runSimulation: Run a Monte Carlo simulation given conditions and simulation...
In philchalmers/SimDesign: Structure for Organizing Monte Carlo Simulation Designs
View source: R/runSimulation.R
runSimulation R Documentation
Run a Monte Carlo simulation given conditions and simulation functions
Description
This function runs a Monte Carlo simulation study given a set of predefined simulation functions,
design conditions, and number of replications. Results can be saved as temporary files in case of
interruptions and may be restored by re-running runSimulation, provided that the respective temp
file can be found in the working directory. runSimulation supports parallel
and cluster computing (with the parallel and future packages; see also
runArraySimulation for submitting array jobs to HPC clusters),
global and local debugging, error handling (including fail-safe
stopping when functions fail too often, even across nodes), provides bootstrap estimates of the
sampling variability (optional), and automatic tracking of error and warning messages
with their associated .Random.seed states.
For convenience, all functions available in the R work-space are exported across all nodes
so that they are more easily accessible (however, other R objects are not, and therefore
must be passed to the fixed_objects input to become available across nodes).
Usage
runSimulation(
design,
replications,
generate,
analyse,
summarise,
fixed_objects = NULL,
packages = NULL,
filename = NULL,
debug = "none",
load_seed = NULL,
save = any(replications > 10),
store_results = TRUE,
save_results = FALSE,
parallel = FALSE,
ncores = parallel::detectCores() - 1L,
cl = NULL,
notification = "none",
beep = FALSE,
sound = 1,
CI = 0.95,
seed = NULL,
boot_method = "none",
boot_draws = 1000L,
max_errors = 50L,
resume = TRUE,
save_details = list(),
control = list(),
progress = TRUE,
verbose = TRUE
)
## S3 method for class 'SimDesign'
summary(object, ...)
## S3 method for class 'SimDesign'
print(x, list2char = TRUE, ...)
Arguments
design
a tibble or data.frame object containing the Monte Carlo simulation
conditions to be studied, where each row represents a unique condition and each column a factor
to be varied. See createDesign for the standard approach
to create this simulation design object
replications
number of independent replications to perform per
condition (i.e., each row in design). Can be a single number, which
will be used for each design condition, or an integer vector with length
equal to nrow(design). All inputs must be greater than 0
generate
user-defined data and parameter generating function (or named list of functions).
See Generate for details. Note that this argument may be omitted by the
user if they wish to generate the data with the analyse step, but for real-world
simulations this is generally not recommended. If multiple generate functions are provided
as a list then the list of generate functions are executed in order until the first valid
generate function is executed, where the subsequent generation functions are then ignored
(see GenerateIf to only apply data generation for specific conditions).
analyse
user-defined analysis function (or named list of functions)
that acts on the data generated from
Generate (or, if generate was omitted, can be used to generate and
analyses the simulated data). See Analyse for details
summarise
optional (but strongly recommended) user-defined summary function
from Summarise to be used to compute meta-statistical summary
information after all the replications have completed within
each design condition. Note that unlike the Generate and Analyse
steps, the Summarise portion is not as important to perfectly organize
as the results can be summarised later on by using the built-in
reSummarise function (provided either
store_results = TRUE or save_results = TRUE were included).
Omitting this function will return a tibble with the Design
and associated results information for all
nrow(Design) * repliations evaluations if the results from each
Analyse() call was a one-dimensional vector.
For more general objects returned by Analyse()
(such as lists), a list
containing the results returned form Analyse.
This is generally only recommended for didactic purposes because the results
will leave out a large amount of
information (e.g., try-errors, warning messages, saving files, etc), can
witness memory related issues if the Analyse function returns larger objects,
and generally is not as flexible internally. However, it may be useful
when replications are expensive and ANOVA-based decompositions involving
the within-condition replication information are of interest, though
of course this can be circumvented by using store_results = TRUE or
save_results = TRUE with or without a supplied summarise
definition.
fixed_objects
(optional) an object (usually a named list)
containing additional user-defined objects
that should remain fixed across conditions. This is useful when including
large vectors/matrices of population parameters, fixed data information
that should be used across all conditions and replications (e.g., including a
common design matrix for linear regression models), or simply control
constant global elements (e.g., a constant for sample size)
packages
a character vector of external packages to be used during the simulation (e.g.,
c('MASS', 'extraDistr', 'simsem') ). Use this input when running code in
parallel to use non-standard functions from additional packages,
otherwise the functions must be made available by using explicit
library or require calls within the provided simulation functions.
Alternatively, functions can be called explicitly without attaching the package
with the :: operator
(e.g., extraDistr::rgumbel())
filename
(optional) the name of the .rds file to save the final
simulation results to. If the extension
.rds is not included in the file name (e.g. "mysimulation"
versus "mysimulation.rds") then the
.rds extension will be automatically added to the file name to ensure
the file extension is correct.
Note that if the same file name already exists in the working
directly at the time of saving then a new
file will be generated instead and a warning will be thrown. This helps to
avoid accidentally overwriting
existing files. Default is NULL, indicating no file will be saved by default
debug
a string indicating where to initiate a browser() call for editing
and debugging, and pairs particularly well with the load_seed argument for precise debugging.
General options are 'none' (default; no debugging), 'error', which
starts the debugger
when any error in the code is detected in one of three generate-analyse-summarise functions,
and 'all', which debugs all the user defined functions regardless of
whether an error was thrown
or not. Specific options include: 'generate'
to debug the data simulation function, 'analyse' to debug the computational function, and
'summarise' to debug the aggregation function.
If the Analyse argument is supplied as a named list of functions then it is also possible
to debug the specific function of interest by passing the name of the respective function in the list.
For instance, if analyse = list(A1=Analyse.A1, A2=Analyse.A2) then passing
debug = 'A1' will debug only the first function in this list, and all remaining analysis
functions will be ignored.
For debugging specific rows in the Design input (e.g.,
when a number of initial rows successfully complete but the kth
row fails) the row number can be appended to the standard
debug input using a '-' separator.
For instance, debugging whenever an error is raised
in the second row of Design can be declared via debug = 'error-2'.
Finally, users may place browser calls within the respective functions for
debugging at specific lines, which is useful when debugging based on conditional evaluations (e.g.,
if(this == 'problem') browser()). Note that parallel computation flags
will automatically be disabled when a browser() is detected or when a debugging
argument other than
'none' is supplied
load_seed
used to replicate an exact simulation state, which is
primarily useful for debugging purposes.
Input can be a character object indicating which file to load from when the
.Random.seeds have
be saved (after a call with save_seeds = TRUE), or an integer vector
indicating the actual
.Random.seed values (e.g., extracted after using store_seeds).
E.g., load_seed = 'design-row-2/seed-1'
will load the first seed in the second row of the design input, or
explicitly passing the
elements from .Random.seed (see SimExtract to extract
the seeds associated explicitly
with errors during the simulation, where each column represents a unique seed).
If the input is a character vector then it is important NOT
to modify the design input object, otherwise the path may not point
to the correct saved location, while
if the input is an integer vector (or single column tbl object)
then it WILL be important to modify the design input in order to load this
exact seed for the corresponding design row. Default is NULL
save
logical; save the temporary simulation state to the hard-drive? This is useful
for simulations which require an extended amount of time, though for shorter simulations
can be disabled to slightly improve computational efficiency. When TRUE,
which is the default when evaluating replications > 10, a temp file
will be created in the working directory which allows the simulation state to be saved
and recovered (in case of power outages, crashes, etc). As well, triggering this flag will
save any fatal .Random.seed states when conditions unexpectedly crash (where each seed
is stored row-wise in an external .rds file), which provides a much easier mechanism
to debug issues (see load_seed for details). Upon completion, this temp file will
be removed.
