ICH E9 for Non-Statisticians: The Operational Contract That Locks Your Trial's Question Before Unblinding

At a glance

• ICH E9(R1) is not a statistics vocabulary lesson. It is the operational contract that ties a trial’s design, conduct, and analysis to one prespecified question, so the answer you report is the answer you asked.
• The estimand framework forces you to define the treatment effect of interest up front, including how intercurrent events (events after treatment initiation, such as discontinuation or rescue medication) are handled, before the trial is designed.
• Analysis populations (full analysis set, per-protocol set, safety set) each guard against a different bias. The choice of who is included must be predefined, not decided after you see the data.
• ICH E6(R3) gives the conduct-side teeth: the statistical analysis plan (SAP) must be consistent with the protocol, analysis-set membership must be predefined, edit access locks before unblinding, and any post-unblinding deviation must be justified and reported in the clinical trial report.
• ICH E8(R1) supplies the frame: estimands refine the study objectives into the critical-to-quality factors a quality-by-design trial protects from the start.
• Where teams get it wrong: finalizing the SAP after a peek at the data, leaning on the per-protocol set to rescue an efficacy claim, handling intercurrent events ad hoc, and breaking blinding inside the analysis pipeline.

ICH E9 in one sentence

For clinical-ops, CRAs, data managers, and QA staff, the most useful way to read ICH E9 is operational, not mathematical. It is the contract that locks a single, precise scientific question into the protocol and statistical analysis plan, and then obligates everyone to protect that question through conduct, data handling, and analysis. ICH E9(R1) frames the whole exercise around alignment: the statistical analysis of clinical trial data should be aligned to the estimand, the precise description of the treatment effect of interest. When that alignment holds, regulatory scrutiny can focus on sensitivity to assumptions rather than on whether the question itself drifted.

You do not have to choose the estimand or write the SAP to be accountable for it. You have to recognize the levers you actually control: protocol prespecification, who locks the analysis and when, blinding integrity through the analysis stage, and deviation and missing-data discipline. Those levers decide whether the trial survives analysis and inspection.

E9 vs E9(R1): which obligations come from which

This is the disambiguation most ranking primers skip. ICH E9 (the 1998 base guideline) introduced the foundational principles, including the Intention-To-Treat (ITT) principle, whereby subjects are followed, assessed, and analysed irrespective of their compliance to the planned course of treatment, on the logic that preserving randomisation provides a secure foundation for statistical tests. That principle is still load-bearing.

ICH E9(R1) is the 2019 Addendum on Estimands and Sensitivity Analysis. It does not replace the 1998 principles; it adds a structured framework on top of them. The addendum’s own framing is that precision in describing a treatment effect is facilitated by constructing the estimand, and that the statistical analysis should then be aligned to that estimand. It also re-opens the question of whether estimating an effect strictly in accordance with the ITT principle always represents the treatment effect of greatest relevance to regulatory and clinical decision making. So when a colleague says “per E9,” ask which layer they mean: the ITT and analysis-set principles, or the estimand and sensitivity-analysis addendum. In practice you are accountable for both.

The estimand and its five attributes

The addendum defines an estimand as a precise description of the treatment effect reflecting the clinical question posed by a given trial objective, and it instructs that clinical questions and their estimands should be specified at the initial stages of planning any clinical trial. That timing is the whole point: define the question, then design the trial to answer it, not the reverse.

ICH E9(R1) builds the estimand from a set of attributes. The treatment attribute is the treatment condition of interest and, where relevant, the alternative condition to which comparison will be made. The population attribute is the patients targeted by the clinical question, which may be the entire trial population, a baseline-defined subgroup, or a principal stratum. The variable (or endpoint) attribute is the measurement obtained for each patient that is required to address the clinical question. The population-level summary provides a basis for comparison between treatment conditions. And the handling of intercurrent events ties the others together: the addendum defines intercurrent events as events occurring after treatment initiation that affect either the interpretation or the existence of the measurements associated with the clinical question of interest.

