# Who stayed in the analysis after people stopped participating?

Draft • Research checked 4 October 2026 • AI-assisted DomDNA editorial content. No independent clinical review or publication is claimed. General education, not individual medical advice.

A programme can look effective among people who completed it and less impressive among everyone who was offered it. Those results need not contradict each other. They can describe different questions, with different opportunities for selection to enter the analysis.

When a trial includes people who missed sessions, stopped an intervention or did not provide a final measurement, the analysis needs more explanation than a single familiar label.

## Assignment and completion are different events

Randomisation occurs at the start. Completion happens later and can depend on experiences after assignment. Restricting an analysis to people who finish may remove some of the protection created by the original random comparison.

The intention-to-treat principle concerns analysing people according to their initial assignment. But that phrase does not magically supply an outcome that was never collected. A [historical survey by Hollis and Campbell](https://www.bmj.com/content/319/7211/670) found differing practices behind the term. The paper is from 1999, not an estimate of how often current trials handle it well.

The [current CONSORT statistical-methods item](https://www.consort-spirit.org/item21a-primaryandsecondaryoutcomes) asks for enough detail to understand what was actually done. “We used intention to treat” is the beginning of that explanation, not the end.

## A deliberately simple example

Imagine two hypothetical digital coaching groups, each with 100 assigned participants. In one group, 60 people provide a final questionnaire; in the other, 90 do. The first group's responders report better average experiences than the second group's responders.

You cannot tell from those averages what happened among the missing 40 and 10 people. Nor can you assume they all did badly or all did well. The different response counts belong beside the averages because they change how confidently you can interpret the comparison.

Now imagine the report highlights only participants who attended every session. Attendance could reflect available time, enthusiasm, tolerability or improvement itself. Comparing those people is no longer simply comparing the groups randomisation produced.

This invented example is about denominators and selection. It gives no real effect estimate and is not a method for filling missing questionnaire answers.

## The research question has to be explicit

Sometimes the useful question is the effect of offering a programme under ordinary participation conditions. Sometimes researchers want to estimate an effect under a different, precisely defined scenario. Those are not interchangeable targets.

[FDA's final ICH E9(R1) guidance](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e9r1-statistical-principles-clinical-trials-addendum-estimands-and-sensitivity-analysis-clinical) uses the estimand framework to make the target effect explicit, including how events after assignment are addressed. Stopping a treatment, using additional care and losing an outcome measurement are not merely interchangeable housekeeping problems.

For a general reader, you do not need to calculate an estimand. You need to recognise that the analysis should match the question. A label such as per-protocol or intention-to-treat cannot tell you every detail about that match.

A sensitivity analysis can explore how conclusions change under alternative assumptions. It does not make the missing outcomes observable. If the main finding depends on one strong assumption, seeing that dependence is useful information rather than an inconvenience to hide.

## Do not treat every exclusion as deception

There can be planned, justified reasons for defining a particular analysis population. A safety analysis may need a different population from an effectiveness analysis. The key is whether the report explains the choice and whether the headline stays within its scope.

It is also possible for a paper to report several analyses honestly. The problem arises when the promotional claim quietly uses the most favourable subset as though it represented everyone initially assigned.

A participant-flow diagram often helps you orient yourself before reading statistical detail. Compare the numbers assigned, followed up and analysed, then locate the explanation for differences. Avoid guessing reasons from the size of the gap alone.

## What you can do with this

When saving a result, write down the starting denominator and the analysed denominator next to the outcome. Add one sentence about how non-completion and missing measurements were handled, if that information is available.

For the imaginary coaching study, a careful note would say “better reports among questionnaire responders, with unequal response counts”, not “the programme helped everybody more”. That phrasing preserves the observation without pretending the study answered a broader question.

This is useful beyond formal trials. Completion-based testimonials can tell you about some people's experiences while leaving the experience of people who left almost entirely out of view. Keep the population in the claim visible.

For an educational starting point, the [DomDNA lifestyle quiz](https://domdna.com/quiz) records lifestyle answers, not DNA analysis or a clinical diagnosis.