Science Education Systems · Article 65. Maya, Jia Jun, Hana and Ethan remain fictional Punggol learners. This article follows the randomisation layer: how Science uses controlled chance to prevent researchers from quietly deciding who receives which condition.
The 50-second parent route
Randomisation is not disorder.
It is a rule for assigning or selecting without favouring a preferred outcome.
The route is:
question → eligible units → random mechanism → assignment → balance → intervention → measurement → comparison → uncertainty → inference → replication
The key question is:
Could the groups have differed systematically before the experiment even began?
This article extends How Scientific Controls Work, How Scientific Bias Works and How Scientific Probability Works.
1. Randomisation protects against allocation bias
If the researcher chooses who receives treatment, conscious or unconscious preferences can distort the groups.
Random assignment removes that discretionary pathway.
2. Random assignment is different from random sampling
Random sampling asks who enters the study.
Random assignment asks which condition each study unit receives.
Sampling supports representativeness.
Assignment supports causal comparability.
3. Random does not mean equal in every detail
Chance can still create imbalances.
One group may contain slightly older participants or slightly larger plants.
Randomisation makes systematic favouring less likely; it does not guarantee perfect matching.
4. Maya’s randomisation error is alternation
She assigns A, B, A, B, A, B.
This looks fair.
But if the investigator can predict the next assignment, selection can still be manipulated.
Her repair:
use a genuinely random or concealed assignment process.
5. Jia Jun’s randomisation error is coin-toss literalism
He thinks every randomised study must use a physical coin.
His repair:
understand that random number generators, shuffled sequences and validated allocation systems can all implement randomisation.
6. Hana’s randomisation error is imbalance panic
The groups are not identical after assignment.
She concludes randomisation failed.
Her repair:
evaluate whether differences are compatible with chance and whether the analysis accounts for important baseline variables.
7. Ethan’s randomisation error is randomising the wrong level
He assigns individual students randomly even though the intervention operates by classroom.
His repair:
randomise at the level where treatment is actually delivered.
8. Randomisation creates a probability model for group differences
Because assignment is governed by chance, statistical inference can describe how much variation might arise even if the treatment has no effect.
Randomisation and probability are tightly connected.
9. Randomisation strengthens the counterfactual comparison
We cannot observe the same person both treated and untreated at the same moment.
A randomised control group approximates what might have happened without treatment.
10. Large randomised groups tend to balance many variables on average
Known variables.
unknown variables.
measured variables.
unmeasured variables.
Randomisation helps distribute them without needing to identify every one in advance.
11. Small randomised studies can still be imbalanced
With few units, chance variation is larger.
Randomisation protects against systematic assignment bias, but small samples remain statistically fragile.
12. Blocking can improve balance on important variables
Group similar-sized plants first.
Then randomise within those blocks.
This preserves chance while reducing imbalance on a known important factor.
13. Stratified randomisation does something similar
Researchers may randomise separately within age groups, sites or other important categories.
The goal is controlled balance where one variable matters strongly.
14. Cluster randomisation assigns groups rather than individuals
Schools.
classrooms.
villages.
clinics.
This can prevent contamination when members of one cluster interact closely.
15. Cluster randomisation changes the statistics
People within one cluster tend to resemble one another.
Ten students from one classroom do not provide the same independent information as ten students from ten unrelated classrooms.
Dependence must be modelled.
16. Crossover randomisation can change treatment order
Some studies allow the same participant to receive multiple conditions in random order.
This controls some individual variation but introduces possible carryover effects.
17. Order can be randomised in laboratory tasks
If every sample is always tested in the same sequence, time drift may bias later measurements.
Randomising order can distribute this effect.
18. Primary Science can learn randomisation through fair selection
Do not choose the biggest seeds for one group and the smallest for another.
Mix them and assign without preference.
The child learns why selection fairness matters.
19. Primary 3 can use simple shuffled assignment
Label objects.
shuffle cards.
assign conditions.
The visible chance mechanism makes fairness concrete.
20. Primary 4 can compare chosen versus random groups
Let one student choose groups deliberately.
Let another randomise.
Discuss how preferences can enter the chosen grouping.
21. Primary 5 can connect randomisation to hidden variables
Plant health differs in ways we may not notice.
Random assignment helps distribute hidden differences rather than placing all strong plants in one condition.
22. Primary 6 can evaluate experimental fairness
Were the groups assigned fairly?
Could the investigator have influenced who received what?
Randomisation becomes part of method evaluation.
23. Secondary Science can formalise randomisation
random allocation.
blocked designs.
cluster designs.
randomised order.
random-number generation.
Students can see chance used as a design tool rather than a nuisance.
24. Allocation concealment is different from blinding
Allocation concealment protects the assignment process before treatment begins.
Blinding protects later behaviour or measurement from knowledge of assignment.
The next article, How Scientific Blinding Works, follows that layer.
25. Predictable assignment can undermine randomisation
If recruiters know the next participant will enter the treatment group, they may delay or accelerate enrolment consciously or unconsciously.
Concealment protects the random sequence.
26. Randomisation does not remove measurement bias
The groups can be assigned perfectly and still be measured differently.
Blinding and standardised measurement remain important.
27. Randomisation does not remove attrition bias
If dropout differs after assignment, the groups may become less comparable over time.
Follow-up quality still matters.
28. Randomisation does not guarantee external validity
A perfectly randomised experiment may involve a narrow population or artificial setting.
