Science Education Systems · Article 79. Maya, Jia Jun, Hana and Ethan remain fictional Punggol learners. This article follows the preregistration layer: how Science protects itself from hindsight by writing down key decisions before the results are visible.
The 50-second parent route
Preregistration does not ban exploration.
It separates what was planned before the data from what was discovered after looking.
The route is:
question → hypothesis → primary outcome → design → sample-size logic → exclusions → analysis plan → timestamped registration → data → deviations → results → interpretation
The key question is:
Which decisions were made before the answer was known?
This article extends How Scientific Hypothesis Testing Works, How Scientific Multiple Comparisons Work and How Scientific Bias Works.
1. Preregistration is a time-order tool
It places the research plan before the outcome.
That sequence matters because knowledge of the result can change analytical choices.
2. The main purpose is transparency
What did the researchers intend to test?
What changed later?
Which analyses were exploratory?
Preregistration makes those distinctions visible.
3. Maya’s preregistration error is thinking the plan can never change
Unexpected measurement failure occurs.
She believes the team must follow the original plan blindly.
Her repair:
change when scientifically necessary, document the deviation, and explain why.
4. Jia Jun’s preregistration error is writing a vague plan
“We will analyse the data appropriately.”
This preserves nearly unlimited flexibility.
His repair:
specify enough detail that another researcher can see what was planned.
5. Hana’s preregistration error is treating exploration as bad
She discovers an unexpected pattern and refuses to discuss it.
Her repair:
label it exploratory and propose a new confirmatory test.
6. Ethan’s preregistration error is registering after looking
He examines the outcomes first, then timestamps a plan matching what he found.
His repair:
the registration must precede access to the relevant results if it is to separate prediction from hindsight.
7. A useful preregistration begins with the research question
Who?
what?
compared with what?
measured how?
over what time?
Question clarity constrains later flexibility.
8. Hypotheses should be testable
State the expected direction or relationship where scientifically justified.
Define what evidence would count against the hypothesis.
9. Primary outcomes should be identified
If twenty outcomes are measured but one is declared primary only after results are known, multiplicity becomes hidden.
Preregistration preserves the original priority.
10. Secondary outcomes can still be important
The point is not to ignore them.
It is to distinguish the main confirmatory question from additional analyses.
11. Sample-size logic should be recorded
Target effect.
expected variability.
power.
available resources.
ethical constraints.
The planned sample should have a reason.
12. Stopping rules matter
Will recruitment end at a fixed sample?
after a fixed time?
under a planned sequential rule?
Unplanned stopping after significance appears changes statistical error behaviour.
13. Inclusion criteria should be prespecified
Which participants, specimens or observations qualify?
Criteria written afterward can become a way to remove inconvenient data selectively.
14. Exclusion criteria should be prespecified too
Instrument failure.
protocol violation.
impossible value.
duplicate record.
The rules should be scientifically motivated and transparent.
15. Outlier handling belongs in the plan
Will outliers be retained?
excluded under a defined rule?
analysed with robust methods?
Several choices may be defensible; the problem is choosing based on the desired result.
16. Variable definitions belong in the plan
What exactly is “improvement”?
What time window defines the outcome?
How is exposure coded?
Operational definitions reduce post-hoc ambiguity.
17. Transformations should be anticipated where possible
Raw scale.
log scale.
percentage change.
standardisation.
Transformations alter the analytical question.
18. Covariates should be justified
Which variables will be adjusted for?
Why?
Adding or removing covariates after seeing significance can create hidden analytical flexibility.
19. Statistical tests should be named where practical
Model family.
comparison.
contrast.
error criterion.
Preregistration works best when the planned analysis can be reconstructed.
20. Multiplicity control belongs upstream
If many outcomes or comparisons are planned, define the hypothesis family and error-control strategy before results appear.
21. Primary Science can learn preregistration as prediction-before-test
Before running the experiment, students write:
what they expect.
why.
what they will measure.
what will stay the same.
Then they collect data.
22. Primary 3 can write one prediction first
Which object will sink?
Write the prediction.
Then test.
This separates prior reasoning from hindsight.
23. Primary 4 can define the fair-test plan
What changes?
what is measured?
what stays constant?
Write it before the first trial.
24. Primary 5 can define what counts as success
Instead of deciding afterward that “the plant looked healthier,” specify height increase, leaf count or another observable measure in advance.
25. Primary 6 can record deviations
The thermometer broke.
one trial was interrupted.
Students note the change instead of hiding it.
Scientific transparency becomes normal.
26. Secondary Science can formalise preregistration
hypotheses.
outcomes.
sample-size plans.
exclusions.
analysis models.
Students can distinguish confirmatory and exploratory reasoning.
27. Preregistration reduces HARKing
Hypothesising After Results are Known makes a discovered pattern look predicted.
A timestamped plan reveals which hypotheses existed beforehand.
28. Preregistration reduces hidden p-hacking
If the main analysis was prespecified, readers can see when alternative models were introduced later.
This does not prevent every bad practice but increases visibility.
29. Preregistration does not guarantee a good study
A badly designed study can be preregistered perfectly.
Transparency does not substitute for scientific quality.
30. Preregistration does not guarantee unbiased analysis
The registered plan may itself contain poor assumptions.
Data collection may fail.
measurement may be biased.
Preregistration protects one layer of the system.
31. Deviations are not automatically misconduct
Science encounters reality.
