Trial Estimands, Intercurrent Events & Missing Data

“What was the treatment effect?” sounds like a single question. In many trials it is several possible questions, depending on what happens after treatment begins. Participants may discontinue the assigned intervention, use rescue treatment, switch therapies, undergo a procedure, or die. Some outcome measurements may also be missing.

The estimand framework makes the intended question explicit. FDA’s final ICH E9(R1) addendum connects the objective, design, data collection, analysis, and interpretation.[1]

An estimand defines the effect being estimated

An estimand precisely describes the treatment effect corresponding to a trial objective. It is neither a statistical model nor the numerical result; it defines the target for estimation.

ICH E9(R1) and current SPIRIT materials describe five linked attributes:[1][2]

Treatment condition or conditions: the assigned regimens or treatment strategies whose effects are being compared.
Population: the target population for the clinical question, defined by relevant participant characteristics rather than by whoever happened to enter one analysis.
Variable or endpoint: the participant-level outcome, including its definition and timing.
Intercurrent-event strategy: how each event after treatment initiation is reflected in the treatment-effect question.
Population-level summary: how participant-level outcomes are summarized and contrasted across treatment conditions, such as a difference in means, risk ratio, or time-to-event summary.

Changing an attribute can change the treatment-effect question. Different analyses may target different estimands, or they may use different estimators for the same estimand; those situations should not be treated as interchangeable.

Intercurrent events are not automatically missing data

An intercurrent event occurs after treatment starts and affects either the interpretation or existence of the measurements needed for the treatment-effect question. Examples include discontinuing assigned treatment, switching to another intervention, receiving rescue medication, or death before an outcome can be measured.

Missing data are observations that would be meaningful for analysis of the chosen estimand but were not collected or are unavailable. Treatment discontinuation is an intercurrent event; it is not itself missing data. Stopping follow-up can create missing data when later measurements remain relevant to the estimand.[1]

The role of a post-event measurement depends on the question. Under a treatment-policy strategy, a post-rescue measurement may directly contribute to the target comparison. Under a hypothetical no-rescue strategy, that measurement does not directly represent the target hypothetical outcome, although it may still inform the estimator or model. An uncollected measurement can therefore be missing for one estimand but not required for another.

Five broad strategies for intercurrent events

ICH E9(R1) describes strategies that can be used alone or in combination across different events:[1]

Strategy	Question represented in plain language
Treatment policy	How do the assigned treatment conditions compare when relevant outcomes are used regardless of whether the intercurrent event occurs?
Hypothetical	How would the treatment conditions compare in a defined scenario in which the event did not occur?
Composite	How do the treatment conditions compare when occurrence of the event is incorporated into the outcome definition?
While on treatment	How do outcomes compare up to occurrence of the intercurrent event, with the observation period for the question ending at that event?
Principal stratum	What is the effect in a subgroup defined by the event’s potential occurrence under the treatment conditions being compared?

A treatment-policy strategy does not mean that the event is ignored, and it does not solve missing data; relevant post-event outcomes still need to be collected. It also cannot manufacture outcomes that cannot exist, such as a post-death symptom score, so terminal events require a clinically coherent outcome definition and strategy. A principal stratum is not simply the observed group of adherers or completers: membership depends on potential event status under treatment conditions and generally requires untestable assumptions. A while-on-treatment strategy is likewise not an estimate of the effect “among adherers”; its interpretation can depend on why and when the event occurred.[1]

No strategy is universally superior. The choice depends on the scientific question and whether it is meaningful and estimable. One trial may use different strategies for different events.

The labels do not remove assumptions. Hypothetical and principal-stratum strategies can require strong assumptions about unobserved scenarios or subgroup membership. Composite outcomes must remain interpretable when components differ in importance.

Estimand, estimator, and estimate

These terms are easy to blur:

The estimand is the target treatment effect.
The estimator is the statistical method used to target it.
The estimate is the numerical result produced from the observed data.

Alignment matters. A sophisticated model cannot rescue an analysis targeting a different question. FDA says the main estimator should align with the estimand, while sensitivity analyses test robustness to assumptions for that estimand.[1]

Why “intention to treat” is not a complete description

Analyzing participants according to randomized assignment helps preserve the randomized comparison. It does not by itself define the estimand, establish a treatment-policy strategy, specify which post-event observations are relevant, or solve missing data.[1][4]

Excluding randomized participants can undermine the benefits of randomization. Excluding observations may instead be required by the chosen estimand, but arbitrary or data-dependent exclusions can change the target question or introduce bias. Reports should therefore identify separately who was included, which observations were used, and why.[4]

A clear report should state:

which randomized participants entered the analysis;
which post-randomization observations were used;
how each intercurrent event was handled;
why the chosen analysis corresponds to the estimand; and
what missing-data assumptions were made.

Missing data are an assumption problem

When relevant outcomes are missing, the unobserved values cannot be recovered from the dataset alone. Complete-case analysis, likelihood methods, weighting, and multiple imputation all rely on assumptions. Planning for a larger sample can address some loss of power, but bias may still be introduced when loss to follow-up is related to response.[3][5]

The National Research Council report emphasizes preventing missing data where possible, clearly stating assumptions, and examining how conclusions change under alternative plausible assumptions.[5] CONSORT 2025 similarly asks reports to describe the amount and reasons for missing data by group, the primary method used, and sensitivity analyses.[3]

Common shorthand includes:

Missing completely at random: missingness is unrelated to observed or unobserved data under the stated model.
Missing at random: after conditioning on observed information, missingness does not depend on the unseen value.
Missing not at random: missingness may still depend on the unseen value after accounting for observed information.

