Invented counts before any missing-outcome coding
Same six-month assessment and seven-day cigarette endpoint in both groups. Unknown means no available outcome, not observed smoking. No real trial data.
| Recorded status | A | B |
|---|---|---|
| Randomised | 100 | 100 |
| Known to meet the endpoint | 20 | 10 |
| Known not to meet the endpoint | 60 | 50 |
| Outcome unknown | 20 | 40 |
First ask what is actually missing
Loss to follow-up is one route to a missing outcome, not its only meaning. A person may attend without supplying the relevant assessment, or a record may be unavailable. Stopping an assigned intervention is also not automatically the same as losing the outcome: a later assessment may still exist.
Keep observed smoking, missing self-report and missing verification separate in the report. If the endpoint requires verification, read that rule too. A coding decision that someone does not meet a documented endpoint is not an observation of that person's smoking or a judgment about their motives.
Why the convention became attractive
A 2012 methodological study of an Internet cessation trial describes the familiar rationale: people who quit might be more likely to report their outcome. Counting only responders could then make the group's proportion look too favourable. That rationale is a hypothesis about response, not a universal description of people who miss follow-up.
Assigning every missing outcome to non-abstinence avoids deleting those people from the denominator. It is simple, but does not recover their results or account for every uncertainty about them. The paper specifically questions treating this rule as automatically conservative for an effect comparison. Its particular trial and model do not establish what happens in all services.
Keep the unknown people visible in a count table
Everything in the table is invented, with no real treatment or study. A and B each have 100 randomised participants. At six months the endpoint is no cigarette smoking in the preceding seven days. Assume every available status is correctly recorded; only the unknown outcomes remain unresolved.
A has 20 meeting the endpoint, 60 not meeting it and 20 unknown. B has 10, 50 and 40 respectively. Coding all unknowns as non-abstinent gives 20/100 = 20% versus 10/100 = 10%: A minus B is ten percentage points. The arithmetic is clear; the missing outcomes are still unknown.
Two low proportions do not make a lower-bound difference
Now imagine, only to test the logic, that none of A's 20 unknown participants meets the endpoint but 15 of B's 40 do. The completed counts would be 20/100 versus 25/100, so A minus B would be minus five percentage points, not plus ten. The original rule understated both groups' completed proportions or left one unchanged, yet overstated their difference in this hypothetical completion.
This is not a claim about the absent participants, a plausible scenario certified for a real trial, or evidence favouring B. It is a mathematical counterexample to ‘lower within each group means conservative between groups’. Real sensitivity scenarios need a study-specific justification; these fabricated numbers supply none.
Deleting unknown outcomes is not a truth-finding alternative. In the same exercise, respondents-only proportions would be 20/80 = 25% and 10/60 ≈ 16.7%, with different denominators. They describe available records; calling them the real full-group rates would require assumptions too.
What a useful sensitivity analysis keeps fixed
Locate the protocol or statistical analysis plan: which population, endpoint and comparison were intended, what missing-outcome rule was primary, and which checks were planned? ICH E9(R1) distinguishes testing assumptions about the same estimand—the effect being estimated—from answering a different question. A result at another follow-up or for another population is not simply an alternative missing-data assumption.
Read whether justified alternative assumptions materially change the conclusion, and whether they were prespecified or added later. Plausibility comes from the study's context and information, not from picking a convenient calculation. There is no universal percentage of missing data that makes bias harmless. Equal missing counts also do not establish equal reasons or unknown outcomes.
For a real report, preserve the group-specific available and missing counts, known reasons, endpoint and assessment time, coding rule and reported sensitivity findings. If those findings are absent, say that the materials read do not show how conclusions depend on missingness. Do not invent a robustness certificate or search for participants' private information.
A defensible summary leaves room for what is not known
For this exercise you can write: ‘With all missing outcomes coded as non-abstinent, the invented proportions are 20% and 10%, a difference of ten percentage points. The rule does not show the absent participants' behaviour or guarantee a lower comparative effect.’ No confidence interval or causal benefit was supplied.
The statistical rule cannot choose your support
For personal help in England, NHS Better Health links to local Stop Smoking Services. A qualified professional can discuss your situation; a trial's missing-data convention cannot select care for you. Service arrangements and eligibility elsewhere need local confirmation.
What to keep in mind
Sources
The central claims on this page were checked against the sources below.
- Cochrane: Cochrane Handbook chapter 8, §8.5 — bias due to missing outcome data
Sources checked: 2026-10-04
- International Council for Harmonisation: ICH E9(R1), adopted 20 November 2019 — estimands, assumptions and sensitivity analysis
Sources checked: 2026-10-04
- BMC Medical Research Methodology: Jackson et al., 2012 — missing-data assumptions in an Internet smoking-cessation trial
Sources checked: 2026-10-04
- NHS Better Health: Ready to quit smoking — local Stop Smoking Service route
Sources checked: 2026-10-04
General research literacy, not an individual smoking judgment, personal forecast or treatment choice. All worked counts are invented; no personal data is requested, saved or sent.