Flexible 'Tidyverse'-Friendly Simulations.

simpr

simpr provides a general, simple, and tidyverse-friendly framework for generating simulated data, fitting models on simulations, and tidying model results. The full workflow can happen in a single tidy pipeline without creating external functions, global values, or using loops. It’s useful for power analysis, design analysis, simulation studies, and for teaching statistics.

• Easily readable simulation specifications. You can specify simulations in a few lines, including referring to other simulation variables and to simulation parameters that you’re varying (such as sample size). simpr takes care of all the details of generating your simulation across varying parameters.

• Sensibly handle errors.simpr has various options to keep going even when simulation or model-fitting hits errors, so that you don’t need to start over if a single iteration hits fatal numerical issues.

• Reproducible workflows. Individual simulations can be reproduced exactly without needed to perform the whole simulation again.

• Easy-to-use parallel processing. Building on furrr, parallel processing for simpr can usually be turned on with a couple lines of code.

The hardest part of any simulation is designing the data-generating process and deciding what values of parameters you want to explore. simpr takes care of the rest so you can focus on these central issues.

Installation and loading

## Install stable CRAN version
install.packages("simpr")

## Install latest development version
remotes::install_github("statisfactions/simpr")

library(simpr)

Example simulation

The simpr workflow, inspired by the infer package, distills a simulation study into five primary steps:

1. specify() your data-generating process

2. define() parameters that you want to systematically vary across your simulation design (e.g. n, effect size)

3. generate() the simulation data

4. fit() models to your data (e.g. lm())

5. tidy_fits() for further processing using broom::tidy(), such as computing power or Type I Error rates

simpr makes no assumptions about your data and is not specialized to any particular type of data generating process or model. If R can generate it and if R can fit models, you can use simpr to run your simulation. (The tidying step is limited by the models supported broom::tidy(), although you can also supply your own tidying function or expression.)

Suppose we are calculating the power for a two-sample t-test where the data is log-normally distributed, which can be generated by stats::rlnorm().

set.seed(100)

## Data-generating mechanism
specify(a = ~ rlnorm(n, mean = 0),
        b = ~ rlnorm(n, mean = 0.5)) %>% 
  ## Vary n from 30 to 100
  define(n = seq(30, 100, by = 10)) %>% 
  ## 100 repetitions
  generate(100) %>% 
  ## fit t-tests
  fit(t_test = ~ t.test(a, b)) %>%
  ## bring model results into a tidy tibble
  tidy_fits()
#> # A tibble: 800 × 14
#>    .sim_id     n   rep Source estimate estimate1 estimate2 statistic p.value
#>      <int> <dbl> <int> <chr>     <dbl>     <dbl>     <dbl>     <dbl>   <dbl>
#>  1       1    30     1 t_test   -0.953      1.73      2.68    -1.60  0.117  
#>  2       2    40     1 t_test   -0.249      1.64      1.89    -0.581 0.563  
#>  3       3    50     1 t_test   -0.616      1.67      2.29    -1.19  0.237  
#>  4       4    60     1 t_test   -1.75       1.28      3.03    -3.30  0.00146
#>  5       5    70     1 t_test   -0.876      1.61      2.48    -1.96  0.0525 
#>  6       6    80     1 t_test   -0.780      1.71      2.49    -2.13  0.0352 
#>  7       7    90     1 t_test   -0.818      1.60      2.42    -2.51  0.0129 
#>  8       8   100     1 t_test   -0.878      1.51      2.38    -2.61  0.00988
#>  9       9    30     2 t_test   -0.487      1.96      2.44    -0.713 0.479  
#> 10      10    40     2 t_test   -2.29       1.37      3.66    -1.76  0.0851 
#> # … with 790 more rows, and 5 more variables: parameter <dbl>, conf.low <dbl>,
#> #   conf.high <dbl>, method <chr>, alternative <chr>

specify() creates two variables a and b that are distributed lognormally (any R expression that generates data can be used here). The specify expressions refer to the sample size, n. define() clarifies that n varies between 30 and 100 by 10s. generate() actually does the data generation, with 100 simulated datasets for each possible value of define(). fit() applies an arbitrary R expression to each simulated dataset, and tidy_fits() brings things together in a tidy tibble that can be easily aggregated and plotted to calculate bias, power, etc.

Further resources

See vignette("simpr") to get started on using the package, or view the simprshowcase for several applied examples.