Estimands and intercurrent events

An estimand states precisely which treatment effect a trial measures. The framework around it turns on why the intention-to-treat default was not enough, the attributes that define an estimand, the intercurrent events that interrupt a trial, the strategies for handling such an event and what each does to the estimated effect, the data a trial has to collect for the estimand it chooses, and who chooses the estimand and in what order.

What an estimand is and why a writer needs the term

An estimand is a clinical entity or parameter that is estimated by performing a clinical study, in other words the target of estimation. ^{1, 2} It precisely describes the treatment effect that reflects the clinical question, and ICH E9(R1) adds it as a bridge between the clinical question and the analysis. ^{3, 4} Estimands are detailed definitions of quantities to be estimated using clinical trial data that make allowance for events that happen after randomisation. ² Compared with endpoints as clinical trial protocols defined them, estimands are more detailed definitions of the quantity to be estimated. ^{5, 6}

Estimand, estimator and estimate name three different things. ⁷ The estimand is the target of estimation, what is to be estimated; the estimator is the method of estimation, how it will be estimated; the estimate is the numerical result the estimator produces, the treatment effect. ⁷ In the statistical analysis of clinical data, an estimate is counterintuitively not an approximate judgment or an informed guess but the result of a defined calculation. ¹ The term highlights that all clinical research can do in evaluating an endpoint such as mean blood sugar is to estimate its true value in the population. ¹ Even large studies, in diabetes for example, can return results only for the patients in the study, so every computation in a clinical study estimates a true population value that can never be known. ¹

The word estimand may look like a spelling mistake, but it represents a new paradigm in clinical research, with a set of concepts that will change the way clinical studies are performed. ^{1, 2} Estimands help to define what kind of treatment effect is to be estimated and how it is to be estimated, which is more or less the statistical analysis. ^{8, 9}

An estimand reflects the clinical question and serves as a guiding thread through a drug's lifecycle documents. ¹⁰ Its five attributes give all the stakeholders a common language, so it bridges disciplines and improves communication between sponsors and agencies. ^{3, 10, 11} One practitioner holds that estimands reflect values of collaboration, communication, leanness and clarity, which play a crucial role in designing better trials and ultimately bring medicines to patients faster. ¹⁰ Estimands will help researchers to formulate more clearly what they really want to get out of a clinical study. ^{12, 13} The framework fosters a precise definition of the treatment effect and is a systematic approach to aligning study objectives, design, conduct, data collection, statistical analysis and interpretation of results. ^{14, 11, 2} It has the potential to improve the quality of clinical research. ² Estimands are recommended and beneficial for all kinds of studies, and especially helpful for phase 3 studies and for study objectives supporting a regulatory decision, above all in confirmatory trials. ⁹

One medical writer holds that it is important for medical writers to understand estimands because they write most of the important documents produced across a drug's lifecycle and because they communicate trial results to professionals and to the public. ¹⁵ Estimands will become increasingly important for medical writers. ²

Why the intention-to-treat default was not enough

ICH E9, issued in 1998, established the intention-to-treat principle as the gold standard: analyse the full analysis set, analyse participants as they were randomised, and follow them up regardless of adherence. ¹⁶

Randomly assigning patients to two parallel groups makes the groups comparable at the start, so that an effect seen at the end can be ascribed to the treatment, but only if the randomisation is maintained, and intercurrent events often prevent that. ^{17, 2, 18} Patients die, stop taking the study medication because of side effects or because they feel they are getting no benefit, or take additional medication that interferes with the efficacy endpoints. ^{17, 2} Different patients experience different events at different times, and because each patient is likely to experience different events depending on the treatment received, their rates and timing may differ between the groups. ^{17, 2} Excluding the patients who experience them may leave groups that are no longer comparable when the results are determined, which is why ICH E9 recommended an intention-to-treat analysis on all randomised patients, or as close to all of them as possible. ^{17, 2, 18} Analysing all patients according to the treatment they were randomised to, and not the treatment they actually received, helps ensure that the treatment effect is not overestimated and its estimates are unbiased, and that statistical tests produce valid results. ^{6, 19}

After ICH E9 was introduced, it was recognised that the intention-to-treat principle has its limitations. ^{16, 20} Trials often lacked clarity between a trial objective and the treatment effect related to it, and limitations were found in cases where patients took rescue medication. ²⁰ The addendum to ICH E9 recognises that the guiding principle has its limitations. ^{17, 18}

