ICH E9(R1) Estimands and Sensitivity Analysis

What it is

— ICH E9(R1), Addendum on Estimands and Sensitivity Analysis to the Guideline on Statistical Principles for Clinical Trials, is a harmonised regulatory guidance adopted at ICH Step 4 in November 2019. It is maintained by the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH) and implemented by member regulators including FDA and EMA (EMA/CHMP/ICH/436221/2017). It is not a reporting checklist and not a critical-appraisal tool; it is a thinking framework that forces a study to state, before any analysis, exactly which treatment effect ("estimand") it is targeting. An estimand is defined by five attributes: (1) the population (target patients), (2) the treatment condition(s) being compared, (3) the endpoint (variable), (4) how intercurrent events (ICEs — events after treatment initiation that alter the existence or interpretation of the outcome, such as treatment discontinuation, switching, rescue medication, or death) are handled, and (5) the population-level summary (e.g., difference in means, risk ratio, hazard ratio). The addendum specifies five strategies for ICEs — treatment-policy, hypothetical, composite, while-on-treatment, and principal-stratum — and requires that the chosen estimator and any sensitivity analyses be aligned to the named estimand rather than chosen by default.

When to use

— Apply ICH E9(R1) whenever a study estimates a treatment effect and that effect must be defended to a regulator, an HTA body, or a critical journal — i.e., the design and analysis stage of pragmatic trials, target-trial emulations, comparative-effectiveness and safety studies in routinely collected data, and post-authorisation safety/efficacy studies (PASS/PAES). It is the governing framework for FDA and EMA submissions where a primary or secondary treatment effect is claimed, and increasingly for HTA dossiers where the relevant decision question (e.g., effectiveness under real-world adherence vs. efficacy under perfect adherence) hinges on how intercurrent events are treated. Decision rule for which tool applies: use ICH E9(R1) to define the question and the effect (what are we estimating, and under what handling of discontinuation/switching/death?); use a reporting guideline such as STROBE/RECORD-PE or a protocol template such as HARPER/STaRT-RWE to document the completed study; use ROBINS-I to appraise its risk of bias. These are complementary, not interchangeable. In RWE specifically, E9(R1) is most powerful when paired with target-trial emulation, which gives the otherwise-abstract intercurrent-event strategies concrete operational meaning (e.g., a treatment-policy estimand maps to intention-to-treat from time zero; a hypothetical "no switching" estimand maps to per-protocol with censoring and inverse-probability-of-censoring weighting).

What it requires

— The addendum requires that a study, at the protocol stage, (i) name the target population and align eligibility and time zero to it (in RWE, the new-user/active-comparator structure and an immortal-time-free index date); (ii) specify the treatment conditions as explicit strategies, including duration, switching, and concomitant therapy; (iii) define the endpoint with its measurement and assessment window; (iv) enumerate the intercurrent events that can occur and assign each an explicit strategy (treatment-policy, hypothetical, composite, while-on-treatment, or principal-stratum) — the analytic heart of the framework, because the same data can yield very different effects depending on how discontinuation, switching, rescue, and especially death are handled; (v) state the population-level summary; and (vi) pre-specify sensitivity analyses that vary the assumptions specific to the chosen estimand (e.g., missing-data mechanisms under a hypothetical strategy, grace periods and censoring rules under a per-protocol estimand) and are distinguished from supplementary analyses that target a different estimand. In real-world data the framework forces several disciplines the catalog implements concretely: data fitness-for-use must support the chosen ICE strategy (e.g., reliable death capture for a composite or while-on-treatment estimand), phenotype/outcome algorithms must be validated, time-zero must be aligned to the population and treatment definitions, attrition and informative censoring must be modeled rather than ignored, and confounding control must target the named estimand (ATT vs. ATE) rather than whatever the default estimator produces.

When NOT to use — limitations and common misapplications

— The dominant failure mode is estimand-as-documentation: writing a five-attribute table to satisfy a template while the analysis is unchanged and the intercurrent-event strategy is never actually operationalized. A second is defaulting to a single treatment-policy estimand for every question without asking whether discontinuation and switching are part of the effect of interest or a nuisance to be removed — the choice is substantive, not administrative, and the wrong default can answer a question no decision-maker asked. A third is confusing the estimand with the estimator: E9(R1) defines what is being estimated; it does not endorse any particular model, and a beautifully specified estimand fitted with a biased estimator (e.g., adjusting for a post-baseline mediator, or ignoring informative censoring under a hypothetical strategy) is still wrong. A fourth, specific to RWE, is importing the framework without the target-trial scaffolding — intercurrent events such as switching and rescue have no operational meaning until eligibility, treatment strategies, and time zero are defined, so applying E9(R1) to a prevalent-user or immortal-time-contaminated cohort produces a precisely-worded estimand for a biased number. Finally, E9(R1) is not a risk-of-bias instrument and not a quality score: a clearly specified estimand does not make an observational comparison causal; residual confounding, positivity violations, and measurement error must still be addressed with the relevant design and analysis methods and appraised with ROBINS-I.

How it maps to this catalog

— Each E9(R1) requirement is implemented by one or more concepts in this repository. The population/treatment/time-zero attributes are operationalized by target-trial-emulation, active-comparator-new-user, and immortal-time-bias-handling. The intercurrent-events and population-level-summary attributes are implemented by estimands-ate-att-intercurrent-events-rwe (the five ICE strategies and ATE/ATT contrasts) and made auditable by estimand-analysis-traceability-rwe (estimand-to-estimator traceability). Confounding control aligned to the chosen estimand is implemented by high-dimensional-propensity-score-hdps-rwe and propensity-score-methods-psm-iptw. Endpoint/phenotype validity is implemented by diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe and claims-outcome-algorithm-ppv-sensitivity-rwe; competing events (notably death as an intercurrent event) by competing-risks-cause-specific-fine-gray-rwe. Data fitness for the chosen ICE strategy is implemented by fit-for-purpose-data-assessment-rwe; attrition and informative censoring by attrition-and-loss-to-follow-up-rwe; the sensitivity-analysis requirement by e-value-sensitivity-analysis; and the decision-question framing by picots-framework-rwe. Applied note for claims/EHR/registry RWE: choosing an intercurrent-event strategy is also a data-fitness decision — a while-on-treatment or composite-with-death estimand demands reliable mortality and discontinuation capture (link to a death index; reconcile MA-only person-time and days-supply gaps), whereas a hypothetical "if patients had not switched" estimand demands defensible censoring-weight models. Write the estimand first, then verify the database can actually support it; the order is what separates decision-grade RWE from an estimand table.