim_asis The simplest of all method. It does nothing (but again, NMsim handles `$INPUT`, `$DATA`, `$TABLE` and more. Use this for instance if you already created a simulation (or estimation actually) control stream and want NMsim to run it on different data sets.

  • NMsim_typical Like NMsim_default but with all ETAs=0, giving a "typical subject" simulation. Do not confuse this with a "reference subject" simulation which has to do with covariate values. Technically all ETAs=0 is obtained by replacing $OMEGA by a zero matrix.

  • NMsim_known Simulates _known_ subjects, meaning that it reuses ETA values from estimation run. This is what is refered to as emperical Bayes estimates. The .phi file from the estimation run must be found next to the .lst file from the estimation.This means that ID values in the (simulation) input data must be ID values that were used in the estimation too. Runs an $ESTIMATION MAXEVAL=0 but pulls in ETAs for the ID's found in data. No $SIMULATION step is run which may affect how for instance residual variability is simulated, if at all.

  • NMsim_VarCov Like NMsim_default but `$THETA`, `$OMEGA`, and `SIGMA` are drawn from distribution estimated in covariance step. This means that a successful covariance step must be available from the estimation. In case the simulation leads to negative diagonal elements in $OMEGA and $SIGMA, those values are truncated at zero. For simulation with parameter variability based on bootstrap results, use NMsim_default.