A trial begins with everyone on the participant-flow diagram. The final result may use fewer people. To judge that result, ask why outcomes are missing and how the analysis dealt with the gap. The percentage missing is a starting point, not the whole diagnosis.
- Stopping treatment and losing outcome follow-up are different events.
- A complete-case percentage can hide different denominators.
- Sensitivity analyses show how conclusions depend on assumptions; they do not recover unknown outcomes.
Outcome availability in the fictional trial
People out of 100 assigned in each group
Follow the denominator through the trial
In an invented trial, 100 people are assigned to A and 100 to B. At week 12, outcomes are available for 80 in A and 90 in B. Of those observed, 48 and 45 respectively meet a fictional improvement criterion.
Among observed participants, improvement is 48/80=60% versus 45/90=50%. That ten-point difference is descriptive of the observed groups. It does not tell us the outcomes of the twenty missing people in A or the ten in B.
Write the initial number, observed number and missing number together. A paper that reports only “60% versus 50%” leaves a substantial part of this example out of view.
Reasons matter, not only percentages
Someone might miss a visit because of travel, because symptoms worsened or because recovery made follow-up feel unnecessary. Those stories imply different possibilities for the unobserved outcome. Equal missing percentages between groups would not guarantee equal consequences.
Cochrane’s risk-of-bias guidance asks whether missingness could depend on the true outcome. In the fictional trial, we would want reasons by group and by visit, not just a total labelled “lost.”
Treat “withdrew from treatment” carefully. A person can stop the intervention and still provide outcome data. A flow diagram should help you distinguish treatment exposure from whether the question’s outcome was actually measured.
Source 1 ↗Calculate what is known before filling any gaps
If nobody missing improved, the observed successes amount to 48/100 in A and 45/100 in B. If every missing person improved, they become 68/100 and 55/100. These are bounds on the complete-group proportions under a binary outcome and our stated counts, not estimates of what actually happened.
The most unfavourable combination for A is 48% versus 55%, a 7-point disadvantage. The most favourable is 68% versus 45%, a 23-point advantage. The gap spans very different stories.
Do not report those extremes as a confidence interval. They are an assumption exercise with no sampling-uncertainty calculation. Their purpose is to show what the observed counts cannot settle.
Try a less extreme sensitivity scenario
Assume, for illustration, that 20% of A’s missing participants improved and 80% of B’s missing participants improved. That adds four successes to A and eight to B, giving 52/100 versus 53/100. The apparent advantage reverses.
This does not prove that the actual result is reversed. The scenario must be judged against clinical context and the reasons outcomes are missing. Our invented trial has no such supporting information, so the scenario remains a teaching assumption.
The methodological article by White and colleagues advocates a main analysis under explicit plausible assumptions, with sensitivity analyses examining departures. A useful report explains why its assumptions are reasonable rather than merely naming a statistical method.
Source 2 ↗Imputation is an analysis, not a recovered measurement
A model may estimate missing outcomes using observed information. It does not discover what an unobserved participant actually experienced. Different models or assumptions can produce different answers, so readers need to know which variables and relationships were used.
Imagine that a report replaces every missed week-12 score with the last recorded score. Our reading question is simple: why would the earlier value represent the later outcome, and how is uncertainty reflected? A convenient replacement rule is not an explanation.
For the fictional trial, we would ask for the primary analysis, the assumed missing-outcome mechanism and scenarios that could change the conclusion. We would keep those results beside the participant-flow counts.
What better follow-up would add
A future study could reduce missing outcomes through accessible visits or continued contact after treatment stops. It could document reasons more clearly and plan sensitivity analyses before results are known. None of these steps guarantees complete data, but each addresses a specific uncertainty.
An additional analysis of our invented dataset would mainly test assumptions; it would not add observed outcomes. A follow-up that actually obtains reliable week-12 information from previously missing participants adds a different kind of evidence.
The useful conclusion is therefore conditional: “The observed participants favour A, but the complete-group comparison depends on missing outcomes.” That wording preserves the finding without pretending the denominator problem has disappeared.
The missing-outcome exercise
| Scenario | A: improved / assigned | B: improved / assigned | A minus B |
|---|---|---|---|
| Observed participants only | 48/80 = 60% | 45/90 = 50% | +10 points; observed denominators |
| All missing did not improve | 48/100 = 48% | 45/100 = 45% | +3 points |
| Least favourable to A | 48/100 = 48% | 55/100 = 55% | −7 points |
| Most favourable to A | 68/100 = 68% | 45/100 = 45% | +23 points |
| 20% of A missing; 80% of B missing improved | 52/100 = 52% | 53/100 = 53% | −1 point |
Invented week-12 binary outcomes. The first row has observed denominators; all later rows use the full assigned groups. Scenarios are not estimates or confidence intervals.
Your questions, answered
Is there a universal safe percentage of missing data?
No. The potential impact depends on why data are missing, how many outcome events occurred, and how strongly plausible missing outcomes could change the comparison. A small percentage can matter in some settings.
Does dropping out always mean the outcome is missing?
No. Participants may stop treatment but remain in follow-up. Check the flow diagram and outcome tables rather than treating discontinuation and missing measurement as synonyms.
Are the −7 to +23 points a confidence interval?
No. They are extreme arithmetic bounds in our fictional binary example: A-minus-B ranges from 48−55 to 68−45. They do not account for sampling uncertainty or assign probabilities to scenarios.
Does the reversal scenario show that A is worse?
No. It shows what follows if the assumed missing-participant improvement rates were 20% in A and 80% in B. Without evidence for those assumptions, it is a sensitivity exercise, not a finding.
Does an intention-to-treat label solve missingness?
No. Keeping people in their assigned groups and obtaining or modelling missing outcomes are separate issues. Read what data actually entered the analysis and which assumptions were required.
What would increase confidence in the result?
Clear reasons for missingness, good follow-up after treatment changes, an explicit main assumption and sensitivity analyses spanning credible alternatives. Agreement across arbitrary scenarios alone is not a substitute for plausible assumptions.
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
- Cochrane Handbook, chapter 8: Assessing risk of bias in a randomized trial
Public HTML, selected sections on randomization, concealment, blinding and missing outcomes; checked by AI. This guide is not a formal RoB 2 assessment. · Accessed 27 Sep 2026
- White, Horton, Carpenter and Pocock (2011): Strategy for intention to treat analysis in randomised trials with missing outcome data
Publisher-indexed abstract and methods-summary passages checked by AI. Direct article and PDF retrieval returned 403. Used for the principle of explicit assumptions and sensitivity analysis, not as a fully reviewed dataset. · 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