To recover your simulation at the last known location (having patched the issues in the
previous execution code) simply re-run the code you used to
initially define the simulation and the external file will automatically be detected and read-in.
Default is TRUE when replications > 10 and FALSE otherwise
store_results
logical; store the complete tables of simulation results
in the returned object? This is TRUE default, though if RAM anticipated to
be an issue see save_results instead. Note that if the Design
object is omitted from the call to runSimulation(), or the number of rows in Design
is exactly 1, then this argument is automatically set to TRUE as RAM storage is no
longer an issue.
To extract these results
pass the returned object to either SimResults or SimExtract with
what = 'results', which will return a named list
of all the simulation results for each condition if nrow(Design) > 1; otherwise, if
nrow(Design) == 1 or Design was missing the results object will be stored as-is
save_results
logical; save the results returned from Analyse to external
.rds files located in the defined save_results_dirname directory/folder?
Use this if you would like to keep track of the individual parameters returned from
the analysis function.
Each saved object will contain a list of three elements containing the
condition (row from design),
results (as a list or matrix), and try-errors.
See SimResults for an example of how to read these .rds files back into R
after the simulation is complete. Default is FALSE.
WARNING: saving results to your hard-drive can fill up space very quickly for
larger simulations. Be sure to
test this option using a smaller number of replications before the full Monte
Carlo simulation is performed.
See also reSummarise for applying summarise functions from saved
simulation results
parallel
logical; use parallel processing from the parallel
package over each unique condition? This distributes the independent replications
across the defined nodes, and is repeated for each row condition in the design
input.
Alternatively, if the future package approach is desired then passing
parallel = 'future' to runSimulation() will use the defined
plan for execution. This allows for greater flexibility when
specifying the general computing plan (e.g., plan(multisession)) for parallel computing
on the same machine, plan(future.batchtools::batchtools_torque) or
plan(future.batchtools::batchtools_slurm) for common MPI schedulers, etc).
However, it is the responsibility of the user to use plan(sequential) to reset the
computing plan when the jobs are completed
ncores
number of cores to be used in parallel execution (ignored if using the
future package approach). Default uses all available minus 1
cl
cluster object defined by makeCluster used to run code in parallel
(ignored if using the future package approach).
If NULL and parallel = TRUE, a local cluster object will be defined which
selects the maximum number cores available
and will be stopped when the simulation is complete. Note that supplying a cl
object will automatically set the parallel argument to TRUE. Define and supply this
cluster object yourself whenever you have multiple nodes to chain together (note in this case
that you must use either the "MPI" or "PSOCK" clusters).
Note that if the future package has
been attached prior to executing runSimulation() then the associated
plan() will be followed instead
notification
an optional character vector input that can be used to send
Pushbullet notifications from a configured
computer. This reports information such as the total execution time, the condition
completed, and error/warning
messages recorded. This arguments assumes that users have already A) registered for
a Pushbullet account,
B) installed the application on their mobile device and computer, and C) created an
associated JSON file of the form
~/.rpushbullet.json using RPushbullet::pbSetup()).
To utilize the RPushbullet in SimDesign first call library(RPushbullet
before running runSimulation() to read-in the default JSON file. Next,
pass one of the following supported
options: 'none' (default; send no notification),
'condition' to send a notification after each condition has completed,
or 'complete' to send
a notification only when the simulation has finished.
beep
logical; call the beepr package when the simulation is completed?
sound
sound argument passed to beepr::beep()
CI
bootstrap confidence interval level (default is 95%)
seed
a vector or list of integers to be used for reproducibility.
The length of the vector must be equal the number of rows in design.
If the input is a vector then set.seed or
clusterSetRNGStream for each condition will be called, respectively,
but will not be run when MPI = TRUE. If a list is provided then these
numbers must have been generated from gen_seeds with the argument
CMRG.seed used to specify the initial. The list approach ensures random number
generation independence across conditions and replications, while the vector input
ensures independence within the replications per conditions but not necessarily
across conditions. Default randomly generates seeds within the
range 1 to 2147483647 for each condition via link{gen_seeds}
boot_method
method for performing non-parametric bootstrap confidence intervals
for the respective meta-statistics computed by the Summarise function.
Can be 'basic' for the empirical bootstrap CI, 'percentile'
for percentile CIs, 'norm' for normal approximations CIs, or 'studentized'
for Studentized CIs (should only be used for simulations with lower replications due to its
computational intensity). Alternatively, CIs can be constructed using the argument 'CLT',
which computes the intervals according to the large-sample standard error
approximation SD(results)/\sqrt{R}. Default is 'none', which performs no CI computations
boot_draws
number of non-parametric bootstrap draws to sample for the summarise
function after the generate-analyse replications are collected. Default is 1000
max_errors
the simulation will terminate when more than this number of consecutive
errors are thrown in any
given condition, causing the simulation to continue to the next unique design condition.
This is included to avoid getting stuck in infinite re-draws, and to indicate that
something fatally problematic
is going wrong in the generate-analyse phases. Default is 50
resume
logical; if a temporary SimDesign file is detected should
the simulation resume from this location? Keeping this TRUE is generally recommended,
however this should be disabled if using runSimulation within runSimulation to avoid
reading improper save states. Alternatively, if an integer is supplied then the simulation
will continue at the associated row location in design (e.g., resume=10).
This is useful to overwrite a previously evaluate element in the temporary files that was detected
to contain fatal errors that require re-evaluation without discarding the originally valid rows
in the simulation
save_details
a list pertaining to information regarding how and where files should be saved
when the save or save_results flags are triggered.
safelogical; trigger whether safe-saving should be performed. When TRUE files
will never be overwritten accidentally, and where appropriate the program will either stop or generate
new files with unique names. Default is TRUE
compnamename of the computer running the simulation. Normally this doesn't need
to be modified, but in the event that a manual node breaks down while running a simulation the
results from the temp files may be resumed on another computer by changing the name of the
node to match the broken computer. Default is the result of evaluating
unname(Sys.info()['nodename'])
out_rootdirroot directory to save all files to. Default uses the
current working directory
save_results_dirnamea string indicating the name of the folder to save
result objects to when save_results = TRUE. If a directory/folder does not exist
in the current working directory then a unique one will be created automatically. Default is
'SimDesign-results_' with the associated compname appended if no
filename is defined, otherwise the filename is used to replace 'SimDesign'
in the string
save_results_filenamea string indicating the name file to store, where the
Design row ID will be appended to ensure uniqueness across rows. Specifying
this input will disable any checking for the uniqueness of the file folder, thereby
allowing independent runSimulation calls to write to the same
save_results_dirname. Useful when the files should all be stored in the same
working directory, however the rows of Design are evaluated in isolation (e.g.,
for HPC structures that allow asynchronous file storage).
WARNING: the uniqueness of the file names are not checked using
this approach, therefore please ensure that each generated name will be unique a priori,
such as naming the file based on the supplied row condition information
save_seeds_dirnamea string indicating the name of the folder to save
.Random.seed objects to when save_seeds = TRUE. If a directory/folder
does not exist
in the current working directory then one will be created automatically. Default is
'SimDesign-seeds_' with the associated compname appended if no
filename is defined, otherwise the filename is used to replace 'SimDesign'
in the string
control
a list for extra information flags for controlling less
commonly used features. These include
stop_on_fatallogical (default is FALSE); should the simulation be
terminated immediately when
the maximum number of consecutive errors (max_errors) is reached? If FALSE,
the simulation will continue as though errors did not occur, however a column
FATAL_TERMINATION will be included in the resulting object indicating the final
error message observed, and NA placeholders will be placed in all other row-elements.