A worked example helps. Suppose the question is the effect of an add-on therapy on symptom severity at 24 weeks. The treatment is “intervention A added to background therapy B, dosed as required”; the population is adults with the diagnosed condition; the variable is the symptom score at 24 weeks; the summary is the difference in mean scores between arms. Now the hard part: what happens when a patient discontinues A and starts a rescue drug at week 10? That is an intercurrent event, and how you handle it changes the question. ICH E9(R1) lays out distinct strategies. Under a treatment-policy strategy, the occurrence of the intercurrent event is considered irrelevant and the value of the variable is used regardless of whether the event occurs. A while-on-treatment strategy restricts the observation of interest to the time before the event. A principal-stratum strategy narrows the population to those in whom the event would (or would not) occur. These are not interchangeable. Picking one after unblinding is picking the answer you want.

Analysis populations decoded

The analysis-set choice is where non-statisticians most often see bias enter, and ICH E9(R1) is direct about it: for superiority trials it strongly recommends that analysis be based on the full analysis set, defined to be as close as possible to including all randomised subjects, and it warns that eliminating some planned measurements on some subjects can have consequences similar to excluding subjects altogether. The addendum also revisits the per-protocol set, asking whether the impact of protocol violations can be addressed in a less biased, more interpretable way than a naive per-protocol analysis. It notes that it is usually appropriate to plan analyses on both the full analysis set and the per-protocol set so that differences between them can be the subject of explicit discussion and interpretation.

Analysis population	What it includes	Bias it guards against	When it is primary
Full analysis set (close to ITT)	As close as possible to all randomised subjects	Selection bias from post-randomisation exclusions; preserves randomisation as the basis for statistical tests	The recommended primary basis for superiority efficacy claims under E9(R1)
Per-protocol set	Subjects who adhered to the protocol without important deviations	Dilution of a true effect by non-adherence, but introduces its own bias by conditioning on post-randomisation behaviour	Supportive only; planned alongside the full analysis set so differences can be interpreted, not as the sole efficacy basis
Safety set	Subjects analysed by treatment actually received	Misattribution of adverse events to the wrong arm	Safety and tolerability summaries

The operational rule that follows: membership criteria for every set must be predefined, and a per-protocol result that looks better than the full-analysis-set result is a flag to investigate and explain, not a headline to promote.

Prespecification and the unblinding line you cannot cross

This is where ICH E9 meets ICH E6(R3), and where the two regulations reinforce each other rather than conflict. ICH E6(R3) section 3.16.2 requires the sponsor to develop a statistical analysis plan that is consistent with the trial protocol and that details the approach to data analysis, unless that approach is sufficiently described in the protocol. The same section requires that the criteria for inclusion or exclusion of trial participants from any analysis set be predefined (for example, in the protocol or SAP), with the rationale for any exclusion clearly documented. That is the conduct-side enforcement of E9(R1)‘s analysis-set principle.

The unblinding line is explicit. ICH E6(R3) section 3.16.1 requires that, prior to provision of the data for final analysis and, where applicable, before unblinding the trial, edit access to the data acquisition tools be restricted. Section 3.16.2 then states that deviations from the planned statistical analysis, or changes made to the data after the trial has been unblinded, should be clearly documented and justified and should only occur in exceptional circumstances, and that post-unblinding data changes and deviations from the planned analyses must be reported in the clinical trial report. ICH E6(R3) section 3.4 also requires a statement in the protocol that any deviation from the SAP will be described and justified in the clinical trial report. Read together: you may change the analysis after unblinding only in exceptional, documented, justified, and reported circumstances. You cannot quietly retrofit the analysis to the result.

ICH E9(R1) supplies the reason this matters. The addendum stresses that an analysis aligned to the estimand and prespecified to a level of detail that a third party could replicate is what allows regulatory interest to focus on sensitivity rather than on whether the question moved. Prespecification is not bureaucracy; it is the thing that makes the eventual result interpretable.

Intercurrent events and missing data: choose the strategy up front

ICH E9(R1) draws a clean line that practitioners frequently blur. Intercurrent events are not the same as missing data, and they are not simply a drawback to be avoided. An intercurrent event is part of the question; missing data is a problem for inference. The addendum notes that having clarity in the estimand gives a basis for planning which data need to be collected, and hence which data, when not collected, present a missing-data problem to be addressed in the statistical analysis. It also recommends a prospective plan to collect informative reasons for why intended data are missing, which helps distinguish genuine intercurrent events from missing data.

