Preregistration in Medical Research: What It Can Prevent—and What It Cannot

TLDR

Preregistration in medical research creates a time-stamped record of what researchers intend to study and how they intend to analyze it. Done before enrollment or before researchers can inspect the relevant data, it helps readers identify outcome switching, selective reporting, and analyses chosen after results were known. It does not guarantee that the research question is important, the design is sound, the statistical methods are appropriate, or the conclusions are reproducible.

The practical test is not simply whether a study was preregistered. Ask when the plan was registered, how much detail it contained, whether later changes are visible and justified, and whether the final report clearly separates prespecified analyses from exploratory findings.

What preregistration in medical research means

Preregistration is the act of publicly recording a research plan before a defined point in the research process. The plan can identify the hypothesis, population, intervention or exposure, comparison group, outcomes, measurement times, sample-size reasoning, and statistical methods. A time stamp makes it possible to compare what investigators planned with what they eventually analyzed and reported.

For an interventional clinical trial, prospective registration ordinarily means creating the public record before the first participant is enrolled. The International Committee of Medical Journal Editors requires registration at or before the first patient gives consent for enrollment as a condition of consideration by its member journals. The World Health Organization likewise defines prospective registration around registration before the first participant is recruited. Read the ICMJE clinical-trial registration policy.

That timing matters because a hypothesis recorded after investigators have viewed outcomes does not provide the same evidence of advance planning. It may still be scientifically useful, but the record can no longer show that the hypothesis was independent of the observed result.

Preregistration, trial registration, protocols, and analysis plans are related but different

These terms are sometimes treated as interchangeable, but each document has a different job. A strong transparency record may include all four.

Document or practice	Main purpose	Typical level of detail
Preregistration	Time-stamps hypotheses, outcomes, and planned methods before a relevant milestone	Varies from a concise plan to a detailed record
Clinical-trial registration	Creates a structured public record of an interventional study	Core design, intervention, eligibility, outcomes, dates, and administrative details
Study protocol	Describes why and how the study will be conducted	Broad operational, clinical, ethical, and methodological detail
Statistical analysis plan	Defines how study data will be transformed and analyzed	Models, analysis populations, missing data, multiplicity, sensitivity analyses, and related decisions

A registry entry may contain only structured summary fields. It should not automatically be treated as a substitute for a complete protocol or statistical analysis plan. Protocol-reporting guidance calls for substantially more detail about trial design, outcomes, methods, and planned analyses.

Requirements also depend on context. Journal policy, funder policy, national law, and registry rules are not identical. ICMJE describes registration expectations for journal consideration, while the National Institutes of Health describes registration and results-reporting obligations applying to NIH-funded clinical trials.

What should be prespecified?

Prespecified means decided and documented before investigators cross the relevant information boundary—usually enrollment or access to outcome data. The plan needs enough detail to make alternative analytical choices visible. Merely writing “we will compare the groups” leaves too many decisions open.

Research question and hypothesis: What effect, association, or difference is being tested?
Population: Who is eligible, who is excluded, and which participants will enter each analysis?
Intervention or exposure: What exactly is being assigned or measured?
Comparator: Placebo, usual care, another treatment, no exposure, or a different exposure level?
Outcomes: What variables will be measured, by which method, at what time points, and using what summary measure?
Sample size: What assumptions, target effect, variability, error rates, or feasibility constraints informed the planned number?
Primary analysis: Which groups, model, covariates, effect measure, and uncertainty interval will be used?
Missing data: How will missing observations, withdrawals, and incomplete measurements be handled?
Multiplicity: How will multiple outcomes, time points, treatment groups, or repeated tests affect error control?
Additional analyses: Which subgroup, sensitivity, per-protocol, interim, or adjusted analyses are planned?

International guidance says analyses should follow prospectively defined plans and departures should be identified in the study report. Statistical guidance also emphasizes advance decisions about core analytical features, while modern protocol guidance calls for planned handling of missing data and additional analyses.

