Compare the Published Trial With the Outcome Plan Written Before Results

Outcome switching occurs when the primary result presented in a trial report does not match the outcome specified before investigators knew the results, and the change is not disclosed and justified. The integrity problem is concealment, not the mere fact that a well-founded amendment occurred.

Key points#

• A primary outcome is defined by its domain, measure, metric, aggregation, and time point, not just a short label.
• Prospective registration creates a dated public record, but registry version history is more informative than the latest entry alone.
• Protocols and statistical analysis plans contain detail that a registry may omit.
• Legitimate outcome amendments should be dated, explained, and disclosed in the final report.
• Selective reporting can bias systematic reviews because unavailable null outcomes cannot enter the pooled estimate.

Why prespecification changes the meaning of a result#

Imagine a trial measures pain at four time points using two scales, physical function using three summaries, sleep, mood, medication use, and several subgroups. Even if the intervention has no true effect, one result may look favorable by chance.

A primary outcome chosen before data are examined limits that flexibility. It identifies the main question that guided the sample size and analysis. A primary outcome chosen after seeing results benefits from hidden multiple testing, even if the final paper presents only one p value. That is what prospective registration is for: it puts the planned question on the record before anyone's outcome can influence the choice, and it gives you something to hold the published report against.

An outcome is more than its topic#

“Depression,” “cardiovascular events,” or “quality of life” is not a complete outcome definition. The WHO trial registration data set calls for primary and secondary outcomes with the measurement method and time point.

A precise specification includes:

• domain, such as pain intensity,
• measurement instrument or event definition,
• metric, such as final value, change from baseline, or time to event,
• method of aggregation, such as mean, proportion above a threshold, or hazard ratio,
• and analysis time point or follow-up window.

Changing any of these can change the result. A paper may retain the label “pain” but move from a 12-month mean score to the most favorable weekly change. That is a substantive switch even though the outcome domain sounds unchanged.

Likewise, promoting a prespecified secondary outcome to primary status after results are known changes its evidentiary role, and omitting the original primary outcome, introducing a new one, or changing the threshold for response can all create selective outcome reporting.

Registration is necessary but not sufficient#

ICMJE recommends public registration of clinical trials at or before enrollment of the first participant as a condition of journal consideration; the timing is what gives the record its force. A record first submitted after enrollment or after follow-up began may reflect information already observed.

The latest registry page is not always the original plan. Investigators can update records for legitimate reasons, and registries preserve change histories to varying degrees; find the version that existed before enrollment, then read every later change, paying particular attention to any made after the trial finished or the blind was broken.

Registry entries can also be vague. A full protocol explains design and outcomes in more detail, while a statistical analysis plan defines transformations, covariates, missing-data methods, multiplicity procedures, populations, and analysis models. The most useful audit compares all dated sources:

1. earliest registry record,
2. original protocol and amendments,
3. statistical analysis plan and its version date,
4. final registry history,
5. journal article and supplements,
6. and other reports from the same trial.

The earliest document is not always the most appropriate final plan, and what you are trying to reconstruct is what changed, when, why, and whether anybody could have seen results before deciding.

Changes can be legitimate#

Trials are planned under uncertainty: an instrument may become unavailable, an outcome definition may be updated by consensus, recruitment may make the original analysis infeasible, or external evidence may require a safety change. A blinded review can reveal distributional features that justify a model revision without revealing treatment effects.

A defensible amendment has a dated record, a reason connected to scientific or operational information, identification of who approved it, and an explanation of whether treatment-group results were available. The publication then reports the change and its rationale.

CONSORT 2025 asks trial reports to identify changes to outcomes after the trial began, explain reasons, provide registration details, and indicate where the protocol and statistical analysis plan can be accessed. A transparent amendment lets you judge the consequences for yourself. An undisclosed one takes that judgment away from you.

What empirical comparisons have found#

Research comparing protocols with publications has documented omitted, introduced, and changed primary outcomes, along with more complete reporting of statistically favorable outcomes. The important denominator is not merely how many articles mention registration. It is how many prespecified outcomes can be traced from plan to report.

The COMPare project performed contemporaneous checks of 58 trials published in five journals. Reviewers compared each report with prespecified outcomes and sent correction letters when discrepancies were identified. The project showed that selective reporting could be audited in real time, and that the arguments it provoked kept returning to the same three questions: which registry version counted, whether entries were precise enough, and whether deviations had been disclosed. All three point the same way. A generic registry label offers weak protection; an operationalized outcome with an accessible history offers much more.

Why systematic reviews inherit the problem#

A meta-analysis can combine only outcomes it can obtain. If trials publish favorable measures and withhold unfavorable ones, the available evidence is a selected sample of all measured results. Pooling those results with flawless arithmetic does not remove the selection.

Selective reporting may appear as:

• complete omission of an outcome,
• incomplete numerical reporting that prevents meta-analysis,
• selection of one scale or time point from several,
• reporting only a favorable subgroup,
• or changing a continuous outcome to a threshold that produces a stronger result.

Reviewers can compare protocols and registries, contact study teams, search other reports, and rate risk of bias from missing results. Statistical sensitivity methods may explore the impact, but they cannot recreate every unreported estimate. So the damage does not stop at the credibility of one abstract: a switched outcome can reach a guideline, and from there a policy, by distorting the evidence that was available to synthesize.

A seven-minute outcome audit#

You do not need the raw data for this. You need the dated documents:

Find the trial identity#

Locate the registration number in the abstract, methods, or end matter. Search the title and investigators if the number is absent. Confirm that the registry record describes the same population, treatment groups, and recruitment dates.

Check timing#

Record the first registration date, first participant enrollment date, study completion date, and dates of relevant registry updates. Registration after enrollment deserves special attention.

Capture the original primary outcome#

Use the earliest informative registry version and available protocol. Record the full domain, measure, metric, aggregation, and time point rather than copying only the short label.

Compare the publication#

Identify the outcome called primary in the article. Check whether the original primary outcome appears, whether its status changed, and whether newly emphasized outcomes were prespecified.

Look for amendments#

Read the protocol amendment table, statistical analysis plan history, methods, and trial-registration section. A transparent explanation may resolve an apparent discrepancy.

Check the analysis hierarchy#

Confirm whether the reported result had alpha assigned in the multiplicity plan. An outcome can be prespecified yet still lack confirmatory status if an earlier hypothesis in a fixed sequence failed.

Calibrate the conclusion#

Classify the result as prespecified confirmatory, transparently amended before relevant results were known, or exploratory and post hoc. The estimate may still be useful in every one of those categories. What changes is the strength of the claim you can make with it.

Outcome switching, p-hacking, and publication bias#

These problems overlap but are not identical. Outcome switching changes which measured outcome receives prominence. P-hacking refers more broadly to flexible analyses selected because they produce favorable results. Publication bias concerns which entire studies become available. Selective outcome reporting operates within a study.

Prospective registration helps with all of them only when entries are timely, specific, complete, and checked against reports. Registration is an audit trail, not a certificate that every published claim was prespecified.