Bayesian vs Frequentist Approaches to Sample Size in Clinical Trials

In clinical trial planning, determining the correct sample size is one of the most critical design decisions. Traditionally, most studies have used the frequentist framework to estimate sample sizes. However, the Bayesian approach is gaining traction, especially in adaptive and complex designs. This article explores both paradigms—highlighting their principles, applications, and implications for regulatory acceptance and scientific robustness.

Understanding how these two frameworks differ and where each excels is essential for trial statisticians, regulatory teams, and QA professionals. We’ll also explore how both approaches interact with guidelines from regulatory bodies like the USFDA and EMA.

Core Philosophy: Bayesian vs Frequentist Thinking

Frequentist Approach

• Parameters are fixed but unknown
• Probability is defined as the long-run frequency of events
• Inferences are based on repeated sampling
• Sample size aims to control type I (α) and type II (β) error rates

Bayesian Approach

• Parameters are random variables with distributions
• Probability reflects the degree of belief, updated with data
• Uses prior and posterior distributions to make inferences
• Sample size is based on predictive probability, utility functions, or credible intervals

Frequentist Sample Size Determination

Inputs Required:

• Type I error (usually α = 0.05)
• Desired power (typically 80–90%)
• Effect size to detect
• Outcome variability or event rate

Typical Formula (for comparing two means):

  n = 2 × (Z1−α/2 + Z1−β)² × σ² / Δ²
  

• σ²: variance
• Δ: clinically relevant difference

Advantages:

• Widely accepted by regulatory agencies
• Straightforward for simple designs
• Established error control methods

Limitations:

• Inflexible in adaptive or sequential trials
• Requires fixed design assumptions
• Cannot incorporate prior knowledge

Bayesian Sample Size Determination

Bayesian methods focus on the probability of achieving a desired posterior result, given the trial data and prior information.

Common Methods:

• Posterior probability criteria: e.g., P(θ > θ₀ | data) ≥ 0.95
• Credible intervals: Ensure the width of a 95% credible interval is below a threshold
• Predictive power: The probability that the posterior result exceeds the success criterion
• Decision-theoretic approaches: Based on expected loss or gain

Inputs Required:

• Priors (informative or non-informative)
• Expected data distributions
• Simulation settings to evaluate trial operating characteristics

Example in R:

  library(BayesFactor)
  result = ttestBF(x = sample_data, y = control_data)
  plot(result)
  

Advantages:

• Can incorporate external data or expert opinion
• Highly adaptable to changing trial conditions
• Well-suited for adaptive designs and rare diseases

Limitations:

• Requires careful selection and justification of priors
• Regulatory familiarity still developing in some regions
• Computationally intensive (needs simulations)

Regulatory Viewpoints

The pharma regulatory compliance landscape is evolving with increasing acceptance of Bayesian methods, particularly in areas like:

• Medical devices (especially by the USFDA’s Center for Devices)
• Rare disease trials with limited subject pools
• Early-phase exploratory studies

However, regulators often require:

• Justification of prior selection
• Extensive simulation-based operating characteristics
• Documentation of robustness to prior sensitivity

Guidance from both the USFDA Bayesian guidance and EMA reflection papers support Bayesian use when clearly justified.

Key Differences at a Glance

Best Practices for Choosing Between Them

1. For simple, fixed designs with large sample sizes, the frequentist approach is sufficient and more universally accepted.
2. For adaptive designs or rare diseases with limited subjects, Bayesian methods offer flexibility and efficiency.
3. Document assumptions and simulations extensively in the protocol and pharma SOP documentation.
4. Use simulation to compare operating characteristics across both approaches.
5. Ensure team training on Bayesian methods for correct implementation and interpretation.

Conclusion: A Complementary Approach for Modern Trials

Neither Bayesian nor frequentist approaches are universally better—they serve different purposes based on the study context. While frequentist methods provide simplicity and regulatory comfort, Bayesian techniques offer adaptability and richer inference capabilities. Understanding both frameworks equips clinical teams to select the right tool for each trial’s complexity, resource, and regulatory landscape.

Explore More:

• GMP validation
• Stability indicating methods
• Computer system validation