Blinded sample size recalculation in randomized controlled trials with analysis of covariance

arXiv:2508.18815 · stat.ME · Submitted 2025-08-26 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Blinded sample size recalculation in randomized controlled trials with analysis of covariance".

Jane: Here's the gist of the paper. In randomized controlled trials, covariate adjustment can improve statistical power and reduce the required sample size compared to unadjusted estimators This study proposes a blinded sample size recalculation method for the ANCOVA estimator based on asymptotic relative efficiency under minimal distributional assumptions.

Tom: Pruning redundant reasoning tokens The authors propose a blinded sample size recalculation method for the ANCOVA estimator that does not require assumptions regarding the joint distribution of the outcome and baseline covariates, nor does it rely on correct specification of the ANCOVA model.

Jane: This method determines the initial sample size based on an unadjusted estimator at design stage and conducts sample size recalculation during an interim analysis.

Tom: Furthermore, because the proposed method does not impose assumptions regarding the baseline covariate distribution, it is applicable even when adjusting for non-normally distributed or discrete covariates.

Paper summary: Tom: To recap, this paper introduces a blinded sample size recalculation method specifically tailored for the ANCOVA estimator in randomized controlled trials. The main thesis is that covariate adjustment can actually improve statistical power and reduce the total number of participants needed compared to just using an unadjusted estimate.

Jane: They address the big hurdle that’s always existed when planning these studies: you need to know how well those baseline covariates relate to your outcome, but figuring out those precise associations at the start of a trial is often incredibly difficult.

Lu: The paper proposes a method where you determine the initial sample size using an unadjusted estimator right at the design stage, but then you recalculate it during an interim analysis using this new approach instead of sticking with your original plan.

Meng: This method is designed to avoid making strong assumptions about the joint distribution of your outcome and those baseline covariates, and critically, it doesn't rely on having a perfectly specified ANCOVA model when you start. That lack of dependency on perfect initial specification is what makes it appealing for real-world applications.

Lalam: They build this method based on the asymptotic relative efficiency between the ANCOVA estimator and the unadjusted estimator, using terms derived from consistent estimators of residual variance and within-arm variance to find that new sample size.

Jane: And they stress that because this approach doesn't assume a specific distribution for your baseline covariates—like assuming everything is multivariate normal—it’s applicable even if you are dealing with non-normal or discrete variables.

Tom: This matters because most standard methods require those strong distributional assumptions, but this paper seems to push the boundary by focusing on minimal assumptions while still controlling the type one error rate during interim analyses.

Lu: The simulations they ran showed that when testing arbitrary model misspecification, like if your working ANCOVA model was wrong, both their proposed method and a simpler recalculation method maintained nominal power of eighty percent in scenario four.

Meng: That’s good for production because it suggests the calculation isn't overly brittle; even when things get messy with the model structure, the required sample size estimate stays reasonable under these test conditions.

Lalam: It also pointed out that the simpler recalculation method might still be okay even past where it was originally derived, which is a bit of a surprise when you’re used to strict derivation settings.

Conclusion: Tom: So, looking at the title of this work, "Blinded sample size recalculation in randomized controlled trials with analysis of covariance," what they’ve really done is provide a practical tool for trial designers who are worried about adjusting for baseline data.

Jane: It’s about taking the challenge of needing covariate adjustment and finding a way to do that without needing perfect upfront knowledge of the data's distribution or the exact model structure used.

Lu: The implication is that researchers can move more flexibly during a trial, using interim data to adjust their sample size requirements in a way that controls error rates even when things aren't textbook perfect.

Meng: For someone shipping models to production, this means if you’re running an adaptive trial where you might need to adjust based on accumulating evidence, this method gives you a more flexible calculation tool instead of being locked into one rigid design.

Lalam: Ultimately, it gives researchers a path forward when they want to use ANCOVA for adjustment but don't have the luxury of perfect upfront assumptions about the underlying data structure.

Takumi Kanata, Yasuhiro Hagiwara, Koji Oba

Department of Biostatistics, School of Public Health, Graduate School of Medicine, The University of Tokyo · Department of Clinical Data Science, Clinical Research & Education Promotion Division, National Center of Neurology and Psychiatry, Tokyo · Interfaculty Initiative in Information Studies, The University of Tokyo

stat.ME

Submitted: 2025-08-26

Updated: 2026-07-21

Journal ref: Biometrics 82(3), ujag147 (2026)

DOI: 10.1093/biomtc/ujag147

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: In randomized controlled trials, covariate adjustment can improve statistical power and reduce the required sample size compared to unadjusted estimators This study proposes a blinded sample size…

