Applying ICH E9 Principles to Bayesian Evidence

Why ICH E9 Principles Matter for Bayesian Analyses

A Bayesian analysis can summarize uncertainty about treatment effects using posterior distributions, but those distributions do not define the clinical question by themselves. A prior, likelihood, and decision rule can be mathematically coherent while still answering a question that is poorly matched to the trial objective, endpoint, or target population. ICH E9 and E9(R1) provide a framework for making the path from the clinical objective to the reported result explicit.

The practical task is not to make Bayesian methods look like frequentist methods. It is to document the question, design, conduct, analysis, and interpretation so a reviewer can see what evidence the analysis represents, what assumptions it uses, and where its conclusions stop. The introductory material on regulatory evidence standards provides the broader context; here the emphasis is on applying the E9 principles to the analysis record.

Connect the Objective to an Estimand

ICH E9(R1) describes an estimand as a precise summary of the treatment effect that a trial seeks to estimate. Its attributes identify the target population, the treatment conditions being compared, the endpoint, how intercurrent events are handled, and the population-level summary. An intercurrent event is an event after treatment initiation that affects either the interpretation or existence of the endpoint, such as starting rescue medication, discontinuing assigned treatment, or switching treatment.

For example, a treatment-policy strategy uses outcomes regardless of whether an intercurrent event occurs. A hypothetical strategy asks what the outcome would have been under a specified hypothetical condition, such as if rescue medication had not been used. These are different clinical questions and can require different data and assumptions. The strategy should be stated in the estimand, not left implicit in the choice of analysis population or model.

The estimand should also make its endpoint and time point unambiguous. “Improvement at follow-up” is not enough if improvement could mean a change score, a threshold response, or a measurement at different visits. Defining these details lets the reader assess whether the chosen likelihood and posterior summary correspond to the intended clinical effect.

Document the Bayesian Analysis as a Traceable Chain

A transparent report links the estimand to the observed data and then to the posterior evidence. It identifies the analysis population, endpoint definition, data cut, treatment of missing outcomes, model, prior distributions, computation, and posterior summaries. It also reports relevant departures from the prespecified analysis and explains their consequences. The goal is for another analyst to understand what was analyzed and why, rather than to infer it from a result table.

For each model component, state the assumption that makes its use meaningful. In a simple binary-outcome model, for example, one may assume that outcomes within each arm are conditionally independent Bernoulli observations with a common arm-specific response probability. The likelihood is then binomial. A Beta prior for each probability gives a conjugate posterior, but the prior is still an assumption about the analysis and must be reported, not hidden by the convenience of the calculation.

The posterior summary must also match the estimand. If the estimand is a difference in response probabilities, reporting only separate arm-specific posterior means can obscure the contrast of interest. Report the posterior for the contrast, or summaries derived from it, and state the direction and units. A posterior probability such as \(P(p_T-p_C>0\mid \text{data})\) is a probability about the parameter conditional on the model and observed data; it is not the probability that the null hypothesis is true.

Show How Intercurrent Events and Missing Data Were Handled

An estimand’s intercurrent-event strategy and the missing-data method are related but not interchangeable. A treatment-policy strategy specifies that outcomes after an event remain relevant to the estimand. If some of those outcomes were not observed, a statistical method is still needed to address their absence. Conversely, analyzing only participants who completed assigned treatment can change the target population and may no longer address the stated estimand.

The report should state how many outcomes are missing by arm, why they are missing when known, and what assumptions the primary analysis makes about them. It should distinguish observed data from modeled or imputed values. Sensitivity analyses can then show how conclusions change under alternative plausible assumptions. A deliberately extreme scenario can be informative as a stress test, but it is not automatically a valid model for the missing outcomes.

Justify the Result and Its Limits

The basic Beta-binomial update illustrates why the model and assumptions belong in the report. If \(Y\mid p\sim\operatorname{Binomial}(n,p)\) and \(p\sim\operatorname{Beta}(a,b)\), then the likelihood is proportional to \(p^y(1-p)^{n-y}\), while the prior density is proportional to \(p^{a-1}(1-p)^{b-1}\). Multiplying these gives a posterior density proportional to

which is the kernel of a \(\operatorname{Beta}(a+y,b+n-y)\) distribution. The update follows from the stated likelihood and prior. It does not establish that the common-probability binomial model, the prior, or the endpoint definition is appropriate for every clinical setting. Those choices require clinical and statistical justification.

A useful report therefore includes checks for data and model problems: verify denominators and outcome coding; compare observed enrollment and follow-up with the intended analysis population; inspect whether outcome patterns differ substantially across sites, time, or clinically relevant subgroups; and assess whether the model’s assumptions are credible. If a model cannot represent important design features or observed data patterns, its posterior precision may be misleading. Computation should also be checked when simulation is used, including convergence and Monte Carlo error for reported summaries.

State limitations directly. A posterior probability is conditional on the model, prior, data, and analysis choices. A credible interval has a posterior probability interpretation conditional on those choices; it is not a guarantee of repeated-trial performance. A posterior mean difference does not establish clinical importance unless the scale and clinically relevant magnitude are considered. And a polished Bayesian analysis cannot compensate for weak endpoint measurement, incomplete follow-up, or a mismatch between the estimand and the available data.

Comparison With Frequentist Reporting

ICH E9 principles apply to the clinical question and evidential transparency, not to one inferential framework alone. A frequentist analysis might report an estimated treatment contrast with a confidence interval and a p-value under a specified testing procedure. A Bayesian analysis might report a posterior distribution, credible interval, or posterior probability for a prespecified contrast or threshold. These summaries have different interpretations, but both should be tied to the same clearly specified estimand and trial data.

Do not translate one framework’s quantity into the other’s language. A 95% credible interval has 95% posterior probability under the stated model that the parameter lies in the interval; a 95% confidence interval has a repeated-sampling coverage interpretation under its procedure. Neither should be presented as evidence that a clinical question was adequately defined. The choice of framework does not remove the need to describe design, conduct, missing data, assumptions, and limitations.

Common Documentation Pitfalls

• Starting with the model instead of the question. A convenient likelihood does not determine the endpoint, target population, or treatment contrast. Define the estimand first.
• Leaving intercurrent events implicit. State whether outcomes after rescue treatment or discontinuation contribute to the target effect and how the strategy is implemented.
• Calling missing outcomes “nonresponders” without qualification. This is an assumption or scenario, not an observed fact. Name it and assess its impact.
• Reporting arm-specific results but not the target contrast. If the estimand is a difference, report the posterior evidence for the difference with units and direction.
• Presenting a posterior probability as unconditional certainty. State that the probability is conditional on the specified model, prior, and observed data.