People do not always continue the treatment assigned at the start of a trial. Some stop, some switch and some need additional care. The analysis must say what those events mean for the question it is trying to answer. A tidy treatment label cannot do that work on its own.
- Identify whether the question concerns assignment, adherence or another defined treatment strategy.
- Stopping treatment does not automatically mean removing someone from follow-up.
- Comparing only adherent participants can lose the protection of the original randomization.
Begin with the decision being studied
Imagine an invented trial comparing offers of two rehabilitation programmes. One question is what happens after people are assigned to programme A rather than B, including the interruptions and changes that actually occur. Another is what would happen if everyone followed the assigned programme as specified.
Those questions are related, but they are not identical. The ICH E9(R1) framework asks researchers to define the treatment effect of interest, including how events such as discontinuation or extra treatment are handled. That target is called an estimand.
Before reading the analysis label, write the practical question in ordinary language. Then check whether the reported comparison actually matches it.
Source 1 ↗Keep the original groups visible
Suppose 100 people are assigned to each programme. Twenty in A and five in B do not complete their assigned course. Reliable week-12 outcomes are nevertheless obtained for everyone. Forty-eight in A and forty in B improve under a fictional criterion.
An analysis by assigned group gives 48% versus 40%, an 8-percentage-point difference. The interrupted participants remain with their original groups. These invented counts illustrate an assignment comparison; they do not establish statistical significance or prove the programmes effective.
An intention-to-treat analysis preserves assignment rather than reorganising people according to subsequent behaviour. Read the denominators to see whether a paper’s use of the label matches that principle.
Source 1 ↗See what a completer-only comparison changes
Now suppose 44 of A’s 80 completers improved and 38 of B’s 95 completers improved. Among completers, the percentages are 55% and 40%, a 15-point difference. The same invented study now appears to have a larger advantage for A.
That change does not automatically reveal the effect of perfect adherence. Completing a programme may depend on early symptoms, transport, motivation, health or perceived benefit. The restricted groups can therefore differ for reasons connected to the outcome.
Our exercise deliberately supplies the same original trial with two denominators. The right question is not “Which result is bigger?” but “What selection produced the second comparison, and what causal interpretation can its method support?”
A treatment change is not the same as absent data
In the example, all outcomes are measured despite treatment interruption. That is different from losing contact and not knowing the outcome. The estimand framework distinguishes events that affect the meaning of the treatment question from missing observations that affect its estimation.
Imagine that additional therapy is allowed when recovery is poor. A result including outcomes after that therapy describes a different strategy from a hypothetical result assuming it had not been used. Both need clear wording and suitable methods.
Neither strategy permits casually deleting inconvenient participants. Ask what the protocol specified and whether the data collected can support the chosen question.
Source 1 ↗Read modified intention-to-treat literally
A report may use the phrase “modified intention-to-treat.” Treat “modified” as an invitation to inspect its definition. Does the analysis exclude people who never started treatment, people without a later measurement, or those found ineligible after assignment? Different rules can produce different study populations.
Create a small participant ledger for our example: assigned, started, completed, outcome observed, analysed. Put the actual count from the report beside each word. If “analysed” is smaller than “assigned,” find the reason rather than assuming the difference is harmless.
A useful summary names the exclusion. “The analysis excluded people without a post-baseline measurement” communicates more than repeating a reassuring acronym.
What the next analysis can and cannot add
A carefully designed adherence analysis may address a worthwhile question, but it needs methods and assumptions that account for why adherence varies. A raw comparison of completers is not an automatic solution. The user-facing conclusion should preserve those conditions.
For our fictional programmes, future evidence could clarify why people stopped, improve delivery, or test the offered programme in another service. Better implementation might alter the assignment effect without changing the biological effect of any component.
Connect papers by their questions: assignment, sustained adherence, or a hypothetical course without additional treatment. If the questions differ, explain that before describing the estimates as agreement or contradiction.
Two analyses of one invented trial
| Comparison | Programme A | Programme B | Difference |
|---|---|---|---|
| By original assignment | 48/100 improved = 48% | 40/100 improved = 40% | 8 percentage points |
| Completers only | 44/80 improved = 55% | 38/95 improved = 40% | 15 percentage points |
| Non-completers, described separately | 4/20 improved = 20% | 2/5 improved = 40% | Descriptive selected groups; not a randomized comparison |
All week-12 outcomes are observed in this hypothetical example. No uncertainty calculation or unbiased adherence-effect estimate is implied.
Your questions, answered
Does intention-to-treat mean pretending everyone took the treatment?
No. It analyses people with their original assigned groups; it does not erase interruptions or invent treatment exposure. Those events should still be described and interpreted in relation to the trial’s question.
Does intention-to-treat always estimate perfect adherence?
No. An assignment comparison and the effect of following a specified treatment course answer different questions. The target must be described before choosing an analysis.
Can per-protocol analyses be useful?
Yes, for an appropriate question and with defensible methods. Simply excluding non-adherent people can create selection bias. Read how the analysis addresses that selection rather than treating its label as sufficient.
Why are the example’s results 8 and 15 percentage points?
Assigned-group proportions are 48/100 and 40/100. Completer proportions are 44/80 and 38/95. The arithmetic uses different selected groups, so the two differences are not interchangeable estimates.
What if someone stops treatment but attends follow-up?
Their outcome may still be observed and relevant to the planned question. Stopping the intervention and withdrawing from data collection should be recorded separately.
What should I write beside an unfamiliar estimand?
Write the people, treatments, outcome, time point, handling of treatment changes and summary comparison in plain language. If you cannot reconstruct those elements, record the specific missing information rather than guessing.
Limits of this interpretation
- This is a selected educational explanation, not a systematic review, validated appraisal instrument or personal care recommendation.
- Numerical examples are hypothetical. Their deliberately simplified assumptions must not be transferred to a real study without checking its methods.
- An AI source check can miss errors; source access and the absence of independent human review are stated explicitly.
Sources & transparency
- ICH E9(R1), adopted by FDA: Estimands and Sensitivity Analysis in Clinical Trials (2021)
Public final guidance PDF, selected sections A.1–A.3 and A.5 checked by AI for treatment discontinuation, withdrawal, the question of interest and sensitivity analyses. · Accessed 27 Sep 2026
Prepared and source-checked with AI on 27 September 2026. Press-news Team is the publication’s collective byline, not a claim of medical credentials or human review. No human editorial or clinical review has taken place. Source access is described below each reference. Worked examples are invented for education and do not report a clinical trial or predict an individual outcome.
Source check: AI source check — selected methods references and worked examples
Clinical review: No human editorial or clinical review
Suggest a correction