Default is FALSE
warnings_as_errorslogical (default is FALSE);
treat warning messages as error messages during the simulation? Default is FALSE,
therefore warnings are only collected and not used to restart the data generation step,
and the seeds associated with
the warning message conditions are not stored within the final simulation object.
Note that this argument is generally intended for debugging/early planning
stages when designing a simulation experiment. If specific warnings are known to
be problematic and should be treated as errors then please use
convertWarnings instead
save_seeds-
logical; save the .Random.seed states prior to performing
each replication into
plain text files located in the defined save_seeds_dirname directory/folder?
Use this if you would like to keep track of every simulation state within each
replication and design
condition. This can be used to completely replicate any cell in the simulation if need be.
As well, see the load_seed input
to load a given .Random.seed to exactly replicate the generated data and
analysis state (mostly useful
for debugging). When TRUE, temporary files will also be saved
to the working directory (in the same way as when save = TRUE).
Default is FALSE
Note, however, that this option is not typically necessary or recommended since
the .Random.seed states for simulation
replications that throw errors during the execution are automatically stored
within the final simulation
object, and can be extracted and investigated using SimExtract.
Hence, this option is only of
interest when all of the replications must be reproducible (which occurs very rarely),
otherwise the defaults to runSimulation should be sufficient
store_Random.seedslogical; store the
complete .Random.seed states
for each simulation replicate? Default is FALSE as this can
take up a great deal of unnecessary RAM (see related save_seeds),
however this may be useful
when used with runArraySimulation. To extract use
SimExtract(..., what = 'stored_Random.seeds')
store_warning_seedslogical (default is FALSE);
in addition to storing the .Random.seed states whenever error messages
are raised, also store the .Random.seed states when warnings are raised? This is
disabled by default
since warnings are generally less problematic than errors, and because many more
warnings messages may be raised
throughout the simulation (potentially causing RAM related issues when constructing
the final simulation object as
any given simulation replicate could generate numerous warnings, and storing the seeds
states could add up quickly).
Set this to TRUE when replicating warning messages is important, however be aware
that too many warnings messages raised during the simulation implementation could cause
RAM related issues.
include_replication_index or
include_repslogical (default is FALSE);
should a REPLICATION element be added to
the condition object when performing the simulation to track which specific
replication experiment is being evaluated? This is useful when, for instance, attempting
to run external software programs (e.g., Mplus) that require saving temporary data sets
to the hard-drive (see the Wiki for examples)
try_all_analyselogical; when analyse is a list, should every generated
data set be analyzed by each function definition in the analyse list?
Default is TRUE.
Note that this TRUE default can be computationally demanding when some analysis
functions require more computational resources than others, and the data should be
discarded early as an invalid candidate (e.g., estimating a model via maximum-likelihood
in on analyze component, while estimating a model using MCMC estimation on another). Hence,
the main benefit of using FALSE instead is that the data set may be rejected earlier,
where easier/faster to estimate analyse definitions should be placed earlier in the list
as the functions are evaluated in sequence
(e.g., Analyse = list(MLE=MLE_definition, MCMC=MCMC_definition))
allow_nalogical (default is FALSE); should NAs be allowed in the
analyse step as a valid result from the simulation analysis?
allow_nanlogical (default is FALSE); should NaNs be allowed in the
analyse step as a valid result from the simulation analysis?
typedefault type of cluster to create for the cl object if no supplied.
For Windows OS this defaults to "PSOCK", otherwise "SOCK" is selected
(suitable for Linux and Mac OSX). This is ignored if the user specifies their own cl object
MPIlogical (default is FALSE); use the foreach package in a
form usable by MPI to run simulation in parallel on a cluster?
print_RAMlogical (default is TRUE); print the amount of RAM
used throughout the simulation? Set to FALSE if unnecessary or if the call to
gc is unnecessarily time consuming
.options.mpilist of arguments passed to foreach() to control the MPI execution
properties. Only used when MPI = TRUE
progress
logical; display a progress bar (using the pbapply package)
for each simulation condition?
This is useful when simulations conditions take a long time to run (see also the
notifications argument). Default is TRUE
verbose
logical; print messages to the R console? Default is TRUE
object
SimDesign object returned from runSimulation
...
additional arguments
x
SimDesign object returned from runSimulation
list2char
logical; for tibble object re-evaluate list elements
as character vectors for better printing of the levels? Note that this
does not change the original classes of the object, just how they are printed.
Default is TRUE
Details
For an in-depth tutorial of the package please refer to Chalmers and Adkins (2020;
\Sexpr[results=rd]{tools:::Rd_expr_doi("10.20982/tqmp.16.4.p248")}).
For an earlier didactic presentation of the package refer to Sigal and Chalmers
(2016; \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1080/10691898.2016.1246953")}). Finally, see the associated
wiki on Github (https://github.com/philchalmers/SimDesign/wiki)
for tutorial material, examples, and applications of SimDesign to real-world
simulation experiments, as well as the various vignette files associated with the package.
The strategy for organizing the Monte Carlo simulation work-flow is to
- 1)
Define a suitable Design object (a tibble or data.frame)
containing fixed conditional
information about the Monte Carlo simulations. Each row or this design object pertains
to a unique set of simulation to study, while each column the simulation factor under
investigation (e.g., sample size,
distribution types, etc). This is often expedited by using the
createDesign function, and if necessary the argument subset
can be used to remove redundant or non-applicable rows
- 2)
Define the three step functions to generate the data (Generate; see also
https://CRAN.R-project.org/view=Distributions for a list of distributions in R),
analyse the generated data by computing the respective parameter estimates, detection rates,
etc (Analyse), and finally summarise the results across the total
number of replications (Summarise).
- 3)
Pass the design object and three defined R functions to runSimulation,
and declare the number of replications to perform with the replications input.
This function will return a suitable
tibble object with the complete simulation results and execution details
- 4)
Analyze the output from runSimulation, possibly using ANOVA techniques
(SimAnova) and generating suitable plots and tables
Expressing the above more succinctly, the functions to be called have the following form,
with the exact functional arguments listed:
Design <- createDesign(...)Generate <- function(condition, fixed_objects = NULL) {...} Analyse <- function(condition, dat, fixed_objects = NULL) {...} Summarise <- function(condition, results, fixed_objects = NULL) {...} res <- runSimulation(design=Design, replications, generate=Generate,
analyse=Analyse, summarise=Summarise)
The condition object above represents a single row from the design object, indicating
a unique Monte Carlo simulation condition. The condition object also contains two
additional elements to help track the simulation's state: an ID variable, indicating
the respective row number in the design object, and a REPLICATION element
indicating the replication iteration number (an integer value between 1 and replication).
This setup allows users to easily locate the rth replication (e.g., REPLICATION == 500)
within the jth row in the simulation design (e.g., ID == 2). The
REPLICATION input is also useful when temporarily saving files to the hard-drive
when calling external command line utilities (see examples on the wiki).
For a template-based version of the work-flow, which is often useful when initially
defining a simulation, use the SimFunctions function. This
function will write a template simulation
to one/two files so that modifying the required functions and objects can begin immediately.
This means that users can focus on their Monte Carlo simulation details right away rather
than worrying about the repetitive administrative code-work required to organize the simulation's
execution flow.
Finally, examples, presentation files, and tutorials can be found on the package wiki located at
https://github.com/philchalmers/SimDesign/wiki.
Value
a tibble from the dplyr package (also of class 'SimDesign')
with the original design conditions in the left-most columns,
simulation results in the middle columns, and additional information in the right-most columns (see below).
The right-most column information for each condition are:
REPLICATIONS to indicate the number of Monte Carlo replications,
SIM_TIME to indicate how long (in seconds) it took to complete
all the Monte Carlo replications for each respective design condition,
RAM_USED amount of RAM that was in use at the time of completing
each simulation condition,
COMPLETED to indicate the date in which the given simulation condition completed,
SEED for the integer values in the seed argument (if applicable; not
relevant when L'Ecuyer-CMRG method used), and, if applicable,
ERRORS and WARNINGS which contain counts for the number of error or warning
messages that were caught (if no errors/warnings were observed these columns will be omitted).
