Preregistration works by creating a timestamped, durable record of important research intentions before the relevant outcomes or analyses are known, so later readers can distinguish what was planned in advance from what was discovered, changed, added or interpreted after contact with the data.
That definition is deliberately about time.
Preregistration does not magically improve a bad hypothesis.
It does not randomise participants.
It does not blind assessors.
It does not increase sample size.
It does not guarantee that the statistical model is correct.
Its central job is subtler:
preserve the difference between decisions made before the result was visible and decisions made after the result could influence them.
The Center for Open Science describes preregistration as posting a timestamped study plan before data collection or analysis. Current OSF guidance likewise frames it as documenting research questions, methods and analysis plans before analysis so the plan can later be matched transparently against the resulting data and conclusions.
Quick Read
QUESTION → PRIOR KNOWLEDGE → HYPOTHESES → METHODS → SAMPLING → OUTCOMES → EXCLUSIONS → ANALYSIS PLAN → TIMESTAMPED REGISTRATION → DATA / ANALYSIS → DEVIATIONS → RESULTS → EXPLORATORY FINDINGS → REPORT → LINK BACK TO PLAN
The central distinction is:
- Confirmatory: the important hypothesis and analysis were specified before the relevant result was known.
- Exploratory: the pattern, subgroup, model or hypothesis emerged after examining the data.
Both can be scientifically valuable.
Preregistration protects the label.
1. Preregistration Is a Timeline Device
Suppose a study measures twenty outcomes.
After seeing the results, researchers choose the strongest one and write:
We hypothesised that the intervention would improve Outcome 17.
Without an earlier record, readers may have no way to know whether Outcome 17 was genuinely the original prediction or selected because it looked impressive.
A preregistration creates evidence about the order of decisions.
It preserves what the researchers said they would do before the relevant data could answer back.
2. Preregistration Does Not Ban Exploration
This misconception damages research.
Unexpected patterns are often the beginning of discovery.
A surprising subgroup may reveal a mechanism. An anomaly may expose a new variable. A model may fail in a way that generates a better hypothesis.
Preregistration does not say “do not look”.
It says:
When you discover something after looking, call it discovery rather than pretending it was an independent prediction before looking.
3. Confirmatory and Exploratory Research Need Each Other
Exploration generates possibilities.
Confirmation subjects selected possibilities to stronger independent tests.
A healthy research programme loops:
EXPLORE → DISCOVER → FORMALISE → PREREGISTER → TEST → UPDATE → EXPLORE AGAIN
The failure is not exploration.
The failure is laundering exploration into confirmation.
4. A Preregistration Should Name the Research Question
“We will study learning” is too vague.
A useful preregistration defines the target question closely enough that later readers can determine whether the final paper answers the same question.
Who is the population?
What exposure or intervention?
What outcome?
What comparison?
What time horizon?
A vague plan can be technically preregistered while preserving enough flexibility to fit almost any result.
5. Hypotheses Need Direction and Conditions Where Theory Justifies Them
A preregistered hypothesis should state what relationship is expected and under what conditions.
If theory predicts a direction, say so.
If theory only predicts a difference, do not invent direction for dramatic effect.
If several hypotheses are primary, distinguish them from secondary or exploratory expectations.
See How Hypotheses Work for the underlying mechanism.
6. Outcomes Should Be Defined Before Their Results Are Known
A study can measure immediate recall, delayed recall, transfer, confidence, completion time and satisfaction.
Which is primary?
If researchers decide after seeing the results, the strongest outcome can become the story while weaker outcomes disappear.
Preregistration preserves which outcomes were intended to bear the main inferential weight.
Additional outcomes can still be reported honestly as secondary or exploratory.
7. Operational Definitions Belong in the Plan
“Accuracy” might mean proportion correct.
“Engagement” might mean time-on-task.
“Improvement” might mean post-test minus pre-test.
If these definitions remain vague until after results are visible, researchers can unconsciously choose the representation that makes the pattern look strongest.
Preregistration therefore connects directly to Measurement and Research Variables.
8. Exclusion Rules Should Be Defined Before Outliers Become Inconvenient
Which participants are eligible?
What counts as failed attention check?
What response time is impossible?
Will participants with incomplete data be excluded?
Outlier decisions can substantially alter results.
Pre-specifying objective rules reduces the temptation to let treatment direction determine which observations look unreasonable.