Primary, secondary, and exploratory outcomes

An outcome label is meaningful only when the outcome is precisely defined. “Pain,” for example, could mean a score at four weeks, change from baseline at 12 weeks, the proportion achieving a threshold improvement, or time until rescue medication. These are different tests and may produce different results.

Primary outcomes

The primary outcome is intended to answer the study’s main question and commonly drives the sample-size calculation and central statistical test. A study may have more than one primary outcome, but multiple primary tests create additional opportunities for a chance-positive finding and therefore require an explicit analytical strategy.

Secondary outcomes

Secondary outcomes address additional questions, such as functional status, adverse events, quality of life, or a different follow-up time. They can be important, but readers should not automatically give a favorable secondary outcome the same evidential weight as a prespecified primary outcome—especially when many outcomes were tested.

Exploratory outcomes and analyses

Exploratory work looks for patterns that can generate new hypotheses. Preregistration does not prohibit it. The key is honest labeling: an analysis developed after looking at the data should not be presented as though it were the study’s original confirmatory test. Reproducibility guidance treats preregistration as a way to distinguish confirmatory analyses from hypothesis-generating exploration, not as a ban on discovery.

What preregistration can help prevent

Preregistration does not physically prevent researchers from changing an outcome or model. It changes what readers can detect. When the original record and its version history are accessible, unnoticed changes become harder to pass off as original plans.

Outcome switching: replacing or redefining the original primary outcome after results are known.
Selective outcome reporting: publishing favorable outcomes while omitting unfavorable or inconclusive ones.
Undisclosed analytical flexibility: trying many models, exclusions, time points, or subgroups but reporting only the preferred result.
Post hoc hypotheses presented as a priori: describing a data-inspired explanation as if it had been predicted in advance.
Selective emphasis: making a secondary or exploratory result the headline after the primary analysis disappoints.

The concern is empirical, not merely theoretical. A cohort study comparing randomized-trial protocols with publications found incomplete reporting and discrepancies between prespecified and published outcomes. That work helped demonstrate why access to original plans matters when evaluating trial reports.

A separate cross-sectional analysis of ClinicalTrials.gov records reported that 30.3% of studies with a registered start date had a primary-outcome change recorded after that date. This percentage identifies changes in registry histories; it does not establish that every change was improper, made after results were seen, or intended to bias reporting.

Changes are allowed, but they should leave a trail

Medical research rarely unfolds exactly as expected. Recruitment can be slower than planned, an assay may fail, clinical practice may change, or an external safety finding may require a revised procedure. A rigid ban on amendments could make a study less ethical or less informative.

The transparent response is to preserve the original plan and document the amendment. A useful amendment record states what changed, when it changed, why it changed, who approved it, and whether investigators had access to relevant outcome data at the time. The final paper should then identify material departures from the prospectively defined analysis plan.

Timing changes their interpretation. An outcome correction made before enrollment because a registry field was entered incorrectly is different from a new endpoint adopted after unblinded results became available. Both may be legitimate, but readers need dates and reasons to judge the risk of result-driven decision-making.

What preregistration cannot guarantee

Preregistration is a transparency tool, not a quality certificate. A plan can be registered on time and still contain a weak hypothesis, unreliable measurements, inadequate sample size, inappropriate model, poor comparator, or clinically irrelevant endpoint.

Sound design: Registration does not create randomization, conceal allocation, prevent attrition, or eliminate confounding.
Appropriate analysis: A prespecified statistical method can still be poorly chosen or incorrectly implemented.
Clinical importance: A statistically significant change in a biomarker may not improve symptoms, function, survival, or quality of life.
Complete reporting: Authors can still describe methods or harms inadequately, although a detailed public plan makes omissions easier to identify.
Freedom from bias: Conflicts of interest, measurement bias, missing data, protocol nonadherence, and publication bias require separate assessment.
Replication: A time-stamped plan does not ensure that another study will obtain the same result.
Truth: Prespecification reduces some opportunities for result-driven choices; it cannot establish that a conclusion is correct.