Note that to extract the specific error and warnings messages see
SimExtract. Finally,
if boot_method was a valid input other than 'none' then the final right-most
columns will contain the labels
BOOT_ followed by the name of the associated meta-statistic defined in summarise() and
and bootstrapped confidence interval location for the meta-statistics.
Saving data, results, seeds, and the simulation state
To conserve RAM, temporary objects (such as data generated across conditions and replications)
are discarded; however, these can be saved to the hard-disk by passing the appropriate flags.
For longer simulations it is recommended to use the save_results flag to write the
analysis results to the hard-drive.
The use of the store_seeds or the save_seeds options
can be evoked to save R's .Random.seed
state to allow for complete reproducibility of each replication within each condition. These
individual .Random.seed terms can then be read in with the
load_seed input to reproduce the exact simulation state at any given replication.
Most often though, saving the complete list of seeds is unnecessary as problematic seeds are
automatically stored in the final simulation object to allow for easier replicability
of potentially problematic errors (which incidentally can be extracted
using SimExtract(res, 'error_seeds') and passed to the load_seed argument). Finally,
providing a vector of seeds is also possible to ensure
that each simulation condition is macro reproducible under the single/multi-core method selected.
Finally, when the Monte Carlo simulation is complete
it is recommended to write the results to a hard-drive for safe keeping, particularly with the
filename argument provided (for reasons that are more obvious in the parallel computation
descriptions below). Using the filename argument supplied is safer than using, for instance,
saveRDS directly because files will never accidentally be overwritten,
and instead a new file name will be created when a conflict arises; this type of implementation safety
is prevalent in many locations in the package to help avoid unrecoverable (yet surprisingly
common) mistakes during the process of designing and executing Monte Carlo simulations.
Resuming temporary results
In the event of a computer crash, power outage, etc, if save = TRUE was used (the default)
then the original code used to execute runSimulation() need only be re-run to resume
the simulation. The saved temp file will be read into the function automatically, and the
simulation will continue one the condition where it left off before the simulation
state was terminated. If users wish to remove this temporary
simulation state entirely so as to start anew then simply pass SimClean(temp = TRUE)
in the R console to remove any previously saved temporary objects.
A note on parallel computing
When running simulations in parallel (either with parallel = TRUE or MPI = TRUE,
or when using the future approach with a plan() other than sequential)
R objects defined in the global environment will generally not be visible across nodes.
Hence, you may see errors such as Error: object 'something' not found if you try to use
an object that is defined in the work space but is not passed to runSimulation.
To avoid this type or error, simply pass additional objects to the
fixed_objects input (usually it's convenient to supply a named list of these objects).
Fortunately, however, custom functions defined in the global environment are exported across
nodes automatically. This makes it convenient when writing code because custom functions will
always be available across nodes if they are visible in the R work space. As well, note the
packages input to declare packages which must be loaded via library() in order to make
specific non-standard R functions available across nodes.
Author(s)
Phil Chalmers rphilip.chalmers@gmail.com
References
Chalmers, R. P., & Adkins, M. C. (2020). Writing Effective and Reliable Monte Carlo Simulations
with the SimDesign Package. The Quantitative Methods for Psychology, 16(4), 248-280.
\Sexpr[results=rd]{tools:::Rd_expr_doi("10.20982/tqmp.16.4.p248")}
Sigal, M. J., & Chalmers, R. P. (2016). Play it again: Teaching statistics with Monte
Carlo simulation. Journal of Statistics Education, 24(3), 136-156.
\Sexpr[results=rd]{tools:::Rd_expr_doi("10.1080/10691898.2016.1246953")}
See Also
SimFunctions, createDesign,
Generate, Analyse, Summarise,
SimExtract,
reSummarise, SimClean, SimAnova, SimResults,
aggregate_simulations, Attach, AnalyseIf,
SimShiny, convertWarnings, runArraySimulation
Examples
#-------------------------------------------------------------------------------
# Example 1: Sampling distribution of mean
# This example demonstrate some of the simpler uses of SimDesign,
# particularly for classroom settings. The only factor varied in this simulation
# is sample size.
# skeleton functions to be saved and edited
SimFunctions()
#### Step 1 --- Define your conditions under study and create design data.frame
Design <- createDesign(N = c(10, 20, 30))
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Step 2 --- Define generate, analyse, and summarise functions
# help(Generate)
Generate <- function(condition, fixed_objects = NULL) {
dat <- with(condition, rnorm(N, 10, 5)) # distributed N(10, 5)
dat
}
# help(Analyse)
Analyse <- function(condition, dat, fixed_objects = NULL) {
ret <- c(mean=mean(dat)) # mean of the sample data vector
ret
}
# help(Summarise)
Summarise <- function(condition, results, fixed_objects = NULL) {
# mean and SD summary of the sample means
ret <- c(mu=mean(results$mean), SE=sd(results$mean))
ret
}
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Step 3 --- Collect results by looping over the rows in design
# run the simulation
Final <- runSimulation(design=Design, replications=10,
generate=Generate, analyse=Analyse, summarise=Summarise)
Final
SimResults(Final)
## Not run:
# reproduce exact simulation
Final_rep <- runSimulation(design=Design, replications=10, seed=Final$SEED,
generate=Generate, analyse=Analyse, summarise=Summarise)
Final_rep
# run with more standard number of replications (note the storage message)
Final <- runSimulation(design=Design, replications=1000,
generate=Generate, analyse=Analyse, summarise=Summarise)
Final
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Extras
# compare SEs estimates to the true SEs from the formula sigma/sqrt(N)
5 / sqrt(Design$N)
# To store the results from the analyse function either
# a) omit a definition of summarise() to return all results,
# b) use store_results = TRUE (default) to store results internally and later
# extract with SimResults(), or
# c) pass save_results = TRUE to runSimulation() and read the results in with SimResults()
#
# Note that method c) should be adopted for larger simulations, particularly
# if RAM storage could be an issue and error/warning message information is important.
# a) approach
res <- runSimulation(design=Design, replications=5,
generate=Generate, analyse=Analyse)
res
# b) approach (store_results = TRUE by default)
res <- runSimulation(design=Design, replications=5,
generate=Generate, analyse=Analyse, summarise=Summarise)
res
SimResults(res)
# c) approach
Final <- runSimulation(design=Design, replications=5, save_results=TRUE,
generate=Generate, analyse=Analyse, summarise=Summarise)
# read-in all conditions (can be memory heavy)
res <- SimResults(Final)
res
head(res[[1]]$results)
# just first condition
res <- SimResults(Final, which=1)
head(res$results)
dplyr::tibble(res$condition, res$results)
# obtain empirical bootstrapped CIs during an initial run
# the simulation was completed (necessarily requires save_results = TRUE)
res <- runSimulation(design=Design, replications=1000, boot_method = 'basic',
generate=Generate, analyse=Analyse, summarise=Summarise)