Key concepts

ANCOVA Estimator
ANCOVA (Analysis of Covariance) is a statistical method used to compare treatment groups while statistically accounting for baseline differences in a continuous variable. The ANCOVA estimator is the specific statistic derived from this analysis that the study focuses on recalculating the required sample size for.
Asymptotic Relative Efficiency
This concept compares how efficient two estimators are as sample sizes get very large. The paper uses this efficiency to determine how much more precise an ANCOVA estimator is compared to a simpler, unadjusted estimator, which helps in calculating the necessary sample size.
Model Misspecification
Model misspecification occurs when the statistical model used (like the ANCOVA model) does not perfectly match the true underlying data-generating process. The proposed method is designed to be robust against this issue because it relies on efficiency under minimal distributional assumptions, making it applicable even if the assumed model structure is slightly incorrect.
Blinded Sample Size Recalculation
This is a procedure where the required sample size for a study is calculated initially at the start and then updated later during an interim analysis, without knowing which group the subjects belong to. The method proposed here achieves this by using observed outcomes and efficiency measures rather than strict distributional assumptions.

Terminology

Summary

In randomized controlled trials, covariate adjustment can improve statistical power and reduce the required sample size compared to unadjusted estimators This study proposes a blinded sample size recalculation method for the ANCOVA estimator based on asymptotic relative efficiency under minimal distributional assumptions. p. 0

Pruning redundant reasoning tokens

The authors propose a blinded sample size recalculation method for the ANCOVA estimator that does not require assumptions regarding the joint distribution of the outcome and baseline covariates, nor does it rely on correct specification of the ANCOVA model. This method determines the initial sample size based on an unadjusted estimator at design stage and conducts sample size recalculation during an interim analysis. p. 1 Furthermore, because the proposed method does not impose assumptions regarding the baseline covariate distribution, it is applicable even when adjusting for non-normally distributed or discrete covariates. p. 1

How the training signal is built

The method is based on the asymptotic relative efficiency of the ANCOVA estimator compared to the unadjusted estimator. The asymptotic relative efficiency of the ANCOVA estimator βbA and the unadjusted estimator αbA is given by equation (6). p. 2 The proposed sample size recalculation formula, Nbrec, incorporates terms derived from consistent estimators of the residual variance and the within-arm variance. p. 4

Where the gains come from

The proposed method is designed to accommodate arbitrary model misspecification by being based on asymptotic relative efficiency under minimal distributional assumptions. The simulation studies demonstrated that the type I error rate was adequately controlled (Section 4). p. 3 In the ACTG 175 trial application, the minimum recalculated sample size was 100 when baseline CD4 count and antiretroviral history were adjusted, and the maximum recalculated sample size was 158. p. 5

Limitations and Future Directions

The proposed method has several limitations, including only using observed outcomes, which excludes pipeline participants. Additionally, the method assumes that treatments were randomly assigned at a 1: 1 ratio independent of baseline covariates. p. 5 The authors recommend that the final sample size should be determined through a comprehensive assessment that also considers safety and other relevant factors, especially for safety outcomes. p. 9

Simulation Robustness

The simulation study evaluated the performance under various scenarios, including those with arbitrary model misspecification where the working ANCOVA model was misspecified. The results showed that both methods achieved nominal power of 80% in scenario (4) even when adjusting for all baseline covariates. p. 7 The proposed method maintained nominal power when adjusting for a larger number of baseline covariates compared to the simple recalculation method. p. 7 The results also showed that the simple recalculation method may remain approximately valid beyond its original derivation settings. p. 9

Code Implementation

The paper provides R and SAS macro codes for implementing the proposed method, including a function n.ancova.recal in R and a macro n ancova recal in SAS. p. 3 These codes allow researchers to apply the proposed method to their data. p. 1 The R code implements the formula Nbrec using the min term in the denominator of the first term. p. 3 The SAS macro implements a similar calculation using ods output commands for ANOVA and summary statistics. p. 3 The R code lists the formula Nbrec, which is an extension of equation (10). p. 3 The SAS macro implements the calculation of n ancova using the formula derived from equation (31). p. 3

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Blinded sample size recalculation for ANCOVA estimators can be implemented to reduce required sample size without inflating type I error rate in final analysis, as proposed by the method that does not require assumptions regarding the joint distribution of the outcome and baseline covariates, nor does it rely on correct specification of the ANCOVA model.

  2. The proposed method allows for sample size recalculation during an interim analysis under blinding, minimizing practical or operational biases that may affect clinical trial validity, which is ideal for minimizing the number of individuals who become aware of the treatment effect during the course of the trial.

  3. The system can perform sample size recalculations based on asymptotic relative efficiency under minimal distributional assumptions, accommodating arbitrary model misspecification where existing methods assume a multivariate normal joint distribution.

  4. The system can handle covariate adjustment for non-normally distributed or discrete covariates, such as sex, because the proposed method is applicable even when adjusting for non-normally distributed or discrete covariates.

  5. The system can dynamically adjust the required sample size at an interim analysis by calculating Nbrec using formula (10), which involves a minimum function to prevent an excessive reduction in the recalculated sample size when the estimated pooled variance overestimates the true variance.

Related papers