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.
NMsim(
file.mod,
data,
subproblems = NULL,
reuse.results = FALSE,
seed.R,
seed.nm,
name.sim,
table.vars,
table.options,
table.format = "s1PE16.9",
carry.out = TRUE,
method.sim = NMsim_default,
typical = FALSE,
inits,
modify,
filters,
sizes,
path.nonmem = NULL,
sge = FALSE,
nc = 1,
execute = TRUE,
script = NULL,
transform = NULL,
order.columns = TRUE,
method.execute,
nmrep,
col.flagn = FALSE,
sim.dir.from.scratch = TRUE,
create.dirs = TRUE,
dir.psn,
args.psn.execute,
args.NMscanData,
as.fun,
system.type = NULL,
dir.sims,
dir.res,
file.res,
wait,
text.sim = "",
auto.dv = TRUE,
clean,
quiet = FALSE,
nmquiet,
progress,
check.mod = TRUE,
format.data.complete = "rds",
text.table,
suffix.sim,
seed,
file.ext = NULL,
method.update.inits,
modify.model,
list.sections,
...
)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.
Path(s) to the input control stream(s) to run the simulation on. The output 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.
The simulation data as a data.frame or a list
of data.frames. If a list, the model(s) will be run on
each of the data sets in the list.
Number of subproblems to use as
SUBPROBLEMS in $SIMULATION block in Nonmem. The
default is subproblem=0 which means not to use
SUBPROBLEMS.
If simulation results found on file, should they be used? If TRUE and reading the results fail, the simulations will still be rerun.
A value passed to set.seed(). It is
recommended to 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.
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).
To avoid changing an exisiting seed in a control stream, use
seed.nm="asis".
In case method.sim=NMsim_EBE, seeds are not used.
Give all filenames related to the simulation a suffix. A short string describing the sim is recommended like "ph3_regimens".
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 very big
difference in execution time. See table.options too.
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 modify that with
table.options. Do not try to modify output filename -
NMsim takes care of that. See `table.format` too.
A format for `$TABLE`. Only used if `table.vars` is provided. Default is "s1PE16.9". NMsim needs a high-resolution format. The Nonmem default "s1PE11.4" is insufficient for simulation data sets of 1e5 rows or more.
Variables from input data that should be included in results. Default is to include everything. If working with large data sets, it may be wanted to provide a subset of the columns here. If doing very large simulations, this may also be a way to save memory.
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.
Run with all ETAs fixed to zero? Technically all
ETAs=0 is obtained by replacing $OMEGA by a zero
matrix. Default is `FALSE`.
Control the parameter values. `inits` is a list. The `method` element controls which method is used to do this, and this corresponds to the old `method.update.inits` argument. If using the new `method=nmsim` you can specify parameter values, fix/unfix them, and edit lower and upper limits for estimation.
`method="nmsim"`, all other arguments are passed to `NMwriteInits`. This is a flexible method that allows for modification of the parameter values and is expected to be the default method in the future. Example which will update the parameter values based on the available estimate, but with `THETA(2)=1.3`: `inits=list(method="nmsim","THETA(2)"=list(init=1.3))`. See `?NMwriteInits` too.
`method="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.
`method="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 does not rely on code-interpretation for generation of simulation control streams. "nmsim" fixes the whole OMEGA and SIGMA matrices as single blocks making the $OMEGA and $SIGMA sections of the control streams less easy to read. On the other hand, this method is robust because it avoids any interpretation of BLOCK structure or other code in the control streams.
`method="none"` Do nothing. This is useful if the model to simulate has not been estimated but parameter values have been manually put into the respective sections in the control stream.
See also `file.ext` which can now be handled by `inits` too. This change collects the update of the "initial" parameter values into one interface rather than multiple arguments.
Named list of additional control stream section edits. Note, these can be functions that define how to edit sections. This is an advanced feature which is not needed to run most simulations. It is however powerful for some types of analyses, like modifying parameter values. See vignettes for further information.
Edit data filters (`IGNORE`/`ACCEPT` statements) before running model. This should normally only be used if no data set is provided. It can be useful if simulating for a VPC but a different subset of data needs to be simulated than the one used for estimation. A common example on this is inclusion of BLQ's in the VPC even if they were excluded in the estimation. See `?NMreadFilters` which returns a table you can edit and pass to `filters`. You can also just pass a string representing the full set of filters to be used. If you pass a string, consider including "IGN=@" to avoid character rows, like the column headers.
If needed, adjust the `$SIZES` section by providing a list of arguments to `NMupdateSizes()`. Example: `sizes=list(PD=80)`. See `?NMupdateSizes` for details. Don't use arguments like `file.mod` and `newfile` which are handled internally.
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.
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.
Number of cores used in parallelization. Only used if `sge=TRUE`.
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.
The path to the script where this is run. For stamping of dataset so results can be traced back to code.
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.
reorder columns by calling
NMdata::NMorderColumns before saving dataset and
running simulations? Default is TRUE.
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_EBE". "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.
Include `NMREP` as counter of subproblems? The default is to do so if `subproblems>0`. This will insert a counter called `NMREP` in the `$ERROR` section and include that in the output table(s). At this point, nothing is done to avoid overwriting existing variables.