res
# alternative bootstrapped CIs that uses saved results via reSummarise().
# Default directory save to:
dirname <- paste0('SimDesign-results_', unname(Sys.info()['nodename']), "/")
res <- reSummarise(summarise=Summarise, dir=dirname, boot_method = 'basic')
res
# remove the saved results from the hard-drive if you no longer want them
SimClean(results = TRUE)
## End(Not run)
#-------------------------------------------------------------------------------
# Example 2: t-test and Welch test when varying sample size, group sizes, and SDs
# skeleton functions to be saved and edited
SimFunctions()
## Not run:
# in real-world simulations it's often better/easier to save
# these functions directly to your hard-drive with
SimFunctions('my-simulation')
## End(Not run)
#### Step 1 --- Define your conditions under study and create design data.frame
Design <- createDesign(sample_size = c(30, 60, 90, 120),
group_size_ratio = c(1, 4, 8),
standard_deviation_ratio = c(.5, 1, 2))
Design
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Step 2 --- Define generate, analyse, and summarise functions
Generate <- function(condition, fixed_objects = NULL) {
N <- condition$sample_size # could use Attach() to make objects available
grs <- condition$group_size_ratio
sd <- condition$standard_deviation_ratio
if(grs < 1){
N2 <- N / (1/grs + 1)
N1 <- N - N2
} else {
N1 <- N / (grs + 1)
N2 <- N - N1
}
group1 <- rnorm(N1)
group2 <- rnorm(N2, sd=sd)
dat <- data.frame(group = c(rep('g1', N1), rep('g2', N2)), DV = c(group1, group2))
dat
}
Analyse <- function(condition, dat, fixed_objects = NULL) {
welch <- t.test(DV ~ group, dat)$p.value
independent <- t.test(DV ~ group, dat, var.equal=TRUE)$p.value
# In this function the p values for the t-tests are returned,
# and make sure to name each element, for future reference
ret <- nc(welch, independent)
ret
}
Summarise <- function(condition, results, fixed_objects = NULL) {
#find results of interest here (e.g., alpha < .1, .05, .01)
ret <- EDR(results, alpha = .05)
ret
}
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Step 3 --- Collect results by looping over the rows in design
# first, test to see if it works
res <- runSimulation(design=Design, replications=5,
generate=Generate, analyse=Analyse, summarise=Summarise)
res
## Not run:
# complete run with 1000 replications per condition
res <- runSimulation(design=Design, replications=1000, parallel=TRUE,
generate=Generate, analyse=Analyse, summarise=Summarise)
res
View(res)
## save final results to a file upon completion, and play a beep when done
runSimulation(design=Design, replications=1000, parallel=TRUE, filename = 'mysim',
generate=Generate, analyse=Analyse, summarise=Summarise, beep=TRUE)
## same as above, but send a notification via Pushbullet upon completion
library(RPushbullet) # read-in default JSON file
runSimulation(design=Design, replications=1000, parallel=TRUE, filename = 'mysim',
generate=Generate, analyse=Analyse, summarise=Summarise,
notification = 'complete')
## Submit as RStudio job (requires job package and active RStudio session)
job::job({
res <- runSimulation(design=Design, replications=100,
generate=Generate, analyse=Analyse, summarise=Summarise)
}, title='t-test simulation')
res # object res returned to console when completed
## Debug the generate function. See ?browser for help on debugging
## Type help to see available commands (e.g., n, c, where, ...),
## ls() to see what has been defined, and type Q to quit the debugger
runSimulation(design=Design, replications=1000,
generate=Generate, analyse=Analyse, summarise=Summarise,
parallel=TRUE, debug='generate')
## Alternatively, place a browser() within the desired function line to
## jump to a specific location
Summarise <- function(condition, results, fixed_objects = NULL) {
#find results of interest here (e.g., alpha < .1, .05, .01)
browser()
ret <- EDR(results[,nms], alpha = .05)
ret
}
## The following debugs the analyse function for the
## second row of the Design input
runSimulation(design=Design, replications=1000,
generate=Generate, analyse=Analyse, summarise=Summarise,
parallel=TRUE, debug='analyse-2')
####################################
## EXTRA: To run the simulation on a MPI cluster, use the following setup (not run)
library(doMPI)
cl <- startMPIcluster()
registerDoMPI(cl)
Final <- runSimulation(design=Design, replications=1000, MPI=TRUE,
generate=Generate, analyse=Analyse, summarise=Summarise)
saveRDS(Final, 'mysim.rds')
closeCluster(cl)
mpi.quit()
## Similarly, run simulation on a network linked via ssh
## (two way ssh key-paired connection must be possible between master and slave nodes)
##
## define IP addresses, including primary IP
primary <- '192.168.2.20'
IPs <- list(
list(host=primary, user='phil', ncore=8),
list(host='192.168.2.17', user='phil', ncore=8)
)
spec <- lapply(IPs, function(IP)
rep(list(list(host=IP$host, user=IP$user)), IP$ncore))
spec <- unlist(spec, recursive=FALSE)
cl <- parallel::makeCluster(type='PSOCK', master=primary, spec=spec)
res <- runSimulation(design=Design, replications=1000, parallel = TRUE,
generate=Generate, analyse=Analyse, summarise=Summarise, cl=cl)
## Using parallel='future' to allow the future framework to be used instead
library(future) # future structure to be used internally
plan(multisession) # specify different plan (default is sequential)
res <- runSimulation(design=Design, replications=100, parallel='future',
generate=Generate, analyse=Analyse, summarise=Summarise)
head(res)
# The progressr package is used for progress reporting with futures. To redefine
# use progressr::handlers() (see below)
library(progressr)
with_progress(res <- runSimulation(design=Design, replications=100, parallel='future',
generate=Generate, analyse=Analyse, summarise=Summarise))
head(res)
# re-define progressr's bar (below requires cli)
handlers(handler_pbcol(
adjust = 1.0,
complete = function(s) cli::bg_red(cli::col_black(s)),
incomplete = function(s) cli::bg_cyan(cli::col_black(s))
))
with_progress(res <- runSimulation(design=Design, replications=100, parallel='future',
generate=Generate, analyse=Analyse, summarise=Summarise))
# reset future computing plan when complete (good practice)
plan(sequential)
####################################
###### Post-analysis: Analyze the results via functions like lm() or SimAnova(), and create
###### tables(dplyr) or plots (ggplot2) to help visualize the results.
###### This is where you get to be a data analyst!
library(dplyr)
res %>% summarise(mean(welch), mean(independent))
res %>% group_by(standard_deviation_ratio, group_size_ratio) %>%
summarise(mean(welch), mean(independent))
# quick ANOVA analysis method with all two-way interactions
SimAnova( ~ (sample_size + group_size_ratio + standard_deviation_ratio)^2, res,
rates = TRUE)
# or more specific ANOVAs
SimAnova(independent ~ (group_size_ratio + standard_deviation_ratio)^2,
res, rates = TRUE)
# make some plots
library(ggplot2)
library(tidyr)
dd <- res %>%
select(group_size_ratio, standard_deviation_ratio, welch, independent) %>%
pivot_longer(cols=c('welch', 'independent'), names_to = 'stats')
dd
ggplot(dd, aes(factor(group_size_ratio), value)) + geom_boxplot() +
geom_abline(intercept=0.05, slope=0, col = 'red') +
geom_abline(intercept=0.075, slope=0, col = 'red', linetype='dotted') +
geom_abline(intercept=0.025, slope=0, col = 'red', linetype='dotted') +
facet_wrap(~stats)
ggplot(dd, aes(factor(group_size_ratio), value, fill = factor(standard_deviation_ratio))) +
geom_boxplot() + geom_abline(intercept=0.05, slope=0, col = 'red') +
geom_abline(intercept=0.075, slope=0, col = 'red', linetype='dotted') +
geom_abline(intercept=0.025, slope=0, col = 'red', linetype='dotted') +
facet_grid(stats~standard_deviation_ratio) +
theme(legend.position = 'none')
## End(Not run)
philchalmers/SimDesign documentation built on April 14, 2024, 6:38 p.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
View source: R/runSimulation.R
| runSimulation | R Documentation |
Run a Monte Carlo simulation given conditions and simulation functions
Description
This function runs a Monte Carlo simulation study given a set of predefined simulation functions,
design conditions, and number of replications. Results can be saved as temporary files in case of
interruptions and may be restored by re-running runSimulation, provided that the respective temp
file can be found in the working directory. runSimulation supports parallel
and cluster computing (with the parallel and future packages; see also
runArraySimulation for submitting array jobs to HPC clusters),
global and local debugging, error handling (including fail-safe
stopping when functions fail too often, even across nodes), provides bootstrap estimates of the
sampling variability (optional), and automatic tracking of error and warning messages
with their associated .Random.seed states.
