Outcome Switching and Selective Outcome Reporting: Detection and Prevention

Outcome switching is when a trial’s published results report a different primary or secondary outcome than the one prospectively registered before the trial started — adding a new outcome, dropping an original one, or promoting a secondary outcome to primary, usually without disclosing the change or its reason. It is detected by comparing the registry entry (ClinicalTrials.gov, ISRCTN, ANZCTR, or another WHO-recognized primary registry) against the published paper, outcome by outcome. This guide covers what counts as switching versus a legitimate pre-specified amendment, how to check for it as a reviewer or reader, what the standards bodies require, and how to avoid it as a trial team.

What counts as outcome switching

An outcome is switched when the published report and the trial’s own registry entry disagree on what was measured as primary or secondary, and the disagreement isn’t explained. Three patterns cover most real cases:

• Silent omission — an outcome that was registered (often a secondary outcome that didn’t reach statistical significance) simply doesn’t appear in the published results, with no explanation for why it was dropped.
• Silent addition — the paper reports an outcome that was never in the original registration, typically one identified after the data were unblinded.
• Reclassification — a secondary outcome is promoted to primary in the paper (or vice versa), which changes which result the trial is judged to have succeeded or failed on.

None of this is inherently misconduct. Trial protocols can legitimately change — a measurement instrument turns out to be unreliable, a regulator requires an added safety outcome, a pandemic disrupts follow-up. The distinction that matters is disclosure: a legitimate amendment is dated, reasoned, and stated in the paper (this is exactly what CONSORT 2010 item 6b asks trial reports to do — describe any changes to outcomes after the trial began, with reasons). An undisclosed discrepancy that a reader only finds by pulling the registry entry themselves is the failure mode this page is about.

Why it matters: the reporting-bias mechanism

Outcome switching is the results-reporting counterpart to publication bias. Publication bias filters which studies get published at all, based on whether the overall result was positive; outcome switching filters which outcomes get emphasized within a study that does get published, based on which ones came out favorably. A trial that pre-specifies six outcomes, finds one of them significant, and reports that one as the headline finding without mentioning the other five is functionally running a multiple-comparisons test and reporting only the result that cleared the bar — while presenting it as a single, pre-planned confirmatory finding. Readers, and downstream systematic reviews and meta-analyses that pool published outcomes, have no way to detect this without checking the registry.

How to detect it: the registry-comparison method

The check itself is mechanical, which is what makes it something a reviewer, journal editor, or systematic-review team can realistically do on every trial rather than relying on trust:

1. Find the trial’s registration. The registration number should be in the paper’s methods section or abstract (this is what CONSORT 2010 item 24 requires). Search the exact number on the registry it names, or search WHO ICTRP if the registry isn’t stated.
2. Pull the registration history, not just the current record. Registries such as ClinicalTrials.gov keep a version history showing every edit and its date. A registration edited to match the published outcomes after the trial’s primary completion date is a strong signal of retrospective outcome switching, not a legitimate pre-specified amendment.
3. List the registered primary and secondary outcomes exactly as first recorded, then list the outcomes actually reported in the paper’s results section and abstract.
4. Compare the two lists directly. Flag anything registered but missing from the paper, anything reported but never registered, and any outcome whose primary/secondary status changed.
5. Check whether the paper discloses the discrepancy. A trial that explains “the pre-specified primary outcome of X was underpowered after enrollment fell short, so Y was designated primary; this amendment was approved by the trial’s oversight committee on [date], before unblinding” is doing exactly what CONSORT asks. A discrepancy with no mention at all is what to report.

This exact method — registry-to-publication comparison across every trial in a defined set of journals, rather than a spot check — is the approach the COMPare project ran out of the University of Oxford’s Centre for Evidence-Based Medicine (now the Bennett Institute for Applied Data Science) between October 2015 and January 2016. The team checked every trial published in five leading medical journals against its registry entry, logged missing and added outcomes for each one, and submitted a correction letter to the journal whenever a discrepancy was found — then separately tracked which of those letters the journals actually published. The project’s central finding, widely cited in the reporting-bias literature since, was that outcome discrepancies were common across the trials checked, and that journals varied considerably in whether they corrected the record once notified. [REPORTED — the project’s existence, timeframe, institutional home, and methodology are confirmed directly against the project’s own site; this page does not cite a specific discrepancy percentage or letter-publication count, since those weren’t independently re-verified this session — re-confirm the exact figures against a COMPare/BMJ primary source before citing a specific number.]