The 2019 ICH E9(R1) addendum broadened the principle with the estimand framework, introduced the term intercurrent event, updated sensitivity analysis, and encourages cross-functional collaboration. ²⁰ Its draft, dated 30 August 2017, was released for public consultation at the end of August 2017, and in the months after, organisations and individuals submitted comments and suggestions on it to the European Medicines Agency. ^{1, 2} By late 2018 the ICH E9(R1) Expert Working Group had released training materials that elaborate on the addendum's content and suggest how to implement it, and the final version was expected in 2019. ² Those training materials include guidance on documenting estimands. ¹³

One dataset, two defensible answers

One trial compared a diabetes drug with placebo in type 2 diabetes, and placebo was acceptable because participants who lacked efficacy received rescue medication. ^{21, 22} The endpoint was absolute change in HbA1c from baseline to week 24, and data were collected to the end of the trial regardless of rescue medication use. ²² The sponsor analysed the data by ANCOVA, excluding the data after rescue medication started, and found a treatment effect of 0.66%. ²² An FDA reviewer analysed the same data regardless of rescue medication use and found 0.45%. ^{22, 23} In that trial the endpoint and the intake of rescue medication belong to the estimand, as endpoint and intercurrent event, ANCOVA is the estimator, and the numerical result, the treatment effect, is the estimate. ⁷

In 2011 a US Food and Drug Administration reviewer used the reasoning behind intention to treat to suggest that the most valid way to analyse the data for the new antidiabetic drug dapagliflozin was to use all data, including values from patients taking rescue medication, in the statistical model. ^{6, 19} That incident was a trigger for the estimands debate. ⁶

The two results could be two different estimands that reflect two different questions: what the treatment effect is if no rescue medication had been given, which assesses direct efficacy in an ideal scenario, and what the treatment effect is whether or not rescue medication was used, which is more a real-world scenario. ²³ The sponsor's analysis did not follow the intention-to-treat principle, which includes all the data, and under the addendum it can also be done. ^{22, 24} Different estimands result in different treatment effects, and what matters is which question is asked. ²⁴ Different ways of dealing with an intercurrent event such as rescue medication use represent different estimands and answer different scientific questions. ⁵ Likewise, where some participants need extra blood-pressure treatment, estimating outcomes regardless of that treatment and estimating them under a hypothetical situation without it are different questions. ⁴

What the addendum changed

Under the addendum, discontinuing treatment counts as an intercurrent event and not as missing data, so less data is lost. ^{24, 4}

In many cases the estimand of interest will not fully reflect a strict intention-to-treat analysis. ⁵ A trial's estimand had tended to be arrived at implicitly and haphazardly, as a consequence of choices about data collection and statistical analysis. ^{5, 6} The ICH E9 addendum suggests that the various scientific questions the trial data could be used to address should instead be considered explicitly and up front. ^{6, 5} The questions, and hence the estimands, can then be chosen for being the most meaningful in the clinical context and the most relevant for patients, their doctors, regulators and payers. ^{5, 6} Using estimands allows intercurrent events to be seen as a source of important additional information on the efficacy and safety of an investigational treatment, rather than as a nuisance or complication. ⁶

The attributes of an estimand

An estimand has five attributes. ^{3, 4} Four are already familiar from traditional protocols, and the intercurrent event is the new one. ^{3, 25}

• Target population — not the analysis set: the two are completely different things. ICH E9(R1) was written for phase 3 trials, where the target population can be pictured as the label claim; for earlier trials it can perhaps be pictured as the key inclusion and exclusion criteria for that endpoint. ^{3, 25, 26, 4}
• Treatment regimen — the intervention being tested. ^{25, 4}
• Endpoint — what is being assessed. ^{25, 4}
• Population-level summary — how the data will be summarised. ^{25, 4}
• Intercurrent events — the new essential component, and how they are handled. ^{25, 27, 4}

The draft addendum described four attributes: the population targeted by the scientific question; the variable or endpoint to be obtained for each patient to address it; how intercurrent events are to be accounted for to reflect it; and the population-level summary, such as a mean or median, on which the treatments are compared. ^{5, 6}

One example of a primary estimand states all five attributes: the mean difference (population-level summary) comparing drug with placebo (treatment) in absolute change in HbA1c from baseline to week 24 (endpoint) as if no emergency medication had been used (intercurrent-event strategy) in patients with type 2 diabetes (target population). ²⁸

