fastrerandomize: An R Package for Ultra-Fast Rerandomization Using Accelerated Computing

What is fastrerandomize? | Installation | Tutorial | Data | CRAN | References | Documentation

Note: fastrerandomize has been successfully tested on CPU, CUDA, and METAL frameworks. Special thanks to Aniket Kamat and Fucheng Warren Zhu for their work on the latest package build!

What is fastrerandomize?

The fastrerandomize contains functions such as randomization_test, which offers a streamlined approach for performing randomization tests after using rerandomization in the research design.

We employ a JAX backend to make exact rerandomization inference possible even for larger experiments where the number of randomizations is in the hundreds of millions or where experimenters seek to maintain balanced randomizations across thousands of features.

Package Installation and Loading

```

Install fastrerandomize stable version from CRAN

install.packages("fastrerandomize")

Install fastrerandomize development version from GitHub

devtools::install_github(repo = "cjerzak/fastrerandomize-software/fastrerandomize")

Load the package

library( fastrerandomize )

Running code the first time, you'll want to create the computational environment

fastrerandomize::build_backend() ```

Tutorial

Let's get started with a tutorial. We're first going to use the package to generate a pool of acceptable rerandomizations. ```

First, specify some analysis parameters

nunits <- 20; ntreated <- 10

Generate covariate data

X <- matrix(rnorm(nunits*5),nrow = nunits)

Generate a set of acceptable randomizations based on randomizationacceptprob.

When randomizationacceptprob = 1, all randomizations are accepted.

When randomizationacceptprob < 1, only well-balanced randomizations are accepted.

When randomizationacceptprob = 1/|Size of cand. randomization set|, 1 randomization is accepted.

candidaterandomizations <- fastrerandomize::generaterandomizations( nunits = nunits, ntreated = ntreated, X = X, randomizationacceptprob = 0.001)

Check out the candidate randomization dimensions

dim( candidate_randomizations ) We can also use `fastrerandomize` to perform a randomization test using those acceptable randomizations.

Setup simulated outcome data

CoefY <- rnorm(ncol(X)) Wobs <- candidaterandomizations[1,] tautrue <- 1 Yobs <- c(X %% as.matrix(CoefY) + Wobstautrue + rnorm(nunits, sd = 0.1))

Perform exact randomization set based on accepted randomizations

ExactRandomizationTestResults <- fastrerandomize::randomizationtest( obsW = Wobs, obsY = Yobs, candidaterandomizations = candidaterandomizations, findFI = F # set to T if an exact fiducial interval needed ) ExactRandomizationTestResults$pvalue # p-value ExactRandomizationTestResults$tau_obs # difference-in-means ATE estimate ```

Replication Data

Replication data for the package is available using the data command. data( QJEData ) data( YOPData )

Development

We welcome new features or bug fixes (you can raise an issue or submit a pull request in the repository). We will keep the package up-to-date with the latest version of the JAX backend.

References

Connor T. Jerzak, Rebecca Goldstein, Aniket Kamat, Fucheng Warren Zhu. fastrerandomize: An R Package for Fast Rerandomization Using Accelerated Computing. ArXiv Preprint, 2025. [PDF] @article{jerzak2025fastrerandomize, title={fastrerandomize: An R Package for Fast Rerandomization Using Accelerated Computing}, author={Jerzak, Connor T. and Rebecca Goldstein and Aniket Kamat and Fucheng Warren Zhu}, journal={ArXiv Preprint}, year={2025} }

Connor T. Jerzak and Rebecca Goldstein. "Degrees of Randomness in Rerandomization Procedures." ArXiv Preprint, 2023. [PDF] @article{JerGol2023, title={Degrees of Randomness in Rerandomization Procedures}, author={Jerzak, Connor T. and Rebecca Goldstein}, journal={ArXiv Preprint}, year={2023}}

_

Package functions

_

Hardware acceleration

_

Key to minimize memory overhead

_