ICH E9 (Statistical Principles for Clinical Trials)

Most people typing “ICH E9” into a search box in 2026 want the estimand framework — and that is not in the 1998 guideline. It is in the 2019 addendum ICH E9(R1). The distinction matters because E9 itself was never revised: there is no E9(R2). The original Statistical Principles for Clinical Trials stands exactly as adopted at Step 4 on 5 February 1998, and E9(R1) sits beside it with separate numbering — references of the form x.y point into E9, references of the form A.x.y into the addendum.

The two documents, dated

E9 was approved under Step 2 on 16 January 1997, reached Step 4 on 5 February 1998, and was re-codified in November 2005 (code change, not content). In the EU it is CPMP/ICH/363/96, in force since 1 September 1998. The addendum — Addendum on Estimands and Sensitivity Analysis in Clinical Trials to the Guideline on Statistical Principles for Clinical Trials — was endorsed at Step 2 on 30 August 2017 and adopted by the Regulatory Members of the ICH Assembly under Step 4 on 20 November 2019 (the adopted document is dated 17 November 2019). In the EU it is EMA/CHMP/ICH/436221/2017, published 18 February 2020, legally effective 30 July 2020. Both are current; neither is superseded.

What E9 contains, section by section

• 3.3.1 / 3.3.2 — trials to show superiority, and trials to show equivalence or non-inferiority. E9 is the origin of the rule that these designs are not interchangeable and that a margin must be justified in advance; see non-inferiority vs superiority trial design.
• 3.5 Sample Size, with 3.4 Group Sequential Designs and 4.4 Sample Size Adjustment — the trio behind nearly every “ICH E9 sample size” search. Sample size must be justified against a pre-specified effect worth detecting; mid-trial adjustment is a planned design feature, not a rescue.
• 5.1 Prespecification of the Analysis — the principle governing the whole guideline. An analysis is E9-consistent because it was fixed before unblinding, not because it was sophisticated.
• 5.2.1 Full Analysis Set, 5.2.2 Per Protocol Set, 5.2.3 Roles of the Different Analysis Sets. E9 introduces the full analysis set as the practical form of the intention-to-treat principle — “as complete as possible and as close as possible to the intention-to-treat ideal of including all randomised subjects” — and gives the two aims governing analysis-set choices: minimise bias, avoid inflation of type I error.
• 5.3 Missing Values and Outliers, 5.6 Adjustment of Significance and Confidence Levels (multiplicity), 4.5 Interim Analysis and Early Stopping, 4.6 Role of Independent Data Monitoring Committee (IDMC).
• 2.3.1 / 2.3.2 — blinding and randomisation, the two design techniques for avoiding bias.

E9(R1): the gap the addendum closes

The addendum’s own diagnosis is that intention-to-treat, as E9 framed it, describes the effect of a treatment policy — subjects followed, assessed and analysed irrespective of compliance with the planned course of treatment — and that this is one legitimate clinical question among several, not the only one. What E9 lacked was a discipline for stating which treatment effect a trial intends to estimate before a method is chosen to estimate it. That is the estimand framework.

Section A.3.3 gives the attributes that jointly define an estimand: the treatment condition, the population targeted by the clinical question, the variable (endpoint) obtained for each patient, the specification of how intercurrent events are reflected, and the population-level summary providing the basis for comparison. Section A.3.2 names the five strategies for addressing intercurrent events: treatment policy, hypothetical, composite variable, while on treatment, principal stratum. An intercurrent event is a post-randomisation occurrence affecting interpretation — discontinuation, rescue medication, death — and the addendum is explicit that a treatment-policy strategy cannot be used for terminal events, because values of the variable after them do not exist.

A.5.2 and A.5.3 separate two things analysis plans routinely conflate: sensitivity analysis stress-tests the assumptions behind the estimate of the same estimand; supplementary analysis addresses a different question. Filing the latter under the former’s heading is a common way a statistical analysis plan drifts out of alignment with E9(R1).

What E9 does not do

E9 is not a menu of approved tests and prescribes no method for any endpoint. It is about pre-specification, bias, and control of type I error in confirmatory trials supporting a submission; exploratory work is treated separately (2.1.3). It does not displace trial-conduct guidance — it sits alongside ICH GCP. A trial can be fully GCP-compliant and still fail E9 by choosing its primary analysis population after unblinding.

Related terms

• Estimand framework — the E9(R1) construct most “ICH E9” searches are aimed at.
• Statistical analysis plan (SAP) — where pre-specification is discharged.
• Intent-to-treat vs per-protocol analysis — sections 5.2.1 and 5.2.2.
• Non-inferiority vs superiority trial design — section 3.3.
• Randomization methods in clinical trials — section 2.3.2.
• Randomized controlled trial (RCT) — the design E9 assumes.
• ICH E2A — the safety-reporting counterpart in the same efficacy series.

Frequently Asked Questions

Is there an ICH E9(R2)?

No. E9 itself was never revised beyond a November 2005 re-codification, which was a code change rather than a content change. The 2019 addendum on estimands and sensitivity analysis is officially designated ICH E9(R1), not E9(R2), and it sits beside the original 1998 guideline with its own separate section numbering rather than replacing it.

What is the difference between the Full Analysis Set and the Per Protocol Set under ICH E9?

Section 5.2.1 defines the Full Analysis Set as the practical operationalization of the intention-to-treat principle — as complete as possible and as close as possible to including all randomised subjects. Section 5.2.2 defines the narrower Per Protocol Set. Section 5.2.3 sets out the role each plays, governed by the same two aims that run through E9: minimise bias and avoid inflation of type I error.

How does ICH E9(R1) differ from ICH E9?

ICH E9 (1998) sets out general statistical principles for confirmatory trials — pre-specification, control of bias, and type I error control — but does not itself provide a discipline for stating which treatment effect a trial intends to estimate before a method is chosen to estimate it. ICH E9(R1), the 2019 addendum, closes that gap by introducing the estimand framework. The two documents use separate numbering — E9 uses section references of the form x.y, the addendum uses A.x.y — and both remain current.

Is ICH E9 the same as ICH GCP?

No. ICH E9 governs statistical pre-specification, bias control, and type I error control in the design and analysis of confirmatory trials, and prescribes no specific test for any endpoint. ICH GCP governs trial conduct. The two guidelines sit alongside each other, and a trial can be fully GCP-compliant while still failing E9 — for example, by choosing its primary analysis population after unblinding.

When should a statistical analysis plan be finalized under ICH E9?

Before unblinding. Section 5.1, described as the principle governing the whole guideline, holds that an analysis is E9-consistent because it was fixed before unblinding, not because it was sophisticated. That standard applies to the full range of choices a statistical analysis plan makes, including hypothesis structure, sample size, analysis populations, and the handling of missing data and multiplicity.

Does ICH E9 specify a required sample size or formula?

No. Section 3.5 requires that sample size be justified against a pre-specified effect worth detecting, and Section 4.4 allows for sample-size adjustment as a planned design feature rather than a mid-trial rescue. E9 does not prescribe a specific formula or minimum sample size, since that depends on the individual trial’s design and endpoint.