Estimands could also be defined for safety questions, such as how long patients are able to remain on treatment before discontinuing because of adverse events. ^{5, 6, 29}

Intercurrent events

An intercurrent event is an event occurring after treatment initiation that affects the interpretation or the existence of a measurement associated with the clinical question of interest. ^{30, 4} It falls after treatment initiation and before the endpoint is observed. ^{30, 27} Intercurrent events are also called post-randomisation events, which has been suggested as the better term. ²

Intercurrent events can include use of additional or alternative medication; use of concomitant rescue medication or medication for acute symptoms; withdrawal or discontinuation of the assigned treatment for toxicity or lack of efficacy; switching from one active treatment to another; death, which is always an intercurrent event; and clinical events that define a principal stratum. ²⁷

Such events are common in clinical practice, are not a drawback, and do not cause missing data. ³⁰ An intercurrent event is not automatically missing data: a participant can stop treatment and still attend every assessment. ⁴ Intercurrent events must be identified before the trial starts. ³⁰

The five handling strategies

ICH E9(R1) defines five strategies, of which treatment policy, hypothetical and composite are the most common. ^{27, 31} Different strategies lead to different treatment effects. ³¹

One illustration is a randomised trial measuring an endpoint from baseline to month two, in which one participant takes rescue medication before the month-two measurement. ^{31, 32} Another compares a drug against placebo, both on top of background therapy, in patients with type 2 diabetes, measuring the reduction in glycated haemoglobin from study start to week 26. ^{33, 19} Rescue medication is standard in such trials: patients whose blood glucose exceeds a predefined limit are allowed additional antidiabetic medication alongside the study treatment, because high blood glucose raises the risk of complications such as cardiovascular problems or damage to the nerves, kidneys and eyes, and it would not be ethical to require patients to continue in the trial with excessive levels. ^{33, 19}

Three estimands for the 26-week trial deal with rescue medication in different ways, although the draft addendum defines five types. ^{5, 29} All three share the same target population, the one in which the drug is planned to be used after approval, as reflected by the trial's inclusion and exclusion criteria. ^{5, 29} They also share the same population-level summary, the difference between the groups in mean change in glycated haemoglobin, or the mean difference in glycated haemoglobin values. ^{5, 29} They differ in how the variable is defined and how the rescue medication is handled. ^{5, 29} Many different estimands are possible, and the situation would grow more complicated still if deaths and discontinuations for adverse events were counted as intercurrent events alongside rescue medication. ^{5, 6}

Treatment policy

In the month-two trial, the treatment policy strategy uses the month-two value for the analysis, no matter when the rescue medication was taken. ³² This is the real-world strategy. ³²

In the 26-week trial it is the change in glycated haemoglobin from baseline to week 26 with all data considered regardless of rescue medication. ^{34, 29} It uses the week-26 values from every patient, including those who started rescue medication, and it is the estimand that corresponds to an intention-to-treat analysis. ^{35, 29, 19} If the drug lowers glycated haemoglobin, rescue use can be expected to be more frequent in the placebo group. ^{19, 29} The estimated effect is then the difference between the drug's effect, occasionally with rescue added, and the average effect in a placebo group where many patients are taking a rescue medication known to work. ^{19, 29} The rescue medication blurs the drug's effect, which may give a modest difference that underestimates the true difference between drug and placebo. ^{19, 29} In exchange the result reflects clinical practice, since it is very likely that some patients in clinical practice will need additional medication whatever they are taking. ^{19, 29} This estimand is likely to be of particular importance to payers and reimbursement agencies who want to know the drug's real-world effectiveness, and it is sometimes called an effectiveness estimand or an effectiveness analysis. ^{19, 29}

Composite

The composite strategy builds the event into the endpoint definition, assigning a specified value to participants who start rescue medication after treatment initiation. ³² The protocol must describe this, and the intercurrent event should appear in the endpoint definition itself. ³²

Hypothetical

The hypothetical strategy asks what would have happened had the event not occurred: use the data up to the intercurrent event, then impute the remainder to obtain the data the participant would have had without the rescue medication. ³⁶