For convenience, all functions available in the R work-space are exported across all nodes
so that they are more easily accessible (however, other R objects are not, and therefore
must be passed to the fixed_objects input to become available across nodes).
Usage
runSimulation(
design,
replications,
generate,
analyse,
summarise,
fixed_objects = NULL,
packages = NULL,
filename = NULL,
debug = "none",
load_seed = NULL,
save = any(replications > 10),
store_results = TRUE,
save_results = FALSE,
parallel = FALSE,
ncores = parallel::detectCores() - 1L,
cl = NULL,
notification = "none",
beep = FALSE,
sound = 1,
CI = 0.95,
seed = NULL,
boot_method = "none",
boot_draws = 1000L,
max_errors = 50L,
resume = TRUE,
save_details = list(),
control = list(),
progress = TRUE,
verbose = TRUE
)
## S3 method for class 'SimDesign'
summary(object, ...)
## S3 method for class 'SimDesign'
print(x, list2char = TRUE, ...)
Arguments
design |
a |
replications |
number of independent replications to perform per
condition (i.e., each row in |
generate |
user-defined data and parameter generating function (or named list of functions).
See |
analyse |
user-defined analysis function (or named list of functions)
that acts on the data generated from
|
summarise |
optional (but strongly recommended) user-defined summary function
from Omitting this function will return a tibble with the |
fixed_objects |
(optional) an object (usually a named |
packages |
a character vector of external packages to be used during the simulation (e.g.,
|
filename |
(optional) the name of the Note that if the same file name already exists in the working
directly at the time of saving then a new
file will be generated instead and a warning will be thrown. This helps to
avoid accidentally overwriting
existing files. Default is |
debug |
a string indicating where to initiate a If the For debugging specific rows in the Finally, users may place |
load_seed |
used to replicate an exact simulation state, which is
primarily useful for debugging purposes.
Input can be a character object indicating which file to load from when the
|
save |
logical; save the temporary simulation state to the hard-drive? This is useful
for simulations which require an extended amount of time, though for shorter simulations
can be disabled to slightly improve computational efficiency. When To recover your simulation at the last known location (having patched the issues in the
previous execution code) simply re-run the code you used to
initially define the simulation and the external file will automatically be detected and read-in.
Default is |
store_results |
logical; store the complete tables of simulation results
in the returned object? This is To extract these results
pass the returned object to either |
save_results |
logical; save the results returned from WARNING: saving results to your hard-drive can fill up space very quickly for
larger simulations. Be sure to
test this option using a smaller number of replications before the full Monte
Carlo simulation is performed.
See also |
parallel |
logical; use parallel processing from the Alternatively, if the |
ncores |
number of cores to be used in parallel execution (ignored if using the
|
cl |
cluster object defined by Note that if the |
notification |
an optional character vector input that can be used to send
Pushbullet notifications from a configured
computer. This reports information such as the total execution time, the condition
completed, and error/warning
messages recorded. This arguments assumes that users have already A) registered for
a Pushbullet account,
B) installed the application on their mobile device and computer, and C) created an
associated JSON file of the form
To utilize the |
beep |
logical; call the |
sound |
|
CI |
bootstrap confidence interval level (default is 95%) |
seed |
a vector or list of integers to be used for reproducibility.
The length of the vector must be equal the number of rows in |
boot_method |
method for performing non-parametric bootstrap confidence intervals
for the respective meta-statistics computed by the |
boot_draws |
number of non-parametric bootstrap draws to sample for the |
max_errors |
the simulation will terminate when more than this number of consecutive
errors are thrown in any
given condition, causing the simulation to continue to the next unique |
resume |
logical; if a temporary |
save_details |
a list pertaining to information regarding how and where files should be saved
when the
|
control |
a list for extra information flags for controlling less commonly used features. These include
|
progress |
logical; display a progress bar (using the |
verbose |
logical; print messages to the R console? Default is |
object |
SimDesign object returned from |
... |
additional arguments |
x |
SimDesign object returned from |
list2char |
logical; for |
Details
For an in-depth tutorial of the package please refer to Chalmers and Adkins (2020;
\Sexpr[results=rd]{tools:::Rd_expr_doi("10.20982/tqmp.16.4.p248")}).
For an earlier didactic presentation of the package refer to Sigal and Chalmers
(2016; \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1080/10691898.2016.1246953")}). Finally, see the associated
wiki on Github (https://github.com/philchalmers/SimDesign/wiki)
for tutorial material, examples, and applications of SimDesign to real-world
simulation experiments, as well as the various vignette files associated with the package.
The strategy for organizing the Monte Carlo simulation work-flow is to
- 1)
Define a suitable
Designobject (atibbleordata.frame) containing fixed conditional information about the Monte Carlo simulations. Each row or thisdesignobject pertains to a unique set of simulation to study, while each column the simulation factor under investigation (e.g., sample size, distribution types, etc). This is often expedited by using thecreateDesignfunction, and if necessary the argumentsubsetcan be used to remove redundant or non-applicable rows- 2)
Define the three step functions to generate the data (
Generate; see also https://CRAN.R-project.org/view=Distributions for a list of distributions in R), analyse the generated data by computing the respective parameter estimates, detection rates, etc (Analyse), and finally summarise the results across the total number of replications (Summarise).- 3)
Pass the
designobject and three defined R functions torunSimulation, and declare the number of replications to perform with thereplicationsinput. This function will return a suitabletibbleobject with the complete simulation results and execution details- 4)
Analyze the output from
runSimulation, possibly using ANOVA techniques (SimAnova) and generating suitable plots and tables
Expressing the above more succinctly, the functions to be called have the following form, with the exact functional arguments listed:
Design <- createDesign(...)Generate <- function(condition, fixed_objects = NULL) {...}Analyse <- function(condition, dat, fixed_objects = NULL) {...}Summarise <- function(condition, results, fixed_objects = NULL) {...}res <- runSimulation(design=Design, replications, generate=Generate, analyse=Analyse, summarise=Summarise)
The condition object above represents a single row from the design object, indicating
a unique Monte Carlo simulation condition. The condition object also contains two
additional elements to help track the simulation's state: an ID variable, indicating
the respective row number in the design object, and a REPLICATION element
indicating the replication iteration number (an integer value between 1 and replication).
This setup allows users to easily locate the rth replication (e.g., REPLICATION == 500)
within the jth row in the simulation design (e.g., ID == 2). The
REPLICATION input is also useful when temporarily saving files to the hard-drive
when calling external command line utilities (see examples on the wiki).
For a template-based version of the work-flow, which is often useful when initially
defining a simulation, use the SimFunctions function. This
function will write a template simulation
to one/two files so that modifying the required functions and objects can begin immediately.
This means that users can focus on their Monte Carlo simulation details right away rather
than worrying about the repetitive administrative code-work required to organize the simulation's
execution flow.
Finally, examples, presentation files, and tutorials can be found on the package wiki located at https://github.com/philchalmers/SimDesign/wiki.
Value
a tibble from the dplyr package (also of class 'SimDesign')
with the original design conditions in the left-most columns,
simulation results in the middle columns, and additional information in the right-most columns (see below).
