Skip to content

Guidance Note: Bayesian Analysis of Clinically Meaningful Survival Benefit Using ESMO-MCBS v2.0

Executive Summary

This note synthesizes principles, thresholds, and implementation steps for evaluating clinically meaningful survival benefits in colorectal cancer clinical trials. It translates ESMO-MCBS v2.0 guidelines into statistical operating characteristics—including ROPE and priors—for Bayesian Piecewise Exponential Additive Models (PAMMs).

1. Context and Research Objective

The aim is to design a transparent and rigorous Bayesian analysis framework to: - Define clinically meaningful effect sizes for overall survival (OS) based on authoritative clinical guidance (ESMO-MCBS v2.0). - Operationalize these thresholds into a statistical analysis pipeline for survival data using Bayesian PAMMs, including ROPE and prior specification.

2. Key Insights from ESMO-MCBS v2.0 Guidelines

2.1. Thresholds for Clinically Meaningful Benefit

Curative (Adjuvant) Setting (Figure 2, p.18): - Grade A (substantial benefit): ≥5% improvement in 5-year OS - Grade B: ≥3% to <5% improvement in 5-year OS - Grade C: <3% improvement in 5-year OS

Non-Curative (Metastatic) Setting: - Thresholds defined primarily as absolute gain in median OS and hazard ratio (HR) with requirements increasing as baseline prognosis improves. - Dual criteria: Both an absolute gain (in months) and sufficient HR reduction required for substantial benefit. - For example, if control OS <12 months: substantial benefit = HR ≤0.7 and ≥3 months OS gain.

Unified Threshold Rationale:
For joint (curative + non-curative) or unclear-population analyses, use the most conservative threshold:
≥3% absolute improvement in 5-year OS (lowest mark for substantial benefit in curative setting).

3. Translating Clinical Guidelines to Statistical Criteria

3.1. ROPE (Region of Practical Equivalence)

  • Definition: Effects smaller than the minimum clinically meaningful threshold.
  • For OS: ±3% absolute difference at 5 years, i.e., 0.03 in either direction.
  • On HR scale: ROPE boundaries should correspond to HRs achieving a 3% difference in 5-year OS, given the population's baseline survival rate.

3.2. Priors for Optimistic Effect

  • For an optimistic prior, center the prior for the treatment effect at the HR (or log-HR) needed for a 5% absolute gain in 5-year OS.

4. Translating Absolute Survival Benefits to HRs and Model Coefficients

4.1. Theoretical Foundation

The relationship between hazard ratios and absolute survival differences depends critically on the baseline survival probability. Under the proportional hazards assumption, if the control group has survival probability S_control(t) at time t, then the treatment group's survival probability is:

\[S_{treat}(t) = [S_{control}(t)]^{HR}\]

Rearranging to solve for the HR required to achieve a target absolute benefit:

\[HR = \frac{\ln(S_{treat}(t))}{\ln(S_{control}(t))}\]

Key Assumption: This derivation assumes exponential survival (constant hazard) or that the HR represents the time-averaged effect over the period of interest.

4.2. Step-by-Step Calculation Method

Step 1: Define your baseline control survival probability at the time point of interest (typically 5 years) - Example: S_control(5 years) = 0.469 (46.9%)

Step 2: Calculate target treatment survival probability - For a 3% absolute gain: S_treat = 0.469 + 0.03 = 0.499 - For a 5% absolute gain: S_treat = 0.469 + 0.05 = 0.519

Step 3: Apply the HR formula - HR = ln(S_treat) / ln(S_control) - log(HR) = ln(HR) for use in statistical models

Step 4: Verify the calculation - Check: [S_control]^HR should equal S_treat

4.3. Worked Example

Control survival: S_control = 0.469
Target gains: 3% → S_treat = 0.499, 5% → S_treat = 0.519

For 3% gain: HR = ln(0.499)/ln(0.469) = -0.696/-0.758 = 0.918
For 5% gain: HR = ln(0.519)/ln(0.469) = -0.656/-0.758 = 0.865

log(HR): 3% gain = -0.085, 5% gain = -0.145

4.4. R Implementation

# Function to calculate HR from absolute survival benefit
calc_hr_from_abs_benefit <- function(s_control, abs_benefit) {
  s_treat <- s_control + abs_benefit
  hr <- log(s_treat) / log(s_control)
  log_hr <- log(hr)

  # Verification
  s_treat_check <- s_control^hr

  return(list(
    hr = hr,
    log_hr = log_hr,
    s_treat_target = s_treat,
    s_treat_check = s_treat_check,
    verification_error = abs(s_treat - s_treat_check)
  ))
}

# Example usage
baseline_os <- 0.469
result_3pct <- calc_hr_from_abs_benefit(baseline_os, 0.03)
result_5pct <- calc_hr_from_abs_benefit(baseline_os, 0.05)

print(paste("3% benefit: HR =", round(result_3pct$hr, 3), 
            ", log(HR) =", round(result_3pct$log_hr, 3)))
print(paste("5% benefit: HR =", round(result_5pct$hr, 3), 
            ", log(HR) =", round(result_5pct$log_hr, 3)))

4.5. Alternative Time Points

The same methodology applies to any time point. For 3-year survival:

# Example for 3-year endpoint
s_control_3yr <- 0.55  # 55% 3-year survival
abs_benefit <- 0.03    # 3% absolute benefit

hr_3yr <- log(s_control_3yr + abs_benefit) / log(s_control_3yr)

5. Bayesian PAMM Implementation

5.1. Model Structure Example

formula = ped_status ~ s(tend) + cohort + offset(offset)

5.2. ROPE Specification

  • On HR scale: [0.92, 1.08] (3% difference in either direction; upper bound is ~1/0.92)
  • On log(HR) scale: [-0.083, +0.083]

5.3. Prior Specification

  • Optimistic prior (e.g., for 5% absolute benefit):
    log(HR) = -0.145
  • In brms R syntax: 
    prior <- set_prior("normal(-0.145, 0.10)", class = "b", coef = "cohort")
    Adjust standard deviation for desired prior informativeness.

5.4. Posterior Analysis

  • Extract posterior for the treatment (cohort) coefficient.
  • Calculate the proportion of the posterior lying below log(HR) = -0.083 (HR < 0.92)—this is the probability of achieving at least a 3% OS benefit.
  • Use the ROPE interval for making decisions about statistical and clinical significance.

6. Clinical and Statistical Notes

  • The mapping between HR and % OS gain depends on the control event rate. Always recalculate if baseline changes.
  • The exponential model is an approximation; PAMMs can flexibly handle non-proportional hazards, but the HR at 5 years remains a useful clinical anchor.
  • Always clearly state assumptions when reporting.

7. Summary Table

5-year OS Gain HR log(HR)
3% 0.92 -0.083
5% 0.87 -0.145

8. Final Guidance & References

  • Reference: ESMO-MCBS v2.0 full document (PDF)
    • See p.16, p.18 (curative), p.22–29 (metastatic).
  • Use a 3% 5-year OS gain (HR 0.92) as the primary clinical threshold for meaningful effect in Bayesian PAMM analysis.
  • Set ROPE and priors accordingly; interpret the posterior in this clinical context.

This note formalizes a statistically rigorous, clinically grounded Bayesian analysis plan for evaluating the survival benefit of treatments in colorectal cancer, aligned with ESMO-MCBS v2.0 benchmarks.