In the 26-week trial it is the change to week 26 with data after rescue modelled as if no patient had taken rescue medication before week 26, an estimate of the effect that would be seen if no patient took it. ^{35, 29, 37} Only values from patients who have not yet started rescue enter the analysis, values are counted as missing from the point rescue begins, and an appropriate method for handling missing data through statistical modelling, such as multiple imputation or mixed models for repeated measures, has to be used. ^{37, 35} The estimate reflects both what actually happened in the patients who reached week 26 without rescue and what the data collected before rescue suggest might have happened in the rest. ^{37, 35} It is hypothetical at the level of a group of patients and relies heavily on modelling for a large proportion of them, so it will never fully reflect a real-life situation. ^{37, 35} Its usefulness has been contested, particularly by health technology assessment agencies responsible for assessing a treatment's value in actual clinical practice. ³⁷ The ICH E9(R1) Expert Working Group advises that hypothetical estimands be based on clinically reasonable situations clearly specified in the protocol. ³⁷

While on treatment

In the month-two trial, the while-on-treatment strategy imputes the month-two value with the last observation taken while the participant was still on treatment, before the intercurrent event, which is also called last observation carried forward. ³⁶ It is a special case of the hypothetical strategy, and one of the five strategies the addendum defines. ^{36, 31} Its estimand is also called the on-treatment estimand, because it estimates the response to treatment before the intercurrent event occurs. ³⁷

In the 26-week trial it is the change to week 26, or to the last value before rescue medication started, with data after rescue not considered. ^{29, 38, 35} If the drug works, few patients in the drug group will need rescue, and those who do are likely to need it late. ^{29, 35} In the placebo group many will need it, and are likely to need it soon after study start, because the background medication will not control blood sugar effectively. ^{29, 35} Taking the last recorded value before rescue therefore uses values for many patients, particularly placebo patients, at a time when glycated haemoglobin is likely to be high, so this estimand tends to overestimate the drug's effect. ^{29, 35} In effect it corresponds to a last-observation-carried-forward analysis, which has been widely used to deal with rescue medication in type 2 diabetes trials but is problematic because the estimates it produces are biased. ^{37, 35} The result is also hard to interpret, because it compares data from widely different time points in the two groups. ³⁷ A while-on-treatment estimand may nonetheless be the most clinically meaningful one in some settings, such as evaluating palliative treatments in end-of-life care, where death is an expected post-randomisation event. ³⁷

The on-treatment and hypothetical estimands attempt in different ways to capture the effect of the drug itself without blurring it by the use of rescue medication. ^{35, 29} With appropriate modelling, the hypothetical estimand is likely to give a less biased answer than the on-treatment estimand to the question that is crucial to an individual patient: if I take this drug as part of my treatment regimen without adding any further drugs, what effect can I expect after 26 weeks? ^{37, 35}

Principal stratum

The principal stratum strategy restricts the analysis to participants who never took rescue medication, excluding anyone who took it at any time. ³⁹

Choosing a strategy for each event

The handling of an intercurrent event depends on the event itself and its clinical importance, and the things to consider include the nature of the medication, the study, and clinical practice in that context. ³⁹ For example, a short course of an oral corticosteroid affects a year-long trial of a systemic drug less than it affects a shorter trial of a localised treatment. ³⁹ For concomitant medication, the right strategy depends on whether the medication is standard of care, where treatment policy would for example be used, and composite would for example be used where the event amounts to treatment failure. ^{39, 28} The disease setting and the aim of treatment will also affect which estimands are chosen. ⁶

An estimand worded "as if no emergency medication had been used" is hypothetical, not treatment policy, because it pretends that the patients take no rescue medication. ^{28, 8}

The data a trial collects

Rescue medication complicates the evaluation of the treatment effect from a scientific point of view, because it raises whether efficacy measurements continue once a patient starts it and whether those measurements are included when the treatment effect is calculated. ^{19, 33, 18} The consequences of that trial design decision for interpreting the results had not always been considered at the design stage. ¹⁸ If the trial is planned so that no data are collected once rescue medication starts, because those patients are withdrawn for instance, few patients may have data at week 26. ¹⁸ The options at the analysis stage are then limited, and a full intention-to-treat analysis becomes impossible. ¹⁸ If data from all patients are collected and used, the measured values reflect both the study treatment and the rescue medication, so the comparison becomes drug plus rescue against placebo plus rescue. ¹⁸ That approach helps ensure statistical validity, but the estimate may not be clinically meaningful, because the drug's effect is blurred by the rescue medication. ¹⁸ The trade-off has been described as the choice between a precise answer to a less relevant question and an approximate answer to the most relevant question, and the estimands discussion makes clear that it can be made in a variety of ways. ^{18, 33} Depending on which option for collecting and analysing the data is chosen, the precise definition of the treatment effect, the estimand, will differ. ¹⁸

