ICH E9(R1): Intent-to-Treat Alone Isn’t an Estimand

ICH E9(R1) reached Step 4 in November 2019; FDA finalized its adoption as guidance in May 2021. Five years on, the framework it introduced — the estimand — still gets treated, on a lot of protocols, as an academic add-on to a statistical analysis plan that was really decided the old way: pick an analysis population, usually intent-to-treat, and move on. That habit misses what the addendum actually changed. An estimand is not an analysis method. It is a precise statement of the treatment effect the trial is designed to estimate — and "intent-to-treat" does not make that statement.

What the addendum actually added

Before ICH E9(R1), a protocol's primary objective and its primary analysis often did the work of an estimand without anyone naming it that. The addendum's contribution was to force those into the open: name the treatment effect of interest with enough precision that population, endpoint, intercurrent-event handling, and summary measure are all pinned down before a single patient is unblinded. That precision matters because different, equally defensible ways of handling the same intercurrent event can point to materially different treatment-effect estimates from the identical trial — a gap that a vague "intent-to-treat, last observation carried forward" specification papers over rather than resolves.

Why intent-to-treat is a default, not a specification

"Intent-to-treat" tells you an analysis population — everyone randomized, analyzed in the group assigned — and it carries a loose implication that intercurrent events are absorbed into the treatment effect rather than analyzed around. What it does not do is name the variable, the summary measure, or, critically, make a deliberate choice among the five ICH E9(R1) strategies for each specific intercurrent event a trial is likely to see. Two protocols that both say "intent-to-treat, primary endpoint at Week 24" can implement completely different handling of a patient who takes rescue medication at Week 10 — one folding the rescue event into a treatment-policy estimate of real-world effectiveness, the other silently imputing a value as though rescue never happened. Both are internally consistent. They are not the same estimand, and they will not produce the same number.

• Treatment policy. The intercurrent event is treated as part of the treatment regimen; its consequences are included in the outcome, regardless of what happened after it.
• Hypothetical. The estimand asks what the outcome would have been in a scenario where the intercurrent event did not occur — a counterfactual that requires its own, separately justified assumptions.
• Composite. The intercurrent event itself becomes part of the outcome variable, often used when the event (treatment failure, discontinuation for toxicity) is clinically meaningful in its own right.
• While on treatment. The variable is defined only up to the point the intercurrent event occurs, without extrapolating beyond it.
• Principal stratum. The population itself is redefined to the subgroup that would, or would not, experience the intercurrent event under a specified treatment condition.

Two teams can both write "intent-to-treat" on a protocol and disagree, without realizing it, about what the trial is actually trying to measure. Why the estimand has to be named explicitly

Where this becomes a regulatory decision, not a statistics footnote

The strategy a sponsor chooses for a given intercurrent event is not a neutral technical setting. A treatment-policy strategy for rescue medication answers a real-world-effectiveness question; a hypothetical strategy for the same event answers an efficacy-under-ideal-adherence question. Reviewers scrutinize whether the chosen estimand actually matches the clinical claim the sponsor wants in labeling, and a mismatch discovered at the end-of-Phase-2 meeting or, worse, at NDA submission is a protocol-level problem no amount of post hoc sensitivity analysis fully repairs. The exposure is sharpest in small, heterogeneous rare-disease trial populations, where a single patient's intercurrent event can visibly shift the result, and where the population attribute itself — who, exactly, the estimand describes — is often doing as much work as the intercurrent-event strategy.

None of this requires a larger statistics team. It requires deciding the estimand as a cross-functional call — regulatory, clinical, and biostatistics together — early enough that it shapes the protocol rather than gets reverse-engineered from an analysis plan someone already drafted. Sponsors who treat the estimand as connected to the same disciplined thinking behind ICH E6(R3)'s quality-by-design principles — naming the factors that are critical to a trial's ability to answer its question, before the trial starts — are the ones whose primary analysis still answers the intended question by the time it reaches an agency review.