The assignment is only the beginning of a trial
Imagine a fictional study assigning 200 people to an evening movement program and another 200 to usual activities. Some assigned participants attend every session. Others attend occasionally, change schedules, or never begin. Some people in the comparison group independently join a different program.
These developments do not automatically make the study useless. They do mean that “Did the program work?” has become more than one question. Did offering this program under these conditions improve outcomes? What would happen if people followed the intended schedule? The first question concerns assignment; the second concerns adherence to a specified intervention.
What intention-to-treat preserves
Under the intention-to-treat principle, people are analyzed in the groups to which they were randomized. Someone assigned to the evening program remains in that group even if they never attend. This preserves the initial comparison created by randomization rather than reorganizing participants according to later choices or circumstances.
In the fictional study, suppose outcome information is available for everyone. Thirty of the 200 assigned participants meet a prespecified improvement criterion, compared with twenty of the 200 comparison participants. The observed proportions are 15% and 10%, a difference of five percentage points. That describes the assignment comparison; it is not a calculation restricted to people who attended.
The result includes the way this particular offer operated: its schedule, accessibility, attendance, and alternatives. Calling it the effect of an offer does not mean it automatically generalizes to every real-world program. A different setting might make participation easier or harder, changing what assignment produces.
Why comparing completers can change the question and the bias
Now suppose only 100 assigned participants completed the schedule, and 25 of them met the improvement criterion. Comparing their 25% improvement rate with the comparison group’s 10% might look like a clearer test of the program. But the completers were not randomly selected from the assigned group.
Perhaps people with flexible work hours could attend more often. Perhaps early improvement encouraged continued participation, while early discomfort discouraged it. If those factors also relate to the outcome, selecting completers can change the comparison in ways unrelated to the program’s causal effect. The apparently stronger result may reflect both participation and who remained.
This does not make adherence an unimportant question. It explains why estimating a per-protocol effect can require methods that account for predictors of adherence, sometimes measured repeatedly over time, along with assumptions about what was measured. Excluding nonadherent participants is not a substitute for that reasoning.
As-treated analysis is another distinction
An as-treated analysis groups participants by the intervention they actually received. Someone assigned to usual activities who joined the program could move into the treated group. A simple per-protocol analysis often excludes certain departures instead. Neither reorganization retains the original randomized comparison automatically.
| Approach | What happens to someone assigned a program who never starts? |
|---|---|
| Intention-to-treat | Remains in the assigned program group |
| Simple per-protocol restriction | May be excluded under the study’s adherence definition |
| As-treated grouping | May be analyzed with those who did not receive the program |
These are simplified descriptions. Actual protocols can define adherence in more detailed ways, and methods intended to estimate per-protocol effects need not consist of a simple restriction. The study’s definitions matter more than assuming that a familiar label describes every analytical choice.
Missing outcomes are a separate problem
Suppose the study never learns whether some participants improved. Keeping their group labels does not reveal the missing outcomes. Analyses may use imputation, weighting, modeling, or sensitivity analyses, each requiring assumptions. A claim that an analysis is intention-to-treat does not show those assumptions are reasonable.
“Modified intention-to-treat” is especially difficult to interpret without the eligibility rules. One report might exclude people with no outcome measurement; another might exclude those who never received a dose or attended a session. Those choices can have different consequences. The participant flow and the reasons for exclusion help explain what was actually estimated.
Two useful answers can still disagree
A program may help when followed and produce a smaller assignment effect because few people can participate as planned. Alternatively, apparent benefits among completers may partly reflect selection. Disagreement between analyses is a reason to examine their questions and assumptions, not an automatic reason to select the larger benefit.
The interpretation becomes clearer when a report specifies the population, intervention, comparator, outcome, and handling of events after assignment. Those details connect the statistic to a real question. They also leave room for the practical reality that participation is shaped by circumstances, rather than treating every departure as a personal failure.
