Building the protocol around the estimand
A protocol that uses the estimand framework is structured around the clinical question of interest. The framework shapes the objectives, the estimands with their attributes, the endpoints and the statistical considerations, and reaches the study design and the schedule of activities. Public protocol templates, example protocols and further references show how estimands can be set out.
Why the protocol carries estimands
ICH E9 is final, and estimands need to be included in clinical trials, so they should be in the protocol, which feeds through to the clinical study report.
Estimands add clarity and alignment to a protocol, to improve the interpretation of its results, and they improve communication among all internal teams. They involve cross-functional collaboration, which starts with the protocol. FDA, EMA and ICH encourage discussion on constructing estimands across disciplines and between sponsors and regulators. They do so because regulatory statisticians often struggled with protocols that contained scattered details and objectives and estimands that were not aligned. Including well-defined estimands can improve clarity and reduce misunderstandings between sponsors and regulatory assessors, and so, ultimately, bring medicines to patients faster.
The clinical question of interest
The protocol and its estimands should be structured around the key clinical question of interest, the red thread of the study objectives. Estimands should be defined first, so as to guide the study design and the analysis. The structure of the protocol should follow the same order: the objectives, then the estimand, which is what is to be estimated, then the estimator, the method of estimation, and then the estimate, the numerical result of the estimator, which is the treatment effect.
Intercurrent events must be identified before the trial starts, so they are best in the protocol from the beginning. Adding a new intercurrent event after the trial has started means a substantial protocol amendment, which is not a good idea.
The sections the framework affects
Of the protocol sections the estimand framework affects, the most important are the objectives, endpoints and estimands, and the statistical considerations.
Among the other sections, collecting and monitoring the anticipated intercurrent events through the schedule of activities is very important. The study design and conduct should be planned around the estimands, detailed reasons for anticipated intercurrent events should be collected, and concomitant and rescue medication should be recorded in relation to the intercurrent-event strategy.
Objectives
A protocol that includes estimands can describe its objectives in two ways: as detailed clinical objectives or as less detailed objectives. Detailed clinical objectives, a method introduced by James Bell, are a very detailed and technical way of describing the objectives, and less detailed objectives are the form known from traditional protocols.
Detailed clinical objectives describe all five attributes of the estimand in prose within a single objective, and they already contain the clinical question. They form the basis for objective-driven trial design, and because they put everything together in one objective, they encourage discussion among stakeholders. Bell's publication on the approach gives text templates for writing them, and is recommended for anyone interested in the approach. Their attributes, however, are somewhat unclear, and the more complex a trial gets, the more difficult they are to understand.
With less detailed objectives, the clinical question should be added, for example: what is the mean difference in the primary endpoint after 24 weeks of treatment with the drug compared with placebo in patients with type 2 diabetes, as if no emergency medication had been used? The question gives context and links the objectives to the endpoints.
Whichever form the objectives take, detailed handling of intercurrent events is needed elsewhere in the protocol.
Estimands
After the objectives, whether they are detailed, less detailed or a mixture of both, the estimands should be added with their five attributes in bullet points or in a table, so that each attribute is clearly identifiable. This layout is recommended, though not necessary, and it makes the protocol very transparent: anyone who reads it can immediately identify which attributes the trial used. Trials are also easier to compare when their attributes are listed, the listing ensures that the intercurrent events are handled properly, and an estimand that differs only in, for example, its intercurrent-event handling strategy is easier to copy to a different section.
Intercurrent events can affect the interpretation of the outcome, which is why they need to be reported properly. Two handling strategies applied to the same endpoint might lead to different results, which can confuse, so it is important to list all the attributes and the intercurrent-event handling strategies. In one example, a treatment policy estimand, which takes all data regardless of rescue medication and describes the real-world scenario, stands beside a trial product estimand, which excludes data after rescue medication and imputes them to describe the ideal efficacy scenario.
An estimand is best named for its role and not for its strategy: labels such as primary estimand or real-world estimand help, but labels linked to the intercurrent-event handling strategies, such as hypothetical estimand or composite estimand, are best avoided. In one protocol whose estimand attributes were neither properly listed nor set out in a table, and whose estimands were labelled hypothetical and composite, finding out which attributes had been used took a long time.
Endpoints
Endpoints stay in their own section even though the endpoint is one of the five attributes, because not every endpoint has an estimand: some secondary endpoints have none. ICH E9(R1) requires estimands for primary and confirmatory secondary endpoints only. A clear endpoint description is needed for disclosure on ClinicalTrials.gov or CTIS.
Statistical considerations
The statistical considerations in section 9 should be structured like section 3, by objectives and endpoints, because not every endpoint is linked to an estimand. A statistical hypothesis should be linked to its estimand and to the tests related to it.
The intercurrent events are best given a separate section on how they are to be handled. That section should list them, or set them out in a table, together with their handling strategies, and either state their statistical implications there or, preferably, cross-reference the sections of the protocol they influence.
The two public templates
The TransCelerate protocol template and ICH M11 are both publicly available, and both cover estimands mainly in section 3 and section 9. The TransCelerate template is somewhat more detailed, with text examples: it shows the clinical question of interest, addresses the estimand attributes and sets them out in bullet points.
The protocol libraries on TransCelerate's website are recommended. They hold protocol templates tailored to specific participant groups, such as healthy participants and possibly pregnant ones, and to therapeutic areas such as diabetes. They give examples of endpoints, for other indications too, and examples of what FDA guidance favours.
In section 9 both templates have a separate section for intercurrent-event strategies, which are needed for any estimand. Both give an example of a hypothesis, and the templates are best checked directly.
Detailed clinical objectives can be used in either template, and where the intercurrent events are then handled in full detail in section 9, all five attributes are covered. The five attributes should still be added in both section 3 and section 9.
Example protocols
A protocol for a study in Alzheimer's disease states the clinical question clearly, describes the endpoint attribute and the other four attributes, and gives a rationale for the estimand, which is always important. One experienced writer still rates it only an acceptable example, because all of that stands in the statistical analysis section, while section 3 holds only the endpoints, with no link to the estimands.
The same writer rates a protocol for a study in patients with diabetes a very good example. Its section 3 holds the objectives, the clinical question the trial asks, a rationale for the estimand and the intercurrent events, together with a table that links the objective to the five attributes and defines a handling strategy for each specific intercurrent event.
Where to look when a template is silent
A protocol author does not necessarily need to be able to fully understand and implement ICH E9 and E9(R1), which are statistically focused, but a strong understanding of what good looks like will help a protocol. Where a template contains no estimand guidance, a reference on estimands in protocol templates is where one experienced writer would look. A two-year perspective on an estimand journey, by some of the original architects of estimands, goes through the rationale for the choices they made, which helps a reader understand them.