Speaker: Connor Jerzak
Paper: doi.org/10.1016/j.softx.2026.102508 // [.bib]
Transcript: Welcome to this presentation on fast re-randomization using accelerated computing, focusing on scaling up experimental design for high-dimensional data.

Randomized controlled trials (RCTs) are widely considered the gold standard for causal inference because they guarantee unbiased estimates of treatment effects on average. For interventions like a pharmaceutical drug or a cash transfer program, the average difference in outcome between the treatment and control groups naturally equals the overall treatment effect.

However, in finite samples, simple randomization frequently produces treatment and control groups with severe chance imbalances. For instance, in an extreme two-unit case where one participant is female and the other is male, a coin flip determining treatment and control makes it impossible to distinguish whether the resulting effect is due to the intervention itself or the background characteristic of gender.

This imbalance issue is significantly exacerbated when working with modern, high-dimensional pre-treatment covariates, such as text embeddings or satellite image descriptors. As the dimensionality of these covariates increases, there is a higher likelihood of chance imbalances that can mask true treatment effects, making conclusions less reliable if those covariates strongly explain the outcome.

To address this, researchers use re-randomization, a process of repeatedly drawing candidate assignments, evaluating their covariate imbalance, and only accepting those that pass a stringent similarity threshold. By shrinking the expected imbalance, re-randomization systematically reduces the overall expected mean squared error (MSE). The MSE in this design is governed by a sample size scaling factor, irreducible sampling noise, and prognostic variance. Re-randomization specifically shrinks this prognostic variance by forcing the treatment and control groups to be highly similar, yielding valid design-based guarantees for improved precision.

Despite its benefits, traditional re-randomization faces severe computational bottlenecks because, as the number of covariates and experimental units scales up, the space of possible randomizations grows exponentially. Achieving meaningful precision gains in high-dimensional settings requires enforcing increasingly stringent acceptance probabilities, meaning standard CPU-bound iterative approaches are far too slow to generate the massive number of candidate draws required.

To solve this, our research leverages hardware acceleration via GPUs and TPUs to evaluate balance criteria across massive pools of candidate assignment vectors in parallel. Furthermore, to prevent the catastrophic memory exhaustion caused by retaining millions of acceptable assignment vectors for large sample sizes, we separate evaluation from retention. Instead of saving the full randomized vectors, we discard them and only retain the compact pseudo-random number generator keys used to create them. Because the generation process is deterministic given a key, we can efficiently regenerate the accepted randomizations on demand. This keys-only storage approach reduces the memory footprint drastically, allowing for theoretically infinite scaling as sample sizes grow.

By resolving both the computational and memory bottlenecks, this framework allows researchers to perform rigorous, design-respecting randomization inference conditional on acceptance. Researchers can compute exact p-values and fiducial intervals without relying on complex sequentially adjusted protocols.

Ultimately, this innovation ushers in an era of representation-rich, high-dimensional experiments, enabling the seamless integration of covariates from deep neural networks, visual embeddings, and text representations. By unifying batch computation and hardware acceleration with a memory-efficient storage design, we can extend the gold standard of experimental design to the massive scale of modern high-dimensional data without compromising on statistical rigor.

–

References

@article{jerzak2026fastrerandomize,
  title={FastRerandomize: Fast Rerandomization Using Accelerated Computing},
  author={Jerzak, Connor T. and Rebecca Goldstein and Aniket Kamat and Fucheng Warren Zhu},
  journal={SoftwareX},
  year={2026},
  volume={33},
  pages={102508},
  publisher={}
}

–

Related Work

@article{jerzak2023degrees,
  title={Degrees of Randomness in Rerandomization Procedures},
  author={Jerzak, Connor T. and Rebecca Goldstein},
  journal={ArXiv Preprint},
  year={2023},
  volume={},
  pages={},
  publisher={}
}