📊 statistics

Fast Power Evaluation under Biased-Coin Minimization: Sampling and Randomization Calibration

This paper introduces the SIGA framework, comprising SIGA-S and SIGA-R procedures, to provide computationally efficient and accurate power and sample-size evaluations for trials using biased-coin minimization by approximating randomization test variances through reusable stratum-imbalance Gaussian models and paired allocation calibrations.

Original authors: Masahiro Kojima

Published 2026-08-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Masahiro Kojima

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but the clues are scattered across different neighborhoods. In the world of medical research, scientists often run "clinical trials" to see if a new medicine works better than a placebo. To make sure the test is fair, they need to balance the groups of people they study. If one group has mostly tall people and the other has mostly short people, the results might be skewed by height, not the medicine. This balancing act is called "randomization."

Usually, scientists use a simple coin flip to decide who gets which treatment. But sometimes, the coin flip isn't enough, especially if there are many different factors to balance, like age, weight, and blood pressure. If they tried to balance every single combination of these factors, they would end up with so many tiny, empty groups that the test would fall apart. So, they use a smarter, more complex method called "minimization." It's like a GPS that constantly checks the traffic (the balance of factors) and steers the next person into the group that needs them most to keep things even.

However, this smart steering creates a problem for the detectives (the statisticians) later on. When they try to calculate the odds that their results happened by chance, the usual math tools break down because the "coin flips" weren't actually random; they were guided by the GPS. To fix this, they have to run a massive computer simulation, re-playing the entire trial thousands of times to see what the results would have been if the GPS had made different choices. This is like trying to predict the weather by running a supercomputer simulation of the entire atmosphere every single time you want to know if it will rain tomorrow. It's incredibly accurate, but it takes so much computing power that it can take days or even weeks to plan a single study.

This is where Masahiro Kojima's paper comes in. He asked a simple question: "Can we predict the weather without running the whole simulation?" He developed two new, super-fast shortcuts called SIGA-S and SIGA-R. Think of these as a "weather forecast" based on the current wind patterns rather than a full atmospheric simulation.

The paper's main finding is that these shortcuts work almost perfectly. In a series of computer simulations involving 56 different scenarios (with different numbers of people and different types of data), the new methods predicted the results of the slow, heavy simulations with incredible accuracy. The difference between the fast shortcut and the slow, perfect method was often less than half a percentage point—so small it's barely noticeable.

But here is the twist: the paper shows that there are actually two different things you might want to predict, and you need the right shortcut for the right job.

SIGA-S is the shortcut for predicting the "real-world" average effect. It's like asking, "If we ran this experiment a million times in the real world, how often would we see this result?"
SIGA-R is the shortcut for predicting the "rulebook" result. It's like asking, "If we followed the strict rules of the coin-flip game exactly as written, how often would we see this result?"

The paper argues that these two questions are not always the same. If you use the wrong shortcut (like using SIGA-S when you need SIGA-R), you might get a slightly wrong answer about how many people you need for your study. However, the paper proves that by using a clever "pairing" trick—simulating two paths at once instead of one—SIGA-R can perfectly mimic the strict rulebook without the heavy computing cost.

The author is very sure about these findings because they tested them rigorously. They didn't just guess; they ran 100,000 simulated trials for each scenario. The results showed that the new methods are not just "good enough"; they are statistically indistinguishable from the slow, perfect method in almost every case.

The most exciting part of the paper is the speed. The author compared their new methods to the old, slow way on a standard computer. The old way took anywhere from 113 minutes to over 2,600 minutes (more than 43 hours!) to run the simulations for one study design. The new SIGA methods? They took between 0.05 minutes and 2.25 minutes. That is a speedup of thousands of times.

In short, this paper doesn't just offer a tiny improvement; it turns a process that used to take days into one that takes minutes. It gives researchers a way to plan complex, fair medical trials quickly and accurately, ensuring that the "GPS" for balancing patients is used correctly without needing a supercomputer to figure out the odds. The paper suggests that by choosing the right shortcut (SIGA-S or SIGA-R) based on what question you are asking, you can save massive amounts of time and computing power while keeping your science just as reliable.

Technical Summary: Fast Power Evaluation under Biased-Coin Minimization

Problem Statement

In clinical trial design, Pocock–Simon minimization is widely used to balance treatment allocation across prognostic factors, particularly when stratified permuted-block randomization would result in sparse strata. However, evaluating statistical power and sample size for trials using this method is computationally intensive.

The core challenge arises because the standard reference procedure—a design-based rerandomization test—requires regenerating treatment paths within every simulated trial to calculate $p$ -values. This creates a "nested Monte Carlo" problem: for each outer simulation trial (representing a potential study outcome), a large number of inner rerandomization paths must be generated to approximate the conditional randomization distribution. When evaluating power across numerous candidate sample sizes, treatment effects, and outcome models, this nested calculation imposes a prohibitive computational burden.

Furthermore, a theoretical distinction exists between the repeated-sampling distribution of an observed statistic and the conditional randomization distribution of a fixed-score test. Under minimization, if the null hypothesis is "weak" (e.g., a marginal mean difference where individual effects vary) rather than "sharp" (constant additive effect), the observed allocation path is embedded in the score statistic. Consequently, the repeated-sampling variance of the observed statistic may differ from the conditional variance of the reference randomization test. An efficient approximation method must therefore be calibrated to the correct inferential target: either the repeated-sampling distribution (for average-effect inference) or the conditional randomization distribution (to reproduce a prespecified randomization test).