The discipline is to decide the intercurrent-event strategy and the missing-data approach before the trial starts, and to align the analysis to that estimand. The addendum is explicit that estimation relying on many or strong assumptions requires more extensive, estimand-aligned sensitivity analysis to explore robustness, and that where the impact of deviations from assumptions cannot be adequately investigated, that combination of estimand and analysis method may not be acceptable for decision making. ICH E6(R3) reinforces the operational side: a prospective plan to retain subjects and reduce missing data protects the analysis you committed to.

Where E9 meets GCP conduct

E9 lives or dies on conduct quality, and ICH E6(R3) makes that link concrete. Randomisation and blinding remain cornerstones of controlled clinical trials, and ICH E6(R3) section 3.15 requires the sponsor to implement measures to safeguard the blinding, including maintaining the blinding during data entry and processing. ICH E6(R3) section 2.11 requires investigators to break the randomisation code only in accordance with the protocol and to promptly document and explain any premature unblinding. If blinding leaks into the data-handling or analysis pipeline, the protections E9 assumes are gone.

Protocol deviations feed the analysis directly. ICH E6(R3) section 3.10 requires the sponsor to determine criteria for classifying protocol deviations as important, defining important deviations as a subset that may significantly impact the completeness, accuracy, or reliability of the trial data or a participant’s rights, safety, or wellbeing. That classification is exactly what drives per-protocol-set membership and the interpretation of full-analysis-set results, which is why deviation discipline is an analysis concern, not just a monitoring one.

ICH E8(R1) closes the loop at the design layer. It frames quality by design as proactively designing quality into the protocol and processes, organised around critical-to-quality factors whose integrity is fundamental to the reliability and interpretability of the results. ICH E8(R1) section 6.1 states that study objectives are further refined through specification of estimands, citing ICH E9(R1), and ICH E8(R1) section 3 lists clear predefined objectives, bias-minimising approaches such as randomisation and blinding, and well-defined endpoints among the design elements that determine study quality. In other words, the estimand is one of the critical-to-quality factors a quality-by-design trial is built to protect.

Operational checklist: E9 obligations by role

• Clinical-ops / project lead: Confirm the protocol states the estimand and intercurrent-event strategy before first-patient-in; confirm the SAP is consistent with the protocol and finalized before unblinding; confirm the protocol carries the required statement that any SAP deviation will be justified in the clinical trial report.
• CRA / site monitor: Verify blinding is maintained at the site; document all protocol deviations and surface candidates for “important” classification, since those feed analysis-set decisions; confirm randomisation codes are broken only per protocol and any premature unblinding is documented.
• Data manager: Ensure analysis-set inclusion/exclusion criteria are predefined and traceable; restrict edit access to data acquisition tools before final analysis and unblinding; preserve audit trails for any post-unblinding data change; maintain blinding through data entry and processing.
• QA / regulatory: Confirm the SAP lock predates unblinding; confirm any post-unblinding analysis deviation is exceptional, justified, authorised by the investigator, and reported in the clinical trial report; confirm the trial’s design and conduct will let the prespecified estimand actually be estimated.

The throughline is simple to state and hard to honour under deadline pressure: prespecify the question, protect it through conduct and blinding, and never let the analysis be chosen after you have seen the answer. No regulation in this set can certify that your trial is compliant. ICH E9(R1), E6(R3), and E8(R1) define what is required; meeting those requirements, and proving it on inspection, remains the sponsor’s responsibility.

For related operational depth, see the sibling guides on protocol deviation classification, essential documents and the trial master file, and blinding and randomisation integrity, and the quality-by-design / RBQM pillar that frames critical-to-quality thinking across the program.

Sources

• ICH E9(R1) Statistical Principles for Clinical Trials (Addendum on Estimands and Sensitivity Analysis), ICH, version r1 (2019)
• ICH E6(R3) Good Clinical Practice, ICH, version r3 (2025) — https://www.ich.org/page/efficacy-guidelines
• ICH E8(R1) General Considerations for Clinical Studies, ICH, version r1 (2021)