Only used if `data` is provided. Use this if you are including an exclusion flag column in data. However, what NMsim will then do is to require that column to equal `0` (zero) for the rows to be simulated. It is often better to subset the data before simulation. See `filters` too.
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.
If the directories specified in dir.sims and dir.res do not exists, should it be created? Default is TRUE.
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.
A charachter string that will be passed as arguments PSN's `execute`.
If table.options is used, NMsim
turns to `NMdata::NMscanData()` for a general method to read
the output tables. Use `args.NMscanData` 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. Better just stick to NMsim's optimized default
`table.options` and related methods for reading results.
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.
A charachter string, either \"windows\" or
\"linux\" - case insensitive. Windows is only experimentally
supported. Default is to use Sys.info()[["sysname"]].
The directory in which NMsim will store all generated files. Default is to create a folder called `NMsim` next to `file.mod`.
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.
Path to an rds file that will contain a table of
the simulated models and other metadata. This is needed for
subsequently retrieving all the results using
`NMreadSim()`. The default is to create a file called
`NMsim_..._MetaData.rds` under the dir.res directory
where ... is based on the model name. However, if multiple
models (file.mod) 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. Notice if file.res is supplied, dir.res
is not used.
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()`. Conveniently, NMreadSim() also takes the `wait` argument too, allowing flexibility to run Nonmem in the background, and then read the results, still waiting for Nonmem to finish.
A character string to be pasted into $SIMULATION. This must not contain seed or SUBPROBLEM which are handled separately. Default is to include "ONLYSIM". You cannot avoid that using `text.sim`. Instead, you can use `onlysim=FALSE` which will be passed to `NMsim_default()`.
Add a column called `DV` to input data sets if a
column of that name is not found? Nonmem is generally
dependent on a `DV` column in input data but this is typically
uninformative in simulation data sets and hence easily
forgotten when generating simulation data sets. If
auto.dv=TRUE and no `DV` column is found, `DV=NA` will
be added. In this case (`auto.dv=TRUE` and no `DV` column
found) a `MDV=1` column will also be added if none found.
The degree of cleaning (file removal) to do after Nonmem execution. If `method.execute=="psn"`, this is passed to PSN's `execute`. If `method.execute=="nmsim"` a similar behavior is applied, even though not as granular. NMsim's internal method only distinguishes between 0 (no cleaning), any integer 1-4 (default, quite a bit of cleaning) and 5 (remove temporary dir completely).
If TRUE, messages from what is going on will be suppressed.
Silent console messages from Nonmem? The default behaviour depends. It is FALSE if there is only one model to execute and `progress=FALSE`.
Track progress? Default is `TRUE` if `quiet` is FALSE and more than one model is being simulated. The progress tracking is based on the number of models completed, not the status of the individual models.
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.
For development purposes - users do not need this argument. Controls what format the complete input data set is saved in. Possible values are `rds` (default), `fst` (experimental) and `csv`. `fst` may be faster and use less disk space but factor levels may be lost from input data to output data. `csv` will also lead to loss of additional information such as factor levels.
Deprecated. Use `table.vars` and `table.options` instead.
Deprecated. Use name.sim instead.
Deprecated. See seed.R and seed.nm.
Depecated. Use `inits=list(file.ext="path/to/file.ext")` instead. Optionally provide a parameter estimate file from Nonmem. This is normally not needed since `NMsim` will by default use the ext file stored next to the input control stream (replacing the file name extension with `.ext`). If using method.update.inits="psn", this argument cannot be used.
Deprecated, please migrate to `inits` instead. 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.
Deprecated. Use modify instead.
Deprecated. Use modify instead.
Additional arguments passed to method.sim.
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 modeify.model argument
is handled in addition to the method.sim. The
subproblems and seed.nm 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_EBE
without quotes. To see the code of that method, type
NMsim_EBE.
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_EBE Simulates _known_ ETAs. By default, the ETA
values are automatically taken from the estimation run. This is
what is refered to as emperical Bayes estimates, hence the name of
the method "NMsim_EBE". However, the user can also provide a
different `.phi` file which may contain simulated ETA values (see
the `file.phi` argument). ID values in the simulation data set
must match ID values in the phi file for this step to work. If
refering to estimated subjects, the .phi file from the estimation
run must be found next to the .lst file from the estimation with
the same file name stem (say `run1.lst` and `run1.phi`). Again, ID
values in the (simulation) input data must be ID values that were
used in the estimation too. The method Runs an $ESTIMATION
MAXEVAL=0 but pulls in ETAs for the ID's found in data. No
$SIMULATION step is run which unfortunately means no
residual error will be simulated.
NMsim_VarCov Like NMsim_default but `$THETA`,
`$OMEGA`, and `SIGMA` are drawn from distribution estimated in
covariance step. This means that a successful covariance step must
be available from the estimation. NB. A multivariate normal
distribution is used for all parameters, including `$OMEGA` and
`$SIGMA` which is not the correct way to do this. In case the
simulation leads to negative diagonal elements in $OMEGA and
$SIGMA, those values are truncated at zero. This method is only
valid for simulation of `$THETA` variability. The method accepts a
table of parameter values that can be produced with other tools
than `NMsim`. For simulation with parameter variability based on
bootstrap results, use NMsim_default.
NMsim_typical Deprecated. Use typical=TRUE instead.