The right-most column information for each condition are:
REPLICATIONS to indicate the number of Monte Carlo replications,
SIM_TIME to indicate how long (in seconds) it took to complete
all the Monte Carlo replications for each respective design condition,
RAM_USED amount of RAM that was in use at the time of completing
each simulation condition,
COMPLETED to indicate the date in which the given simulation condition completed,
SEED for the integer values in the seed argument (if applicable; not
relevant when L'Ecuyer-CMRG method used), and, if applicable,
ERRORS and WARNINGS which contain counts for the number of error or warning
messages that were caught (if no errors/warnings were observed these columns will be omitted).
Note that to extract the specific error and warnings messages see
SimExtract. Finally,
if boot_method was a valid input other than 'none' then the final right-most
columns will contain the labels
BOOT_ followed by the name of the associated meta-statistic defined in summarise() and
and bootstrapped confidence interval location for the meta-statistics.
Saving data, results, seeds, and the simulation state
To conserve RAM, temporary objects (such as data generated across conditions and replications)
are discarded; however, these can be saved to the hard-disk by passing the appropriate flags.
For longer simulations it is recommended to use the save_results flag to write the
analysis results to the hard-drive.
The use of the store_seeds or the save_seeds options
can be evoked to save R's .Random.seed
state to allow for complete reproducibility of each replication within each condition. These
individual .Random.seed terms can then be read in with the
load_seed input to reproduce the exact simulation state at any given replication.
Most often though, saving the complete list of seeds is unnecessary as problematic seeds are
automatically stored in the final simulation object to allow for easier replicability
of potentially problematic errors (which incidentally can be extracted
using SimExtract(res, 'error_seeds') and passed to the load_seed argument). Finally,
providing a vector of seeds is also possible to ensure
that each simulation condition is macro reproducible under the single/multi-core method selected.
Finally, when the Monte Carlo simulation is complete
it is recommended to write the results to a hard-drive for safe keeping, particularly with the
filename argument provided (for reasons that are more obvious in the parallel computation
descriptions below). Using the filename argument supplied is safer than using, for instance,
saveRDS directly because files will never accidentally be overwritten,
and instead a new file name will be created when a conflict arises; this type of implementation safety
is prevalent in many locations in the package to help avoid unrecoverable (yet surprisingly
common) mistakes during the process of designing and executing Monte Carlo simulations.
Resuming temporary results
In the event of a computer crash, power outage, etc, if save = TRUE was used (the default)
then the original code used to execute runSimulation() need only be re-run to resume
the simulation. The saved temp file will be read into the function automatically, and the
simulation will continue one the condition where it left off before the simulation
state was terminated. If users wish to remove this temporary
simulation state entirely so as to start anew then simply pass SimClean(temp = TRUE)
in the R console to remove any previously saved temporary objects.
A note on parallel computing
When running simulations in parallel (either with parallel = TRUE or MPI = TRUE,
or when using the future approach with a plan() other than sequential)
R objects defined in the global environment will generally not be visible across nodes.
Hence, you may see errors such as Error: object 'something' not found if you try to use
an object that is defined in the work space but is not passed to runSimulation.
To avoid this type or error, simply pass additional objects to the
fixed_objects input (usually it's convenient to supply a named list of these objects).
Fortunately, however, custom functions defined in the global environment are exported across
nodes automatically. This makes it convenient when writing code because custom functions will
always be available across nodes if they are visible in the R work space. As well, note the
packages input to declare packages which must be loaded via library() in order to make
specific non-standard R functions available across nodes.
Author(s)
Phil Chalmers rphilip.chalmers@gmail.com
References
Chalmers, R. P., & Adkins, M. C. (2020). Writing Effective and Reliable Monte Carlo Simulations
with the SimDesign Package. The Quantitative Methods for Psychology, 16(4), 248-280.
\Sexpr[results=rd]{tools:::Rd_expr_doi("10.20982/tqmp.16.4.p248")}
Sigal, M. J., & Chalmers, R. P. (2016). Play it again: Teaching statistics with Monte
Carlo simulation. Journal of Statistics Education, 24(3), 136-156.
\Sexpr[results=rd]{tools:::Rd_expr_doi("10.1080/10691898.2016.1246953")}
See Also
SimFunctions, createDesign,
Generate, Analyse, Summarise,
SimExtract,
reSummarise, SimClean, SimAnova, SimResults,
aggregate_simulations, Attach, AnalyseIf,
SimShiny, convertWarnings, runArraySimulation
Examples
#-------------------------------------------------------------------------------
# Example 1: Sampling distribution of mean
# This example demonstrate some of the simpler uses of SimDesign,
# particularly for classroom settings. The only factor varied in this simulation
# is sample size.
# skeleton functions to be saved and edited
SimFunctions()
#### Step 1 --- Define your conditions under study and create design data.frame
Design <- createDesign(N = c(10, 20, 30))
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Step 2 --- Define generate, analyse, and summarise functions
# help(Generate)
Generate <- function(condition, fixed_objects = NULL) {
dat <- with(condition, rnorm(N, 10, 5)) # distributed N(10, 5)
dat
}
# help(Analyse)
Analyse <- function(condition, dat, fixed_objects = NULL) {
ret <- c(mean=mean(dat)) # mean of the sample data vector
ret
}
# help(Summarise)
Summarise <- function(condition, results, fixed_objects = NULL) {
# mean and SD summary of the sample means
ret <- c(mu=mean(results$mean), SE=sd(results$mean))
ret
}
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Step 3 --- Collect results by looping over the rows in design
# run the simulation
Final <- runSimulation(design=Design, replications=10,
generate=Generate, analyse=Analyse, summarise=Summarise)
Final
SimResults(Final)
## Not run:
# reproduce exact simulation
Final_rep <- runSimulation(design=Design, replications=10, seed=Final$SEED,
generate=Generate, analyse=Analyse, summarise=Summarise)
Final_rep
# run with more standard number of replications (note the storage message)
Final <- runSimulation(design=Design, replications=1000,
generate=Generate, analyse=Analyse, summarise=Summarise)
Final
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Extras
# compare SEs estimates to the true SEs from the formula sigma/sqrt(N)
5 / sqrt(Design$N)
# To store the results from the analyse function either
# a) omit a definition of summarise() to return all results,
# b) use store_results = TRUE (default) to store results internally and later
# extract with SimResults(), or
# c) pass save_results = TRUE to runSimulation() and read the results in with SimResults()
#
# Note that method c) should be adopted for larger simulations, particularly
# if RAM storage could be an issue and error/warning message information is important.
# a) approach
res <- runSimulation(design=Design, replications=5,
generate=Generate, analyse=Analyse)
res
# b) approach (store_results = TRUE by default)
res <- runSimulation(design=Design, replications=5,
generate=Generate, analyse=Analyse, summarise=Summarise)
res
SimResults(res)
# c) approach
Final <- runSimulation(design=Design, replications=5, save_results=TRUE,
generate=Generate, analyse=Analyse, summarise=Summarise)
# read-in all conditions (can be memory heavy)
res <- SimResults(Final)
res
head(res[[1]]$results)
# just first condition
res <- SimResults(Final, which=1)
head(res$results)
dplyr::tibble(res$condition, res$results)
# obtain empirical bootstrapped CIs during an initial run
# the simulation was completed (necessarily requires save_results = TRUE)
res <- runSimulation(design=Design, replications=1000, boot_method = 'basic',
generate=Generate, analyse=Analyse, summarise=Summarise)
