Prerequisites
Research protocols, study versus report units, effect estimates, and missing-data mechanisms.
Begin with a prespecified question and a study-level inventory. Compare registrations and protocols with reports. Selectively unavailable results can distort a synthesis; a symmetrical plot or nonsignificant diagnostic does not prove their absence.
Worked example
Three reports from one cohort are not three independent studies. A registered outcome missing from a report is a reason to investigate, not evidence that its effect was zero.
Model and Assumptions
An invented registry lists six independent studies of one educational outcome. Four reported study estimates are 1, 2, 3, and 6. Two estimates are unavailable. For a deliberately simple sensitivity exercise, weight all six studies equally and let each unavailable estimate lie between -4 and 4. This is not an inverse-variance meta-analysis.
Further inquiry
Derive bounds for the six-study average when missing estimates instead lie in [a,b]. Discuss how the conclusions depend on those bounds, equal weighting, and the possibility of unregistered studies.
Review criteria
- Show the interval [(12+2a)/6, (12+2b)/6].
- Do not equate a sensitivity range with sampling uncertainty.
- Define a reproducible study-level search and discrepancy log.