Modified Intention-to-Treat: When Randomization Starts Losing Patients After the Fact

What Intention-to-Treat Actually Protects

For the effect of treatment assignment, the intention-to-treat principle has two practical commitments: analyze participants in the groups to which they were randomized, and try to include outcome information from every randomized participant. The first prevents treatment switching from rewriting allocation. The second prevents post-randomization selection from rewriting the trial population.

Intention-to-treat does not mean pretending missing outcomes exist. It means designing follow-up around the effect of interest, preserving allocation, and addressing unavailable outcomes transparently rather than redefining inconvenient participants out of the analysis.

“Modified ITT” Is Not One Reproducible Method

One paper may use modified ITT to mean everyone with at least one post-baseline measurement. Another may require one dose, a minimum exposure duration, a confirmed diagnosis after randomization, or no major protocol deviation. Those rules target different populations and open different bias pathways.

A reviewer should never accept “mITT population” as a sufficient denominator description. Replace the acronym with the exact inclusion rule, then ask what happened between randomization and eligibility for that analysis.

When Can a Post-Randomization Exclusion Be Defensible?

Some exclusions may create little bias, but the burden of proof is narrow. A delayed finding that a participant violated an entry criterion is most defensible when the criterion concerns information fixed before randomization, detection is objective, every arm receives equal scrutiny, and the decision is made without access to treatment allocation or outcomes.

“Never dosed” is not automatically harmless. If patients or clinicians know the assignment before deciding whether to begin treatment, the no-dose subset can differ systematically across arms. “No data” is not automatically harmless either. Missing outcomes require a missing-data argument, not deletion by vocabulary.

Clinical Example: The First Post-Baseline Visit Gate

Imagine a randomized trial of two preventive therapies. The primary modified ITT analysis includes only participants who take one dose and return for a four-week assessment. In the experimental arm, early side effects lead some patients to stop immediately and skip the visit. In the control arm, missed visits are mostly logistical.

Excluding both groups under one neat rule does not make the mechanism equal. The experimental analysis has removed patients whose early experience may predict a poor outcome, while the control exclusions may be less prognostic. The retained arms can therefore look more comparable on paper precisely because the clinically informative difference was deleted.

A credible report would show randomized and analyzed denominators by arm, give reasons for each loss, preserve outcomes after treatment discontinuation when possible, and stress-test assumptions about the outcomes that remain unavailable.

A Reviewer Checklist for Post-Randomization Exclusions

Also inspect the participant flow diagram and outcome-specific denominators. CONSORT 2025 asks reports to show losses and exclusions after randomization with reasons, and to report how many participants were included and had available data for each outcome and time point. One headline sample size is not enough.

What Not to Conclude

1. “The groups still looked balanced.” Balance on measured baseline variables cannot rule out selection on unmeasured prognosis or post-baseline events.
2. “Only a small percentage was excluded.” Direction, reasons, and outcome risk matter; a modest strategically selected loss can matter more than a larger random one.
3. “Both arms used the same rule.” A symmetric rule can act asymmetrically when treatment changes the probability of satisfying it.
4. “ITT is always conservative.” Missingness and post-randomization selection can bias toward or away from the null. The direction must be argued, not assumed.

Where Aqrab Fits

Analysis-set language often sounds routine enough to escape scrutiny. Aqrab helps reviewers turn a label such as “modified ITT” into the questions that matter: who was randomized, who disappeared, when the rule was applied, whether allocation influenced inclusion, and which estimand the remaining analysis can support.

If you are reviewing a trial manuscript or protocol, try Aqrab on the analysis-population section. If you want the same methodological checks inside a research workflow, start with the developer documentation.

Sources and Further Reading

• CONSORT 2025 explanation and elaboration, including participant-flow and outcome-denominator reporting.
• ICH E9(R1) addendum on estimands and sensitivity analysis, including follow-up after treatment discontinuation and preservation of randomization.
• ICH E9 statistical principles for clinical trials, including full analysis sets and conditions around exclusions.
• Cochrane Handbook, Chapter 8, on intention-to-treat effects, missing outcomes, and post-randomization exclusions.

The Practical Bottom Line

Modified intention-to-treat is not automatically wrong. It is automatically incomplete. The modifier must be unpacked into an exact rule, a timeline, arm-specific reasons, and a sensitivity analysis.

Randomization is a design achievement, not a permanent force field. Once the analysis removes patients based on what happened after allocation, the paper must earn back the credibility that the acronym implies.