res
# alternative bootstrapped CIs that uses saved results via reSummarise().
# Default directory save to:
dirname <- paste0('SimDesign-results_', unname(Sys.info()['nodename']), "/")
res <- reSummarise(summarise=Summarise, dir=dirname, boot_method = 'basic')
res
# remove the saved results from the hard-drive if you no longer want them
SimClean(results = TRUE)
## End(Not run)
#-------------------------------------------------------------------------------
# Example 2: t-test and Welch test when varying sample size, group sizes, and SDs
# skeleton functions to be saved and edited
SimFunctions()
## Not run:
# in real-world simulations it's often better/easier to save
# these functions directly to your hard-drive with
SimFunctions('my-simulation')
## End(Not run)
#### Step 1 --- Define your conditions under study and create design data.frame
Design <- createDesign(sample_size = c(30, 60, 90, 120),
group_size_ratio = c(1, 4, 8),
standard_deviation_ratio = c(.5, 1, 2))
Design
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Step 2 --- Define generate, analyse, and summarise functions
Generate <- function(condition, fixed_objects = NULL) {
N <- condition$sample_size # could use Attach() to make objects available
grs <- condition$group_size_ratio
sd <- condition$standard_deviation_ratio
if(grs < 1){
N2 <- N / (1/grs + 1)
N1 <- N - N2
} else {
N1 <- N / (grs + 1)
N2 <- N - N1
}
group1 <- rnorm(N1)
group2 <- rnorm(N2, sd=sd)
dat <- data.frame(group = c(rep('g1', N1), rep('g2', N2)), DV = c(group1, group2))
dat
}
Analyse <- function(condition, dat, fixed_objects = NULL) {
welch <- t.test(DV ~ group, dat)$p.value
independent <- t.test(DV ~ group, dat, var.equal=TRUE)$p.value
# In this function the p values for the t-tests are returned,
# and make sure to name each element, for future reference
ret <- nc(welch, independent)
ret
}
Summarise <- function(condition, results, fixed_objects = NULL) {
#find results of interest here (e.g., alpha < .1, .05, .01)
ret <- EDR(results, alpha = .05)
ret
}
#~~~~~~~~~~~~~~~~~~~~~~~~
#### Step 3 --- Collect results by looping over the rows in design
# first, test to see if it works
res <- runSimulation(design=Design, replications=5,
generate=Generate, analyse=Analyse, summarise=Summarise)
res
## Not run:
# complete run with 1000 replications per condition
res <- runSimulation(design=Design, replications=1000, parallel=TRUE,
generate=Generate, analyse=Analyse, summarise=Summarise)
res
View(res)
## save final results to a file upon completion, and play a beep when done
runSimulation(design=Design, replications=1000, parallel=TRUE, filename = 'mysim',
generate=Generate, analyse=Analyse, summarise=Summarise, beep=TRUE)
## same as above, but send a notification via Pushbullet upon completion
library(RPushbullet) # read-in default JSON file
runSimulation(design=Design, replications=1000, parallel=TRUE, filename = 'mysim',
generate=Generate, analyse=Analyse, summarise=Summarise,
notification = 'complete')
## Submit as RStudio job (requires job package and active RStudio session)
job::job({
res <- runSimulation(design=Design, replications=100,
generate=Generate, analyse=Analyse, summarise=Summarise)
}, title='t-test simulation')
res # object res returned to console when completed
## Debug the generate function. See ?browser for help on debugging
## Type help to see available commands (e.g., n, c, where, ...),
## ls() to see what has been defined, and type Q to quit the debugger
runSimulation(design=Design, replications=1000,
generate=Generate, analyse=Analyse, summarise=Summarise,
parallel=TRUE, debug='generate')
## Alternatively, place a browser() within the desired function line to
## jump to a specific location
Summarise <- function(condition, results, fixed_objects = NULL) {
#find results of interest here (e.g., alpha < .1, .05, .01)
browser()
ret <- EDR(results[,nms], alpha = .05)
ret
}
## The following debugs the analyse function for the
## second row of the Design input
runSimulation(design=Design, replications=1000,
generate=Generate, analyse=Analyse, summarise=Summarise,
parallel=TRUE, debug='analyse-2')
####################################
## EXTRA: To run the simulation on a MPI cluster, use the following setup (not run)
library(doMPI)
cl <- startMPIcluster()
registerDoMPI(cl)
Final <- runSimulation(design=Design, replications=1000, MPI=TRUE,
generate=Generate, analyse=Analyse, summarise=Summarise)
saveRDS(Final, 'mysim.rds')
closeCluster(cl)
mpi.quit()
## Similarly, run simulation on a network linked via ssh
## (two way ssh key-paired connection must be possible between master and slave nodes)
##
## define IP addresses, including primary IP
primary <- '192.168.2.20'
IPs <- list(
list(host=primary, user='phil', ncore=8),
list(host='192.168.2.17', user='phil', ncore=8)
)
spec <- lapply(IPs, function(IP)
rep(list(list(host=IP$host, user=IP$user)), IP$ncore))
spec <- unlist(spec, recursive=FALSE)
cl <- parallel::makeCluster(type='PSOCK', master=primary, spec=spec)
res <- runSimulation(design=Design, replications=1000, parallel = TRUE,
generate=Generate, analyse=Analyse, summarise=Summarise, cl=cl)
## Using parallel='future' to allow the future framework to be used instead
library(future) # future structure to be used internally
plan(multisession) # specify different plan (default is sequential)
res <- runSimulation(design=Design, replications=100, parallel='future',
generate=Generate, analyse=Analyse, summarise=Summarise)
head(res)
# The progressr package is used for progress reporting with futures. To redefine
# use progressr::handlers() (see below)
library(progressr)
with_progress(res <- runSimulation(design=Design, replications=100, parallel='future',
generate=Generate, analyse=Analyse, summarise=Summarise))
head(res)
# re-define progressr's bar (below requires cli)
handlers(handler_pbcol(
adjust = 1.0,
complete = function(s) cli::bg_red(cli::col_black(s)),
incomplete = function(s) cli::bg_cyan(cli::col_black(s))
))
with_progress(res <- runSimulation(design=Design, replications=100, parallel='future',
generate=Generate, analyse=Analyse, summarise=Summarise))
# reset future computing plan when complete (good practice)
plan(sequential)
####################################
###### Post-analysis: Analyze the results via functions like lm() or SimAnova(), and create
###### tables(dplyr) or plots (ggplot2) to help visualize the results.
###### This is where you get to be a data analyst!
library(dplyr)
res %>% summarise(mean(welch), mean(independent))
res %>% group_by(standard_deviation_ratio, group_size_ratio) %>%
summarise(mean(welch), mean(independent))
# quick ANOVA analysis method with all two-way interactions
SimAnova( ~ (sample_size + group_size_ratio + standard_deviation_ratio)^2, res,
rates = TRUE)
# or more specific ANOVAs
SimAnova(independent ~ (group_size_ratio + standard_deviation_ratio)^2,
res, rates = TRUE)
# make some plots
library(ggplot2)
library(tidyr)
dd <- res %>%
select(group_size_ratio, standard_deviation_ratio, welch, independent) %>%
pivot_longer(cols=c('welch', 'independent'), names_to = 'stats')
dd
ggplot(dd, aes(factor(group_size_ratio), value)) + geom_boxplot() +
geom_abline(intercept=0.05, slope=0, col = 'red') +
geom_abline(intercept=0.075, slope=0, col = 'red', linetype='dotted') +
geom_abline(intercept=0.025, slope=0, col = 'red', linetype='dotted') +
facet_wrap(~stats)
ggplot(dd, aes(factor(group_size_ratio), value, fill = factor(standard_deviation_ratio))) +
geom_boxplot() + geom_abline(intercept=0.05, slope=0, col = 'red') +
geom_abline(intercept=0.075, slope=0, col = 'red', linetype='dotted') +
geom_abline(intercept=0.025, slope=0, col = 'red', linetype='dotted') +
facet_grid(stats~standard_deviation_ratio) +
theme(legend.position = 'none')
## End(Not run)
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.