9. The Analysis Plan Should Be Specific Enough to Reconstruct the Main Test
“We will analyse the data statistically” is not an analysis plan.
A useful plan may specify:
- the primary model;
- outcome and predictor coding;
- covariates;
- interaction terms;
- transformations;
- missing-data handling;
- multiple-testing correction;
- decision threshold;
- and robustness analyses.
The exact resolution should match the method. Complex qualitative, computational and exploratory projects may require different templates from conventional hypothesis-testing experiments.
10. Sample Size and Stopping Rules Belong in the Plan
How many observations will be collected?
Why that number?
Will recruitment stop on a calendar date, at a target n, after a sequential boundary or when a resource limit is reached?
If researchers repeatedly inspect p-values and stop when significance appears, the inferential procedure changes.
Preregistering a valid stopping rule preserves the actual sampling design.
See How Statistical Power Works.
11. Preregistration Should Happen Before the Relevant Result Is Knowable
For newly collected data, preregistration is ordinarily done before data collection or before analysis, depending on the design and registration claim.
For secondary data, the important question is whether researchers have already accessed or examined the data in a way that could reveal the relevant result.
One cannot recover genuine prospective independence by writing a plan after knowing what the dataset contains.
The timestamp matters because it anchors the plan to an information state.
12. Existing Data Can Still Be Preregistered Under the Right Information Boundary
Secondary-data research is essential.
A researcher may preregister an analysis of an existing dataset if the relevant results have not already been examined or can be protected through appropriate held-out data or restricted access.
The preregistration should be transparent about what is already known.
The question is not “Are the files old?”
It is “Could the researchers’ current knowledge already have been shaped by the outcomes they are about to test?”
13. A Registration Is a Durable Snapshot
OSF registrations create timestamped snapshots of a project state.
Current OSF documentation describes registrations as formal, transparent records of study plans. The registered version preserves the state at submission so the ongoing project can continue changing while the earlier snapshot remains recoverable.
This separation between live project and preserved snapshot is the core architecture.
A research plan can evolve.
The original plan does not need to disappear.
14. Embargoes Allow Timing Evidence Without Immediate Public Disclosure
Researchers can have legitimate reasons not to reveal a plan immediately.
They may fear being scooped, need blinded peer review, handle commercially sensitive work or require time before public release.
OSF supports embargoed registrations, preserving the timestamp while delaying public visibility for a defined period.
Transparency does not always require immediate universal visibility.
It requires a recoverable record whose timing and later release can be audited.
15. Preregistration Is Not a Prison
Reality breaks plans.
A recruitment site closes.
A variable is badly distributed.
A software bug is discovered.
A measurement fails.
An assumption is violated.
Researchers should change the plan when the original plan becomes scientifically wrong.
The requirement is honesty about the change.
Deviate when science requires it. Preserve the fact that you deviated.
16. Deviations Should Be Versioned With Reasons
A useful final report can state:
- what the original plan specified;
- what changed;
- when it changed;
- why it changed;
- whether the change occurred before or after relevant outcomes were visible;
- and how the deviation affects interpretation.
This is scientifically stronger than following a broken plan merely to claim perfect compliance.
Preregistration values inspectability, not obedience theatre.
17. Undisclosed Deviations Are the Real Problem
Suppose a primary outcome proves noisy, so researchers switch to another outcome that shows a clear effect.
That switch may be scientifically sensible.
If the paper presents the new outcome as though it were always primary, readers overestimate the independence of the evidence.
If the paper says the original outcome failed and the new outcome was selected after inspection, the finding remains useful but is correctly labelled as less confirmatory.
18. Preregistration Reduces HARKing
HARKing means hypothesising after the results are known while presenting the hypothesis as though it came before them.
A result appears.
A plausible explanation is constructed afterward.
The explanation is then narrated as prediction.
Preregistration makes the original hypothesis recoverable and therefore makes retrospective storytelling easier to detect.
19. Preregistration Reduces Outcome Switching
If primary and secondary outcomes are declared beforehand, readers can compare the final paper with the original plan.
This reduces the ability to quietly promote a favourable secondary outcome after the original primary outcome disappoints.
The same logic applies to time points.
If a treatment is tested at 1 week, 1 month, 3 months and 6 months, the time point should not become “primary” only because it looked best.