Once collection stops at rescue, or the later data are ignored, a large proportion of the data may be missing, and there are various options for the patients who started rescue: substitute the last value before rescue for the week-26 value, as last observation carried forward; exclude them and run a completers analysis on the patients with complete observations and no rescue use; or model the results likely at week 26 had no patient used rescue from the data all patients provided before rescue, with mixed models for repeated measures or multiple imputation. ¹⁹ Compared with the treatment policy estimand, these approaches have the potential to give better estimates of the drug's own efficacy, which may be of particular interest to a patient and the treating physician. ¹⁹

For many estimands relevant to patients and physicians, reasons for treatment discontinuation will need to be recorded more rigorously than has tended to be done, and captured in the study protocol and the case report form with greater granularity. ^{13, 37} Broad categories such as "lost to follow-up" or "investigator decision" do not permit important estimands to be calculated, and questions about how likely a patient is to discontinue treatment because of tolerability issues or because of lack of efficacy can only be answered if detailed reasons are captured. ¹³

Who chooses the estimand, and in what order

Choosing the appropriate estimand for a given trial objective is primarily a medical and clinical question and not a statistical one, and some prominent statisticians go so far as to proclaim that estimands are not a statistical topic at all. ^{13, 40} Discussion between medical and statistical experts is still necessary, so that the estimands chosen reflect questions of clinical interest and can also be estimated statistically. ^{13, 40} The concept is not really a statistical idea but pertains more generally to the medical evaluation of trial results. ^{1, 2}

Because the addendum revised a statistical guideline and not ICH E8 on general considerations for clinical studies, it risked being perceived as a statistical topic only, even though it describes a goal shared by statisticians, clinicians and others involved in a trial. ⁴¹ Specialists anticipated that getting all researchers to view the specification of estimands as a shared goal would be challenging, which is why an estimand implementation working group launched a free estimand training academy. ³

In many settings a single estimand is unlikely to meet the different needs of all stakeholders. ^{5, 6} It has even been proposed that the most helpful way to give physicians and patients what they need would be to include a lay description of several estimands in the prescribing information. ⁶

The paradigm shift is a different sequence of activities. ^{13, 35} Researchers first need to think about the objectives, what the trial is meant to show. ^{13, 35} They then need to consider the precise scientific questions of interest and choose the estimands that answer them, a choice that should be made in discussion with regulatory authorities and in accordance with available guidance so that key stakeholders' needs are met. ^{13, 35} In 2018, indication-specific guidelines on the appropriate use of estimands were likely to become available in the future. ¹³ Once the estimands have been defined, the trial can be designed so that all the necessary data are collected, and the statistical analysis methods chosen to address those estimands. ^{13, 40} The study objectives and estimands should be defined first, to guide the study design and the analysis. ⁴² In the past researchers often tried to work out what exactly they had evaluated after a study had been completed, when very little can be done: once the decision not to collect data after rescue medication starts has been taken, it cannot be reversed after the trial is completed. ^{13, 35} The proposed sequence runs from trial objectives to the scientific questions of interest, then to estimands, then to trial design, then to analysis methods. ^{12, 40, 13} In good clinical research, researchers always started planning a trial by defining its objectives and choosing endpoints and a statistical methodology, but the potential influence of intercurrent events on the interpretation of the endpoints was rarely considered. ^{13, 40} Estimands close the gap between the trial objectives and the main estimates by clarifying exactly how intercurrent events will be considered, or how the interpretation changes when those events are considered in different ways. ^{13, 40} Depending on how such events are accounted for, a trial may be estimating the effect of the study drug itself or evaluating a treatment policy. ^{13, 35}

How far the framework has spread

By 2025 more and more agencies had implemented the addendum, Japan in 2024, and working groups such as an estimand implementation working group address the practical implementation of the framework. ²¹ By 2025 estimands had developed over the previous five or six years and were still developing, in something like a transition phase. ¹⁶

By September 2019 the European Medicines Agency's Scientific Advice Working Party had released a draft qualification opinion on clinically interpretable treatment effect measures based on recurrent event endpoints that allow for efficient statistical analyses, which is relevant to developing estimands, that is, clinically interpretable treatment effect measures. ⁴³ Medicine development was then becoming increasingly estimand-based, and writers were urged to share that opinion with their statistical colleagues. ⁴³

All the trials of the PIONEER programme were run with the estimand framework. ¹¹