Why Science? | ICH E9(R1) Estimands, Intercurrent Events and Sensitivity Analysis

E9(R1) uses an estimand to describe the treatment effect targeted by the clinical question. It is distinct from the estimator, which is the analytical method, and the estimate, which is the numerical result. Keeping the three separate prevents a model from quietly defining the question after data are collected.

For an estimand is the treatment effect of interest, one extra word can change the scientific question. E9(R1) uses an estimand to describe the treatment effect targeted by the clinical question. It is distinct from the estimator, which is the analytical method, and the estimate, which is the numerical result. Keeping the three separate prevents a model from quietly defining the question after data are collected. Keep this boundary beside the estimate: A precise calculation can answer the wrong question when the estimand was not defined first.

A strong science note separates observation from interpretation. Write one sentence each for the estimand, estimator and estimate in a hypothetical trial. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Next chapter

FDA identifies E9(R1) as final Level 1 guidance dated May 2021 under docket FDA-2017-D-6113, issued by CDER and CBER. It supplements the 1998 E9 Statistical Principles guideline and strengthens dialogue among clinicians, statisticians, regulators and patients.

An estimand makes the may 2021 addendum is final level 1 explicit before the model is run. FDA identifies E9(R1) as final Level 1 guidance dated May 2021 under docket FDA-2017-D-6113, issued by CDER and CBER. It supplements the 1998 E9 Statistical Principles guideline and strengthens dialogue among clinicians, statisticians, regulators and patients. Preserve this guardrail: The addendum focuses on estimands and sensitivity analysis rather than replacing the broader E9 framework.

A strong science note separates observation from interpretation. Create a source receipt with revision, status, date, docket, issuing centres and relationship to E9. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The treatment condition of interest describes the interventions being compared, such as investigational drug versus placebo or standard care. It should include relevant features of treatment and use so that the comparison matches the clinical decision.

To understand the treatment attribute defines the comparison, separate the desired effect from the calculation. The treatment condition of interest describes the interventions being compared, such as investigational drug versus placebo or standard care. It should include relevant features of treatment and use so that the comparison matches the clinical decision. The conclusion stops here: A treatment label can hide important differences in dose, duration, adherence support or accompanying care.

A strong science note separates observation from interpretation. Specify both treatment conditions and list every co-intervention needed to interpret the comparison. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The target population is described through clinical and demographic characteristics relevant to the question. Trial eligibility and recruitment should allow the observed sample to represent that population well enough for inference.

For the population attribute defines who the effect concerns, one extra word can change the scientific question. The target population is described through clinical and demographic characteristics relevant to the question. Trial eligibility and recruitment should allow the observed sample to represent that population well enough for inference. Keep this boundary beside the estimate: The analysed participants are not automatically identical to the target population.

A strong science note separates observation from interpretation. Draw the path from target population to eligible, randomised, followed and analysed participants. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The variable may be a measurement at a time point, change from baseline, event occurrence, time to event or another participant-level outcome. Its definition should include method, timing and how events such as death affect meaning.

An estimand makes the variable attribute defines what is measured explicit before the model is run. The variable may be a measurement at a time point, change from baseline, event occurrence, time to event or another participant-level outcome. Its definition should include method, timing and how events such as death affect meaning. Preserve this guardrail: A familiar endpoint name can hide different timing, scoring or event rules.

A strong science note separates observation from interpretation. Build a variable card with concept, measurement, time, scale, range and interpretation. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The population-level summary may be a mean difference, odds ratio, hazard ratio, risk difference or another measure suited to the variable and question. The summary should be interpretable for the clinical decision and not chosen solely because a model is convenient.

To understand the summary attribute combines individual outcomes, separate the desired effect from the calculation. The population-level summary may be a mean difference, odds ratio, hazard ratio, risk difference or another measure suited to the variable and question. The summary should be interpretable for the clinical decision and not chosen solely because a model is convenient. The conclusion stops here: Different summaries can emphasise different aspects of the same outcome distribution.

A strong science note separates observation from interpretation. Compare two summaries for the same variable and explain what each tells a patient or clinician. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

An intercurrent event is an event after treatment initiation that affects either the interpretation or existence of measurements associated with the clinical question. Examples include treatment discontinuation, rescue medication, switching therapy, terminal events or changes in background treatment.

For intercurrent events occur after treatment starts, one extra word can change the scientific question. An intercurrent event is an event after treatment initiation that affects either the interpretation or existence of measurements associated with the clinical question. Examples include treatment discontinuation, rescue medication, switching therapy, terminal events or changes in background treatment. Keep this boundary beside the estimate: An intercurrent event is not the same as a missing measurement, although it can cause missing data.

A strong science note separates observation from interpretation. List possible post-randomisation events and separate their effect on meaning from their effect on data availability. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A treatment-policy strategy addresses the effect of treatment regardless of whether the intercurrent event occurs, resembling an intention-to-treat style question for that event. It may fit decisions about starting a treatment in real practice where discontinuation or rescue remains part of experience.

An estimand makes treatment-policy strategy uses outcomes regardless of the event explicit before the model is run. A treatment-policy strategy addresses the effect of treatment regardless of whether the intercurrent event occurs, resembling an intention-to-treat style question for that event. It may fit decisions about starting a treatment in real practice where discontinuation or rescue remains part of experience. Preserve this guardrail: The strategy still needs adequate outcome collection after the event; it does not make missing data disappear.

A strong science note separates observation from interpretation. State which outcomes must continue after treatment discontinuation and how follow-up will be preserved. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A hypothetical strategy targets the outcome under a scenario in which a specified intercurrent event would not happen, such as no rescue medication. It requires assumptions and analytical methods that estimate unobserved counterfactual outcomes.