Nor does preregistration turn every prespecified subgroup finding into strong evidence. Subgroup credibility also depends on factors such as biological plausibility, adequate sample size, a direct statistical test of interaction, and whether the subgroup was one of many examined. “Prespecified” answers an important timing question, but not every question about validity.

Can observational and secondary-data studies be preregistered?

Yes. Researchers can preregister cohort, case-control, cross-sectional, diagnostic, and secondary-data analyses. The relevant deadline may differ from a clinical trial’s enrollment-based deadline. For an existing database, the crucial point is often before the analyst accesses outcomes, runs the focal models, or otherwise learns the patterns that could influence analytical choices.

That boundary must be described honestly. Researchers may already know the dataset, have conducted preliminary cleaning, or have seen descriptive statistics. A useful record explains what information was available at registration. If the data have already been extensively explored, a registered analysis can still organize subsequent work, but it should not be portrayed as fully independent confirmation.

How to compare the registration with the published paper

Do not stop at the statement “this study was preregistered.” Open the record, check its history, and compare it with the paper using the following sequence.

Check the dates. Was the record created before enrollment or before access to the relevant data? Look for later updates, not just the most recent version.
Match the primary outcome. Compare the variable, measurement instrument, time point, scoring rule, and summary measure—not merely the broad outcome name.
Compare analysis populations. Did the paper use all randomized participants, treated participants, protocol-adherent participants, or a newly defined subset?
Compare the statistical model. Look for changes in covariates, transformations, exclusion rules, missing-data methods, and significance thresholds.
Count the tests. Consider all outcomes, time points, subgroups, and models, not only the analysis highlighted in the abstract.
Look for amendments and deviations. Check whether changes are dated, explained, and acknowledged in the publication.
Separate confirmation from exploration. Give the prespecified primary analysis the clearest confirmatory interpretation; treat data-inspired findings as hypotheses needing further testing.
Assess ordinary study quality. Examine randomization, blinding, measurement validity, missing data, effect magnitude, uncertainty, harms, and applicability.

If a registry summary is vague, look for a dated protocol and statistical analysis plan. Their absence does not automatically invalidate the study, but it limits how confidently a reader can reconstruct which decisions came before the results.

Frequently asked questions

Does preregistration prevent p-hacking?

It can deter or expose some forms of p-hacking by recording analytical choices in advance. It cannot stop undisclosed analyses, guarantee adherence, or eliminate judgment calls that were left vague in the plan. Detail and version history matter.

Is retrospective registration useless?

No. A late record can improve discoverability and disclose that a study exists. It cannot provide the same protection against result-informed choices as a record created before enrollment or data access, so it should be identified as retrospective rather than treated as prospective evidence.

Are researchers allowed to run analyses they did not preregister?

Yes. Exploratory analyses can reveal unexpected safety signals, generate mechanisms, and suggest questions for future studies. They should be labeled as exploratory, with enough information about the number and nature of analyses to interpret chance findings.

Does an outcome change prove research misconduct?

No. Outcomes may change for defensible scientific, operational, or ethical reasons. The important questions are when the change occurred, why it occurred, whether relevant results were already known, and whether the change was disclosed.

Is a registered study automatically more trustworthy than an unregistered one?

Not automatically. Prospective, detailed registration provides useful evidence of transparency, but trust also depends on design, conduct, measurement, analysis, reporting, effect size, uncertainty, and consistency with the broader evidence.

The practical takeaway

Treat preregistration as an audit trail. Its value lies in showing which questions and analyses came before the results, which arose afterward, and where the study departed from its original plan.

When evaluating a medical study, ask four questions first: Was the plan registered early enough? Was it specific enough to constrain analytical flexibility? Are amendments visible and justified? Does the final report clearly distinguish prespecified findings from exploration? Then assess the study’s design and results on their own merits. A transparent plan makes research easier to evaluate, but it cannot make weak research strong.