Methodology

The author develops a reusable Stratum-Imbalance Gaussian Approximation (SIGA) framework. This approach decomposes a fixed-score statistic into two orthogonal components:

Joint-stratum imbalance component: Driven by treatment imbalances within the joint strata formed by all minimization factors.
Orthogonal within-stratum component: Driven by score contrasts within strata.

Based on this decomposition, the paper proposes two distinct procedures:

1. SIGA-S (Sampling-Calibrated)

Target: Repeated-sampling inference on the marginal average treatment effect.
Mechanism: SIGA-S uses an allocation-only calibration. It generates $B_0$ independent allocation paths to estimate the covariance of the joint-stratum imbalance vector.
Variance Estimation: The variance estimator combines the calibrated covariance of the imbalance vector with a common variance for the orthogonal within-stratum subspace.
Validity: Theorem 3.1 establishes that SIGA-S consistently estimates the repeated-sampling variance and yields asymptotically valid inference at a marginal mean- or risk-difference boundary, regardless of whether the sampling and randomization variances align.

2. SIGA-R (Randomization-Calibrated)

Target: Efficient reproduction of the conditional randomization distribution of a prespecified fixed-score randomization test.
Mechanism: SIGA-R addresses the "pair-path" structure created when a fixed score (containing the observed allocation path) is evaluated against regenerated paths. It uses a reusable paired allocation-only calibration. In each calibration replicate, two independent treatment paths are generated conditional on the same factor sequence.
Variance Estimation: It estimates the "pair-path covariance" (the covariance between products of treatment signs from two independent paths). This allows the calculation of a variance-gap term ( $\hat{C}_n$ ) that adjusts the SIGA-S variance to match the conditional randomization variance of the reference test.
Validity: Theorem 3.2 shows that SIGA-R is asymptotically equivalent to the ideal reference randomization test. It reproduces the test's conditional distribution, including any weak-null size distortion inherent to the reference test itself.

Computational Strategy

Both procedures replace the nested rerandomization loop with a single, reusable allocation-only calibration step. The calibration estimates (covariance matrices and treatment-count probabilities) are computed once per design and sample size and reused across all outer simulation trials, treatment effects, and null boundaries.

Key Contributions

Exact Orthogonal Decomposition: The paper derives an exact projection of the treatment-score statistic into joint-stratum imbalance and within-stratum components. This leads to a design-level covariance calibration based solely on allocation simulations.
Formalization of Variance Discrepancy: The author formalizes the dependence created by holding a score fixed at a nonsharp boundary. They derive a design-specific identity for the variance gap between the repeated-sampling and conditional randomization variances, characterizing when the reference randomization test is asymptotically valid under a weak null and when it suffers from size distortion.
Reusable Pair-Path Calibration: The paper introduces a method to estimate the pair-path covariance using independent allocation-only replicates. This enables SIGA-R to approximate the conditional reference distribution of the prespecified randomization test without regenerating treatment paths within every simulated trial.
Dual-Procedure Framework: The author clarifies that SIGA-S and SIGA-R serve different inferential targets. SIGA-S is for average-effect inference, while SIGA-R is a computational surrogate for a prespecified randomization test. They are not competing methods for the same target.

Results

The author conducted two independent simulation studies across 56 scenario configurations (varying outcome types, minimization factors, sample sizes, and testing settings) and a trial-inspired illustration.

Accuracy:
- SIGA-S: In the sampling-targeted study, the maximum absolute difference between SIGA-S and the reference randomization test rejection probabilities was 0.42 percentage points (0.18 pp for Type I error, 0.42 pp for power).
- SIGA-R: In the randomization-targeted study, the maximum absolute difference between SIGA-R and the reference test was 0.32 percentage points.
- The estimated variance ratios ( $\hat{\rho}_n$ ) between the two targets ranged from 1.000 to 1.042, indicating that while discrepancies exist, they are often small in magnitude but can be practically relevant for sample size selection.
Computational Efficiency:
- The timing comparison demonstrated a reduction in computation by several orders of magnitude.
- For 100,000 analyses, SIGA-S required 0.050–2.259 minutes, whereas the reference randomization test required 113.04–2625.76 minutes.
- The one-time allocation-only calibration took between 2.8 and 47.5 seconds.

Significance and Claims

The paper claims that the SIGA framework provides a computationally feasible route for design-stage power and sample-size evaluation under biased-coin minimization without sacrificing inferential accuracy.

Separation of Targets: The author emphasizes that computational acceleration and inferential target selection cannot be separated. SIGA-S provides valid inference for the marginal average effect, while SIGA-R provides a first-order approximation to the conditional randomization distribution of a prespecified test.
No Correction of Distortion: The paper modestly notes that SIGA-R does not correct the weak-null size distortion of the reference randomization test; it merely reproduces the test's behavior. If the reference test is invalid under a weak null, SIGA-R will reproduce that invalidity.
Reusability: A central claim is the reusability of the calibration. The allocation-only simulations are independent of the outcome model, treatment effects, and null boundaries, allowing a single calibration to serve multiple design-stage evaluations.
Practical Impact: The framework enables researchers to perform extensive power analyses for complex minimization designs that would otherwise be computationally prohibitive due to the nested rerandomization requirement.

The author concludes that while SIGA-R is not a generally preferable successor to SIGA-S for all purposes, it is an essential tool when the design-stage objective is to approximate the rejection probability of a specific, prespecified randomization test.