20. Preregistration Reduces Analysis Shopping
Many datasets permit several reasonable analyses.
Raw outcome or transformed outcome?
With covariate or without?
Exclude outlier or retain it?
One interaction or another?
If the analysis is chosen because it produces the desired result, ordinary inferential guarantees no longer match the reported procedure.
A preregistered primary analysis preserves a clean confirmatory route while still allowing alternative models to be reported as sensitivity or exploratory analyses.
21. Preregistration Makes Null Results Easier to Interpret
A null result is much more informative when readers know the study genuinely planned to test that question, collected the intended outcome, followed the declared analysis and had adequate power.
Without a plan, readers may wonder whether the null analysis was chosen after more interesting results failed elsewhere.
Preregistration therefore increases the visibility of negative and inconclusive evidence even before publication incentives are addressed.
22. Preregistration Does Not Solve Publication Bias by Itself
A preregistered study can still remain unpublished if the results are disappointing.
The registration makes the existence of the planned study more discoverable once public, but it does not force a journal to publish the completed results.
Registered Reports go further by linking peer review and in-principle publication commitment to the study plan before results are known.
23. Preregistration and Registered Reports Are Not the Same
| Mechanism | Main job |
|---|---|
| Preregistration | Timestamp and preserve the research plan before the relevant result is known |
| Registered Report | Peer-review the question and methods before results and provisionally commit to publication based on methodological quality rather than outcome |
In May 2026, Nature announced an expansion of Registered Reports across the fields it publishes and beyond confirmatory hypothesis testing to additional research types. The logic is precisely this upstream shift: evaluate importance and methodological strength before knowing whether the result is exciting.
24. Registered Reports Turn Peer Review Into Design Review
Conventional peer review often arrives after data collection is complete.
A reviewer can identify an underpowered design but cannot go back in time and recruit more participants.
A reviewer can identify an invalid primary outcome but cannot reconstruct data that were never collected.
Registered Reports move expert criticism upstream, when methods can still change.
25. Preregistration Does Not Guarantee a Good Study
You can preregister a confounded observational design.
You can preregister an underpowered experiment.
You can preregister an invalid measure.
You can preregister a statistical model that does not match the data.
Preregistration makes the plan visible.
It does not make the plan wise.
This is why preregistration should be paired with good design, domain knowledge and appropriate peer or methodological review.
26. Preregistration Does Not Prove Researchers Followed the Plan
A registration preserves what researchers intended.
The final report must still be compared with that plan.
Compliance is an empirical question.
A preregistered badge or link should therefore invite inspection rather than end it.
Open science becomes meaningful when the plan, data, code and publication can be connected across the lifecycle.
27. Preregistration Does Not Guarantee Transparency if the Plan Is Vague
“We will analyse the effect of treatment on outcomes using appropriate statistics” gives enormous freedom.
A timestamp makes the sentence old.
It does not make it informative.
The preregistration needs enough specificity that readers can distinguish adherence, deviation and exploration.
28. Over-Specification Can Also Become Theatre
A forty-page preregistration full of details nobody can interpret is not necessarily better than a shorter, precise plan.
Researchers can also create excessive decision trees covering every possible outcome, leaving so much flexibility that the registration ceases to constrain interpretation.
The right level of resolution depends on the research design.
The plan should clarify the decisions whose timing matters to evidential interpretation.
29. Qualitative Research Can Preregister Without Pretending to Be Confirmatory Quantitative Research
Qualitative studies can document sampling logic, research questions, interview procedures, analytic framework, reflexive assumptions and intended stopping considerations.
The purpose is not to freeze themes before fieldwork.
It is to preserve what was planned and make methodological evolution visible.
Different methods need different preregistration grammars.
30. Exploratory Data Analysis Can Be Preregistered at the Level of Procedure
A researcher may not know which pattern will emerge.
They can still preregister the exploration procedure: which variables, which transformations, which clustering or dimensionality-reduction methods, which validation steps and how discoveries will be separated from confirmatory follow-up.
Preregistration is not synonymous with predicting the answer.
It can preserve a method for searching.
31. Computational Research Can Register Code and Analysis Intentions
Machine-learning studies can preregister datasets, splits, preprocessing, metrics, hyperparameter search limits, exclusion criteria and final test procedures.
This can reduce benchmark overfitting and metric shopping.