Strong internal validity and broad generalisation are different questions.
29. Randomisation supports intention-to-treat reasoning
Analyse participants according to the group to which they were originally assigned, regardless of perfect adherence.
This preserves important benefits of random allocation when estimating the effect of assignment.
30. Per-protocol analysis answers a different question
It focuses on participants who followed the intended treatment sufficiently.
This may estimate a different effect but can reintroduce selection bias because adherence is not random.
31. Both analyses can be informative when interpreted correctly
The important point is to match the analysis to the scientific question and preserve transparency about what population the estimate describes.
32. Randomisation and controls are inseparable
A control provides the comparison condition.
Randomisation strengthens how units enter those conditions.
Together they build a more credible causal contrast.
33. Randomisation and confounding are inseparable
Confounders threaten causal interpretation in observational studies.
Random assignment helps break systematic links between treatment and both known and unknown confounders.
See How Scientific Confounding Works.
34. Randomisation and blinding solve different problems
Randomisation protects assignment.
Blinding protects behaviour, treatment, assessment or analysis from expectation effects.
Strong experiments often need both.
35. Randomisation and probability support inference
Because assignment is chance-governed, we can reason about how likely the observed group difference would be under no treatment effect.
This is one foundation of randomisation-based statistical tests.
36. Randomisation tests can use the actual assignment mechanism
Imagine reshuffling treatment labels many times according to the original randomisation rule.
How often would a difference as large as the observed one appear by chance?
This directly links design and inference.
37. Randomisation can be audited
Was the sequence generated correctly?
Was allocation concealed?
Were there unexplained deviations?
Scientific provenance applies to randomisation too.
38. Random seeds can support computational reproducibility
If software generates the assignment, recording the seed can reproduce the exact sequence when appropriate.
The seed should not be exposed in ways that compromise concealment during active assignment.
39. Ethical randomisation requires genuine uncertainty and acceptable options
Researchers should not assign participants randomly to conditions known to be unacceptably harmful merely for methodological purity.
Ethics constrains experimental design.
40. Equipoise is one ethical idea behind some randomised studies
There should be meaningful uncertainty about which intervention is better for the population under study.
Specific ethical standards vary by domain, but uncertainty and participant welfare matter.
41. Randomisation is not always feasible
Climate exposure.
lifetime habits.
historical events.
rare disasters.
When assignment is impossible or unethical, observational and quasi-experimental methods become important.
42. Natural randomisation can sometimes be exploited
Lotteries, thresholds or institutional rules may create near-random variation.
Scientists can sometimes use these natural structures for stronger causal inference.
43. AI experiments need randomisation
If one group of users receives a new AI tool and another receives the old workflow, assignment should not depend on motivation or prior skill if the goal is a fair causal comparison.
44. Online A/B tests are randomised experiments
Users are assigned to variants.
Outcomes are compared.
But instrumentation, interference, repeated testing and subgroup effects can still complicate interpretation.
45. AI can help generate randomisation plans
Useful prompts:
“Propose a blocked randomisation scheme for this experiment.”
“Identify the correct level of randomisation.”
“Explain whether cluster randomisation is needed.”
“List ways allocation could accidentally become predictable.”
46. AI can also create fake randomness
A language model generating labels casually is not necessarily a validated random-number source.
Important experimental randomisation should use a method appropriate to the scientific stakes.
47. Parents can teach randomisation through simple home experiments
Testing two study methods?
Randomise which topic uses which method rather than assigning the easier topic to the preferred method.
Fair assignment makes the comparison more meaningful.
48. Small-group tuition can make allocation bias visible
Ask one student to create two “equal” groups by judgement.
Then randomise.
Compare the assumptions that entered each process.
49. A compact randomisation checklist
- What units are eligible for assignment?
- At what level should randomisation occur?
- What random mechanism will be used?
- Can anyone predict the next assignment?
- Should important variables be blocked or stratified?
- Could contamination occur across groups?
- Are baseline imbalances plausible by chance?
- Will measurement be blinded where appropriate?
- How will deviations from assignment be handled?
- Does the analysis preserve the randomised comparison?
- Does randomisation support the causal question but not overpromise generalisation?
50. Frequently asked questions
What is scientific randomisation?
It is the use of a chance-governed procedure to assign or select study units so systematic human preference is reduced.
Why does randomisation help causal inference?
It helps distribute known and unknown baseline differences across groups without relying entirely on researcher judgement.
Is random assignment the same as random sampling?
No. Random sampling concerns who enters a study; random assignment concerns which condition study units receive.
Does randomisation guarantee identical groups?
No. Chance can still create imbalance, especially in small studies.
How does randomisation help PSLE Science?
It teaches fair selection, unbiased grouping and why experimental comparison should not favour one condition before testing begins.
How does randomisation change in Secondary Science?
It expands into blocked, stratified, cluster and computer-generated allocation, with stronger links to probability and causal inference.
51. Continue the Science Education Systems series
- How Scientific Blinding Works
- How Scientific Confounding Works
- How Scientific External Validity Works
Conclusion: Randomisation uses chance to make the comparison less dependent on choice
Maya wants to choose the groups.
Jia Jun generates the sequence.
Hana checks concealment and imbalance.
Ethan asks whether the randomisation level matches the system.
Science needs all four.
Define the units.
randomise fairly.
protect the allocation.
measure consistently.
analyse according to the design.
Then let chance serve as a guardrail against hidden preference.
