Publication bias in quit-smoking research: what evidence is actually missing?

01

Three fictional records, one six-month question

All records and arrangements here are invented; no trial identifier, institution or effect is real. Imagine adult studies of a diary added to a brief consultation versus the same consultation alone. You seek the between-group result for no cigarette smoking in the previous seven days, assessed at six months. Your document search ends on 4 October 2026.

Record A says the trial was completed in June 2024 and last updated in August 2026, but you locate neither a results record nor a report. Paper B's methods name the six-month smoking assessment, yet the results you can see provide only four-week confidence. Paper C provides a three-month smoking result; its plans list both three and six months and multiple analyses, without the six-month result in the material read.

These are different gaps, not three negative trials. A needs a linked report/results search and confirmation of status; B needs the target outcome and an explanation; C needs the six-month result and the rationale for what was reported. A three-month result cannot fill a question specifically defined at six months. You do not know an absent effect, its reason for absence or whether every planned assessment took place.

02

Ask what the review searched for and what it could use

Read the review's eligibility definition, information sources and each source's last search date. Journal articles are not the only possible source: registries, supplementary reports and other public records may contain results. More than one report can describe the same study, so additional documents are not automatically additional trials.

Locate the review's account of eligible studies with unavailable results. A registration entry can identify a study without supplying usable outcome data. ‘Not statistically significant’ alone is not a numerical comparison with uncertainty; a confidence score is not an abstinence outcome. Check full reports and linked supplements before calling something absent.

Compare the original plans, later versions and actual methods where available. Prespecification does not prove an outcome was measured; legitimate changes or measurement failures can explain a gap. A discrepancy warrants clarification, not an invented allegation. A defensible summary is ‘the target result was not located in these sources by this date’, with the unresolved documents named—not ‘the missing trials failed’. Missing-evidence concerns apply to a specific synthesis, not a blanket verdict on every result in the review.

[1][3]

03

A funnel plot shows a pattern, not the missing records

The plot places study effect estimates against size or precision. Smaller and larger studies may differ because of selective availability, methods, populations or how the intervention was delivered; the effect measure itself and chance can also create asymmetry. Ask which explanation the review examined, rather than identifying empty space as suppressed studies.

With few studies, tests often have little power. A non-significant asymmetry test does not rule out non-reporting bias; a significant one does not establish it. Selection-model sensitivity analyses explore specified assumptions about availability, not recovered unseen results. Stability under a model is limited by its assumptions and does not certify completeness. Do not manufacture a ‘corrected quit rate’ from the picture.

[1]

04

Evidence availability does not select personal help

This exercise uses public documents, not personal smoking history or symptoms, and does not require contacting participants. For individual support or treatment questions, consult a qualified local professional. NHS information describes UK service routes; readers elsewhere need locally appropriate services, not a care choice derived from an incomplete evidence table.

[4]

05

What to keep in mind

• Name the target result before naming a gap.
• Unavailable does not mean negative or deliberately hidden.
• Funnel asymmetry and model corrections do not recover records.

06

Common questions