But if a benchmark has already been used repeatedly by the team, formal preregistration cannot make it an untouched test set again.
Information exposure leaves history.
32. Preregistration Is Especially Useful Where Analytical Flexibility Is High
A study with one obvious outcome and one analysis may gain modestly from preregistration.
A study with many outcomes, subgroups, preprocessing choices and model specifications has a much larger researcher-decision space.
Preregistration can create a clean primary path through that garden while preserving the rest for transparent exploration.
33. Preregistration Helps Meta-Research Reconstruct Invisible Studies
If registrations are public or later released, researchers studying the scientific system can compare planned studies with published studies.
Which studies disappeared?
Which outcomes changed?
Which hypotheses were reported?
This turns otherwise invisible publication and reporting bias into something more measurable.
34. Preregistration and Blinding Control Different Information Flows
Blinding asks who should not see treatment information during the study.
Preregistration asks which decisions should be recorded before researchers can see the results that might influence those decisions.
One is role-based information control.
The other is time-based information control.
35. Preregistration and Randomisation Control Different Failure Modes
Randomisation separates treatment assignment from baseline prognosis.
Preregistration separates important analytic choices from outcome knowledge.
A randomised trial can still outcome-switch.
A preregistered observational study can still have confounding.
Strong research uses complementary controls rather than expecting one method to solve every problem.
36. Preregistration and Power Planning Reinforce Each Other
A power calculation contains assumptions about effect size, sample size, variance, outcome and analysis.
Preregistering those assumptions makes it easier to see whether the final study used the design that justified its sample size.
If researchers switch to a much noisier outcome or different model, the original power argument may no longer apply.
Design logic should remain connected to analysis logic.
37. Preregistration and Peer Review Should Not Be Confused
A preregistration can be posted without peer review.
That means no expert has necessarily evaluated whether the plan is sensible.
A Registered Report adds peer review of the plan.
The difference matters because timestamping proves timing, not methodological quality.
38. Preregistration Does Not Prevent Fraud
A dishonest researcher can preregister and later fabricate data.
They can misreport deviations.
They can run unreported analyses.
Preregistration changes the audit surface.
It does not abolish the need for data provenance, code review, replication, peer scrutiny and research integrity systems.
39. The Hostile Test: Preregistered After Looking
A researcher examines the dataset, finds a striking association, then writes a detailed preregistration and performs the same analysis again.
The registration is timestamped.
The plan is specific.
The confirmatory independence is gone because the researcher already knew the pattern.
The correct label is replication or confirmation on independent data—not “preregistered discovery” in the same dataset.
40. The Second Hostile Test: A Vague Preregistration
The plan says:
We predict meaningful relationships among the main variables and will use appropriate statistical analyses.
The sentence is timestamped.
Almost any later result can fit it.
Preregistration works only when the plan narrows the decision space enough to distinguish planned tests from after-the-fact choices.
41. The Third Hostile Test: Blindly Following a Broken Plan
A preregistered analysis assumes normality.
The collected outcome is severely zero-inflated.
The original model is inappropriate.
Running it anyway is not scientific virtue.
The stronger response is to document the problem, explain the deviation, use an appropriate model and distinguish the revised analysis from the original confirmatory plan.
42. The Fourth Hostile Test: Preregistered Primary Outcome Quietly Disappears
The registration names delayed retention as primary.
The final paper highlights immediate confidence because delayed retention shows no effect.
The paper never mentions the change.
The registration itself is fine.
The reporting violates the transparency contract.
43. The Fifth Hostile Test: Registration Used as a Prestige Badge
A paper says “preregistered” prominently.
Readers stop asking questions.
The sample is biased, measure weak and analysis inappropriate.
Preregistration solved none of those problems.
Open-science signals should open inspection, not close it.
44. Primary School: Preregistration Begins as “Write What You Think Before You Look”
A child predicts which paper aeroplane will fly farther.
Before throwing them, the child writes:
I think Plane A will fly farther because its wings are wider.
After the throws, Plane B wins.
The child does not rewrite the original prediction.
They update it.
That is the seed of preregistration: preserve what you thought before the world answered.
45. Secondary School: Separate Prediction From Discovery
Students can write their hypothesis, variables, sample plan and main analysis before collecting results.
Then, after analysis, they can add a separate section:
Unexpected pattern discovered after looking at the data.