Equipment fails.
assumptions break.
unexpected distributions appear.
The proper response is transparent deviation and sensitivity analysis.
32. Deviation logs preserve provenance
Original plan.
change.
date.
reason.
effect on analysis.
This allows readers to reconstruct the research path.
33. Exploratory analyses should be welcomed and labelled
Unexpected patterns are often where new Science begins.
The distinction is simple:
discovery now, confirmation later.
34. Preregistration and replication are natural partners
An exploratory result generates a hypothesis.
A new study preregisters the hypothesis and tests it on fresh data.
The scientific loop becomes visible.
35. Preregistration and systematic review are connected
Review protocols can also be registered before screening begins.
This reduces flexibility in inclusion criteria and synthesis decisions after seeing study results.
36. Preregistration and meta-analysis are connected
Prospective meta-analysis can prespecify outcomes and pooling plans across studies before individual results are known.
This reduces selective synthesis.
37. Preregistration and statistical significance are connected
A p-value is easier to interpret when the hypothesis, outcome and stopping rule were chosen before the data were examined.
38. Preregistration and multiple comparisons are inseparable
Readers need to know which analyses were planned and how many opportunities existed for chance findings.
Registration exposes the denominator.
39. Preregistration and effect size are connected
The minimum meaningful effect can be declared before seeing results.
This prevents practical thresholds from drifting toward whatever the data happened to show.
40. Preregistration and statistical power are connected
Sample-size planning should follow the scientifically meaningful effect, expected variability and design.
The logic belongs before recruitment.
41. Registered Reports go one step further
Preregistration records the plan.
Registered Reports add formal peer review of that plan before results determine publication.
The next article follows this model.
42. Registered analysis plans can be detailed
Some projects preregister code or simulated-data workflows so the entire analysis can be tested before real outcomes are opened.
This strengthens reproducibility.
43. Blinded analysis can reinforce preregistration
Analysts may work on coded treatment groups or altered outcomes while checking model assumptions.
Final group labels are revealed only after the analysis pipeline is frozen.
44. Holdout data creates a practical preregistration boundary
Develop the model on training data.
freeze the procedure.
then open the untouched test set.
The holdout functions as fresh evidence.
45. Reusing the holdout weakens confirmation
Every time the result is inspected and the model is changed, the test set becomes part of development.
A new holdout may eventually be needed.
46. AI makes preregistration more important
AI can generate hundreds of hypotheses, transformations and models rapidly.
Without a frozen evaluation plan, search flexibility becomes enormous.
47. AI can help draft preregistrations
Useful prompts:
“Turn this research question into explicit hypotheses and outcomes.”
“List ambiguous decisions that should be fixed before data collection.”
“Generate a deviation log template.”
“Check whether the analysis plan leaves hidden flexibility.”
48. AI should not invent scientific justifications
A polished preregistration can still contain assumptions without evidence.
Every planned threshold, outcome and exclusion rule should have a scientific reason.
49. AI-generated protocols need version control
Which version was approved?
What changed?
When?
Automatic rewriting can erase the distinction between original and revised plans if version history is not preserved.
50. Parents can teach preregistration through study experiments
Before trying a new revision method, write:
which topic.
how long.
what outcome will be checked.
what counts as improvement.
Then test without changing the rules afterward.
51. Small-group tuition can run prediction-before-data exercises
Give students an experimental setup but hide the results.
Each writes:
hypothesis.
prediction.
variables.
analysis.
Then reveal the data.
Hindsight bias becomes visible.
52. Preregistration helps examinations too
Before calculating, state the relationship you expect.
Before reading the answer key, commit to the explanation.
Then compare.
The habit makes error diagnosis more honest.
53. A compact preregistration checklist
- What is the scientific question?
- What hypothesis is confirmatory?
- What is the primary outcome?
- What secondary outcomes are planned?
- What sample-size logic is used?
- What are inclusion and exclusion rules?
- How will outliers be handled?
- Which variables and transformations are planned?
- Which statistical model or test will be used?
- What stopping rule applies?
- How will multiplicity be controlled?
- When was the plan registered relative to data access?
- How will deviations be documented?
- Which analyses will be labelled exploratory?
54. Frequently asked questions
What is preregistration?
Preregistration is the timestamped recording of key research questions, hypotheses, outcomes, design and analysis decisions before the relevant results are known.
Does preregistration stop researchers changing the plan?
No. Changes can be scientifically necessary. The important requirement is to document deviations and distinguish them from the original plan.
Does preregistration ban exploratory analysis?
No. Exploration remains valuable. Preregistration helps readers distinguish planned confirmation from post-hoc discovery.
Does preregistration guarantee reliable Science?
No. It increases transparency but cannot repair poor measurement, bias, confounding or weak theory by itself.
How does preregistration help students?
It teaches prediction before observation, clear variable definitions and honest comparison between what was expected and what actually happened.
55. Continue the Science Education Systems series
- How Scientific Systematic Review Works
- How Scientific Meta-Analysis Works
- How Scientific Registered Reports Work
Conclusion: Preregistration protects the difference between prediction and hindsight
Maya writes what she expects.
Jia Jun fixes the analysis.
Hana records the deviation.
Ethan saves fresh data for the hard test.
Science needs all four.
Plan first.
timestamp the plan.
collect evidence.
report changes honestly.
keep exploration alive.
Then let a new experiment decide whether the discovered pattern survives outside the data that revealed it.