These are model assumptions, not diagnoses. “Missing at random” does not mean values disappeared haphazardly, and it requires a context-specific rationale.[7]

Why single imputation can be misleading

Replacing every missing outcome with one value and then analyzing the completed dataset as though those values were observed fails to propagate imputation uncertainty and can make the result look more precise than it is. Last observation carried forward additionally assumes that the last observed value remains an appropriate value for later unobserved times, an assumption that may be scientifically implausible. CONSORT and the National Research Council caution against such simple primary approaches unless their assumptions are scientifically justified.[3][5]

Multiple imputation or model-based methods can better represent uncertainty in suitable settings, but they are not assumption-free. Readers need the variables included, imputation and analysis models, analysis groups, number of imputations when applicable, compatibility with the estimand, handling of intercurrent events, and rationale for the assumed missing-data mechanism.[3][7]

Sensitivity analysis versus supplementary analysis

A sensitivity analysis examines how robust the inference for the same estimand is to departures from assumptions underlying the main estimator. For example, it may examine progressively less favorable outcomes among participants with missing values.[1][3]

A supplementary analysis may answer a different question or target another estimand. FDA’s E9(R1) addendum distinguishes these roles.[1] Calling every alternative model a sensitivity analysis can hide the fact that the scientific question changed.

Agreement across scientifically credible sensitivity analyses can show that the inference is not sensitive to the particular assumptions examined; it does not validate those assumptions or eliminate bias shared by all analyses. Material divergence under plausible alternatives indicates assumption sensitivity and should be reported and explained rather than discarded.[3]

A reading framework for trial reports

Write the trial objective in a full sentence.
Identify the population, treatments, endpoint, time point, and summary measure.
List each important event after treatment initiation that affects the interpretation or existence of the measurements.
Determine which strategy the estimand uses for each event.
Identify which measurements directly address the estimand, which remain informative to its estimator, and which relevant measurements are unavailable.
Define the analysis population, randomized-group handling, and observations included.
Record the main estimator and its missing-data and model assumptions.
Separate sensitivity analyses for the same estimand from supplementary analyses that answer another question.
Check the number available for every outcome and time point; CONSORT recommends reporting both the number analyzed and the number with observed data.[6]

For adjacent interpretation steps, use the confidence intervals, p values, and multiplicity guide, the research-study reading framework, and the negative and null-results guide.

Important limitations

An estimand clarifies the question; it does not guarantee unbiased estimation.
Different estimands can be legitimate, but their estimates should not be pooled or compared as if they answer the same question.
Intercurrent events and missing data are related but not interchangeable.
“Intention to treat” is incomplete without the analysis population, observations used, and handling of post-randomization events.
Missing-data methods depend on assumptions that may not be verifiable from observed data.
Sensitivity analyses should probe the main assumptions, not merely repeat the preferred model with cosmetic changes.
A treatment-policy strategy does not remove the need to collect relevant outcomes after treatment discontinuation or another intercurrent event.
A principal stratum is defined through potential event status and is not automatically the observed subgroup of adherers, completers, or survivors.
Agreement across several sensitivity analyses does not prove their shared assumptions or rule out shared bias.
Trial results remain specific to the studied population, interventions, endpoint, follow-up, and material.

Frequently asked questions

Is an estimand the same as an endpoint?

No. The endpoint is one attribute. An estimand also specifies the population, treatment conditions, handling of intercurrent events, and population-level summary.

Is treatment discontinuation always missing data?

No. Treatment discontinuation is ordinarily handled as an intercurrent event when it affects the treatment-effect question; it is not missing data by itself. Missing data arise when measurements that remain relevant to the chosen estimand are unavailable.

Can a trial have more than one estimand?

Yes. Different objectives or outcomes may require different estimands. Each should be labeled so readers know which question an estimate answers.

Does intention-to-treat solve every post-randomization problem?

No. Retaining randomized groups is important, but the report must still explain which measurements were collected and how discontinuation, switching, rescue treatment, death, and missing values were handled.

What does “missing at random” mean?

It means that, after accounting for specified observed data, the probability of missingness does not depend on the unobserved value. It is an assumption requiring justification, not a synonym for accidental loss.

Why is last observation carried forward often criticized?

It assumes the earlier value remains appropriate at later times. When a singly imputed dataset is analyzed as though imputed values were observed, the analysis also fails to propagate imputation uncertainty.

What should a sensitivity analysis tell me?

It should show whether the conclusion for the same estimand remains stable under credible departures from the main analysis assumptions.

Primary and standards-based sources

Review and corrections

Sources checked: September 1, 2026. Questions or corrections can be sent through the Serum Research contact page.

Editorial responsibility: Serum Research prepared and maintains this educational page. No independent scientific, statistical, regulatory, or clinical review is claimed. See how Serum Research reviews evidence or return to the Research Peptide and Compound Library.

Are you over 21?

Trial Estimands, Intercurrent Events, and Missing Data