This teaches that exploration is legitimate when its timing is honest.
46. JC and University: Preregistration Becomes Decision Provenance
At higher levels, preregistration should preserve enough detail to reconstruct important decisions:
- question;
- hypotheses;
- sampling;
- outcomes;
- variables;
- exclusions;
- sample size;
- stopping;
- analysis model;
- multiplicity;
- missing data;
- and robustness analyses.
The preregistration becomes a provenance record for reasoning rather than an administrative form.
47. Where Preregistration Fits in the eduKateSG “How Works” Landscape
- How Hypotheses Work — what predictions exist before testing.
- How Research Bias Works — selective reporting and analytic flexibility that preregistration can reduce.
- How Statistical Power Works — the sample-size assumptions that can be preserved before data collection.
- How Randomisation Works — assignment control rather than decision-timing control.
- How Blinding Works — role-based information control.
- How Peer Review Works — expert scrutiny, including Registered Reports.
- How Scientific Research Works — the full evidence lifecycle.
Preregistration owns one precise job: make the temporal provenance of research decisions visible.
48. What This Article Does Not Claim
- Preregistration does not make a weak study automatically rigorous.
- Preregistration does not ban exploratory analysis.
- Preregistration does not require researchers to follow an invalid plan after reality reveals a problem.
- Justified deviations are not misconduct when they are disclosed transparently.
- Preregistration does not replace peer review.
- Preregistration does not replace randomisation, blinding, adequate power or valid measurement.
- A timestamp alone is not useful if the plan is too vague to distinguish decisions.
- Preregistering after seeing the relevant result does not restore prospective independence.
- Preregistration does not guarantee publication of null results.
- Registered Reports are a related but stronger publishing design, not a synonym for preregistration.
49. A Compact Preregistration Audit
- What information was known when the plan was registered?
- Was the relevant data already collected?
- Had researchers already examined the outcomes?
- What is the primary research question?
- What hypotheses are confirmatory?
- Which outcomes are primary?
- How are variables operationalised?
- What sampling and eligibility rules apply?
- What exclusion rules are specified?
- What sample-size and stopping rules apply?
- What primary analysis is specified?
- What covariates and interactions are included?
- How are missing data handled?
- How is multiplicity handled?
- Which analyses are explicitly exploratory?
- Is the registration public or embargoed?
- What deviations occurred?
- When did each deviation occur?
- Why was it necessary?
- Does the final report link clearly back to the original plan?
50. Frequently Asked Questions
What is preregistration?
Preregistration is the practice of creating a timestamped record of a research plan before the relevant data or analyses are known, preserving the timing of hypotheses, methods, outcomes and analysis decisions.
Can I change a preregistered plan?
Yes. Research plans sometimes need to change. The scientifically important step is to preserve the original registration and report important deviations, timing and reasons transparently rather than rewriting history.
Does preregistration stop exploratory research?
No. It helps distinguish analyses planned before results from analyses inspired by the observed data. Exploratory findings can remain valuable and can generate future confirmatory studies.
Can existing data be preregistered?
Yes in appropriate circumstances, but researchers should state what they already know and whether the relevant outcomes have been examined. Prospective confirmatory claims require an information boundary that prevents result knowledge from shaping the plan.
What is the difference between preregistration and a Registered Report?
Preregistration timestamps a study plan. A Registered Report additionally sends the plan through journal peer review before results and can provide in-principle acceptance based on the question and methods rather than the eventual outcome.
51. Authoritative Research Corridor
- OSF Support — Registrations and Preregistrations
- OSF Support — Simplifying the Preregistration Process, updated August 2026
- Center for Open Science — Lifecycle Open Science
- Center for Open Science — OSF registration infrastructure
- Nature — Expanding Registered Reports Across Fields, 27 May 2026
- Scientific Reports — Registered Reports policy and two-stage review
Final Thought: Preregistration Preserves the Before
After a result appears, the path toward it starts to feel inevitable.
Of course that variable mattered.
Of course that subgroup was interesting.
Of course that model was the sensible one.
Human memory is generous to the present.
Preregistration leaves a receipt from the past.
This is what we thought.
This is what we planned.
This is what reality changed.
This is what we discovered afterward.
The value is not rigidity.
It is provenance.
Good science should be allowed to change its mind. Preregistration makes sure it does not quietly change its memory.