What the standards require

• ICMJE requires prospective trial registration (before or at first participant enrollment, not at manuscript submission) on a WHO-qualifying registry, which is what makes registry-to-publication comparison possible at all — a trial registered only after results are known gives reviewers nothing to compare against.
• CONSORT 2010 item 6a requires “completely defined pre-specified primary and secondary outcome measures,” item 6b requires disclosing any change to those outcomes after the trial began and why, and item 24 requires the registration number and registry name in the report itself. [REPORTED tier for the specific item numbers — standard, widely-cited CONSORT 2010 structure, not independently re-fetched from consort-statement.org this session after two fetch attempts returned a redirect and a 404; re-verify the exact item numbers against a working copy of the checklist before treating them as certain in a future draft.]
• PROSPERO extends the same logic to systematic reviews and meta-analyses — registering the review’s planned outcomes and analysis before the literature search is completed, so a reviewer switching which outcomes get reported in the final review is detectable the same way.
• Core outcome sets (see CASRAI’s guide on Core Outcome Sets and the COMET Database) address a related but distinct problem: even a trial that reports exactly what it registered can still choose outcomes that don’t match what other trials in the same condition measure, making cross-trial comparison and meta-analysis difficult. A core outcome set doesn’t stop switching, but it narrows the space of outcomes a trial can legitimately choose from in the first place.

What to do as a reviewer, editor, or systematic reviewer who finds a switch

Finding a discrepancy is not the same as proving misconduct, and the response should match that:

• As a peer reviewer: flag the discrepancy to the editor rather than silently accepting the paper’s framing. Ask the authors, through the editor, to explain and date the change, or to report the originally registered outcomes alongside the current ones.
• As a journal editor: CONSORT compliance is a condition of publication at journals that have adopted it, which gives an editor real leverage to require a stated explanation before acceptance — this is the correction-letter mechanism the COMPare project relied on, applied prospectively at the review stage rather than after publication.
• As a systematic reviewer extracting data from a trial for a risk-of-bias assessment: RoB 2’s fifth domain covers exactly this — “bias in selection of the reported result.” A registry-versus-publication mismatch you find while extracting data is directly reportable evidence for that domain judgement, not a separate side note.
• As a reader without reviewer access: a discrepancy is still worth a letter to the editor or a comment through the journal’s post-publication commenting system; several of the corrections publicly documented in the reporting-bias literature originated exactly this way.

How to prevent it as a trial team

• Register the trial prospectively, with the complete outcome set, before enrolling the first participant — not as a formality, but because a complete pre-registration is what makes any later change visible as a change rather than invisible as “how it was always going to be.”
• If an outcome genuinely needs to change after the trial starts, amend the registration with a dated note explaining why, and route that amendment through whatever oversight body governs the trial (a data and safety monitoring board, an ethics committee, or equivalent) before unblinding — the “before unblinding” timing is what separates a legitimate amendment from a result-driven one.
• State the amendment explicitly in the published paper, per CONSORT item 6b, rather than leaving a reader to discover it by cross-checking the registry independently.
• Report every registered outcome in the results — including the ones that didn’t reach significance — rather than reporting only the outcomes that support the paper’s headline conclusion.

Frequently asked questions

Is any difference between the registry and the published paper automatically a problem?

No. A disclosed, dated, reasoned amendment made before unblinding is the system working as intended — CONSORT explicitly asks trials to report changes, not to pretend they never happened. The problem is an undisclosed discrepancy a reader has to find independently.

How is outcome switching different from p-hacking?

They’re related but not identical. P-hacking usually describes flexible analysis choices within a single, disclosed outcome (which subgroup, which covariate adjustment, which cutoff) to reach significance. Outcome switching is specifically about which outcome gets reported as primary or secondary at all, checkable against an external, dated record — the registry — independent of the analysis itself.

Does registering a trial on ClinicalTrials.gov alone prevent this?

Registration makes switching detectable; it doesn’t prevent it on its own. Detection depends on someone actually doing the registry-to-publication comparison — a reviewer, an editor, a systematic reviewer extracting data, or an interested reader — which is why this remains a real, ongoing reporting-quality problem even in a research landscape where prospective registration is now close to universal at major journals.