To understand hypothetical strategy imagines the event did not occur, separate the desired effect from the calculation. A hypothetical strategy targets the outcome under a scenario in which a specified intercurrent event would not happen, such as no rescue medication. It requires assumptions and analytical methods that estimate unobserved counterfactual outcomes. The conclusion stops here: The answer can be highly sensitive to unverifiable assumptions about what would have happened.

A strong science note separates observation from interpretation. Describe the hypothetical world and list evidence supporting each assumption used to predict it. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A composite-variable strategy incorporates the intercurrent event into the endpoint, for example defining treatment failure when rescue medication is used. This can align the endpoint with clinical importance when the event itself is meaningful.

For composite strategy makes the event part of the outcome, one extra word can change the scientific question. A composite-variable strategy incorporates the intercurrent event into the endpoint, for example defining treatment failure when rescue medication is used. This can align the endpoint with clinical importance when the event itself is meaningful. Keep this boundary beside the estimate: Coding unlike events into one result can hide why participants failed or how severe outcomes differed.

A strong science note separates observation from interpretation. Design a composite rule and report each component separately beside the combined result. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A while-on-treatment strategy considers response before an intercurrent event, such as discontinuation. It can address effects while treatment is being taken, but follow-up duration may differ and the occurrence of the event can relate to prognosis.

An estimand makes while-on-treatment strategy limits the relevant period explicit before the model is run. A while-on-treatment strategy considers response before an intercurrent event, such as discontinuation. It can address effects while treatment is being taken, but follow-up duration may differ and the occurrence of the event can relate to prognosis. Preserve this guardrail: The estimate may not answer what happens over the full intended treatment period.

A strong science note separates observation from interpretation. Define the observation window and show how early discontinuation changes the amount of information. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A principal-stratum strategy concerns participants who would experience, or not experience, an intercurrent event under specified treatment conditions. Because membership may depend on unobservable potential outcomes, estimation can require strong assumptions.

To understand principal-stratum strategy targets a post-baseline subgroup, separate the desired effect from the calculation. A principal-stratum strategy concerns participants who would experience, or not experience, an intercurrent event under specified treatment conditions. Because membership may depend on unobservable potential outcomes, estimation can require strong assumptions. The conclusion stops here: A post-randomisation subgroup observed under one treatment is not automatically a valid principal stratum.

A strong science note separates observation from interpretation. Write the stratum definition using potential outcomes and identify which membership information is unobserved. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

One trial can use different strategies for different intercurrent events because rescue medication, discontinuation, switching and death change interpretation in different ways. The protocol should justify each strategy and ensure the combined estimand remains coherent.

For different events may need different strategies, one extra word can change the scientific question. One trial can use different strategies for different intercurrent events because rescue medication, discontinuation, switching and death change interpretation in different ways. The protocol should justify each strategy and ensure the combined estimand remains coherent. Keep this boundary beside the estimate: Choosing whichever strategy gives the most favourable result after seeing data undermines interpretability.

A strong science note separates observation from interpretation. Create an event-strategy table before unblinding and give a clinical reason for every pairing. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Statistical methods, data collection and assumptions should target the defined estimand. An estimator may use regression, survival analysis, imputation or other techniques, but its properties must be understood under the planned intercurrent-event strategy and missing-data pattern.

An estimand makes the estimator should align with every estimand attribute explicit before the model is run. Statistical methods, data collection and assumptions should target the defined estimand. An estimator may use regression, survival analysis, imputation or other techniques, but its properties must be understood under the planned intercurrent-event strategy and missing-data pattern. Preserve this guardrail: A standard software command does not prove alignment with the clinical question.

A strong science note separates observation from interpretation. Trace each estimand attribute to a dataset field, model term, assumption and output. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Missing data concern unavailable measurements, whereas intercurrent events concern interpretation or existence of measurements. The analysis should prevent missingness where possible, document reasons and use assumptions appropriate to the estimand.

To understand missing data require a separate plan, separate the desired effect from the calculation. Missing data concern unavailable measurements, whereas intercurrent events concern interpretation or existence of measurements. The analysis should prevent missingness where possible, document reasons and use assumptions appropriate to the estimand. The conclusion stops here: Labeling all missing values the same can mix administrative loss, illness, withdrawal and event-related absence.

A strong science note separates observation from interpretation. Classify missingness reasons and show which are compatible with the primary model assumptions. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A sensitivity analysis explores robustness of inference to deviations from assumptions used in the main estimator while targeting the same estimand. It should focus on assumptions capable of materially changing the conclusion and present a credible range rather than cosmetic variations.

For sensitivity analysis tests robustness for the same estimand, one extra word can change the scientific question. A sensitivity analysis explores robustness of inference to deviations from assumptions used in the main estimator while targeting the same estimand. It should focus on assumptions capable of materially changing the conclusion and present a credible range rather than cosmetic variations. Keep this boundary beside the estimate: A sensitivity analysis that changes the estimand answers a different question instead of testing robustness.

A strong science note separates observation from interpretation. Name the main unverifiable assumption, vary it across plausible values and record where the conclusion changes. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The receipt records objective, treatment conditions, population, variable, each intercurrent event and strategy, population summary, data collection, estimator, assumptions, missing-data handling, sensitivity analyses, estimates and limits. It lets readers see exactly which treatment effect was learned.

An estimand makes a complete estimand receipt prevents interpretive drift explicit before the model is run. The receipt records objective, treatment conditions, population, variable, each intercurrent event and strategy, population summary, data collection, estimator, assumptions, missing-data handling, sensitivity analyses, estimates and limits. It lets readers see exactly which treatment effect was learned. Preserve this guardrail: A complete estimand cannot make an unimportant outcome clinically meaningful.

A strong science note separates observation from interpretation. Give the receipt to a clinician and a statistician and ask whether both describe the same treatment effect. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter