Randomisation works by assigning eligible study units to comparison conditions through an unpredictable chance process, while concealing future assignments from the people enrolling those units, so that treatment allocation is not systematically determined by baseline prognosis, preference or recruiter judgement. This makes the groups comparable in expectation and gives later outcome differences a stronger causal interpretation.
Randomisation is often taught with a coin.
Heads: Treatment A.
Tails: Treatment B.
The coin is memorable.
The deeper mechanism is not.
Randomisation is not valuable because randomness is scientifically mystical. It is valuable because an unpredictable assignment rule closes a dangerous human pathway. If neither the participant nor the recruiter can determine which eligible person receives which intervention, baseline differences have less opportunity to be intentionally or unintentionally routed into one group.
That single design move can transform a comparison.
The governing question: before outcomes happen, can treatment assignment be made independent of the characteristics that would otherwise predict those outcomes?
Quick Read
ELIGIBILITY → CONSENT / ENROLMENT → RANDOM SEQUENCE → ALLOCATION CONCEALMENT → ASSIGNMENT → IMPLEMENTATION → FOLLOW-UP → OUTCOME MEASUREMENT → ANALYSIS BY ASSIGNED GROUP WHERE APPROPRIATE → EFFECT ESTIMATE → UNCERTAINTY → CAUSAL INTERPRETATION → GENERALISATION
CONSORT 2025 separates two mechanisms that are often blurred together:
- Random sequence generation: create an unpredictable sequence of treatment assignments.
- Allocation concealment: prevent the people enrolling participants from knowing the upcoming assignment before the participant is irreversibly entered.
Both matter. A mathematically random list can still be subverted if recruiters can see what comes next.
1. Randomisation Solves a Counterfactual Problem
Suppose one student receives a new tutoring method and later scores 80.
Did the tutoring cause the score?
We cannot observe the same student, in the same week, under the same world history, both with and without the tutoring.
The missing outcome under the unreceived condition is the counterfactual.
Randomised trials solve this at group level. They create comparison groups whose assignment mechanism is unrelated to baseline prognosis by design, so the outcomes of one group can serve as an estimate of what might have happened to the other group under the alternative intervention, subject to the design and analysis assumptions.
See How Counterfactuals Work for the larger causal architecture.
2. Randomisation Is Random Assignment, Not Random Sampling
This distinction is essential.
| Mechanism | Question | Main inferential job |
|---|---|---|
| Random sampling | Who enters the study? | Connect the observed sample to a target population under a probability sampling design |
| Random assignment | Which intervention does an enrolled unit receive? | Create baseline-comparable intervention groups for causal inference |
A clinical trial may recruit a non-random convenience sample of eligible volunteers and then randomise those volunteers to treatment groups.
The trial can have strong internal causal validity for enrolled participants while still leaving an external-validity question about the broader patient population.
Sampling and randomisation therefore own different gates.
3. “Random” Has a Technical Meaning
CONSORT’s current explanation is unusually clear: random assignment means each participant has a known probability of receiving each intervention before assignment, and the assignment is determined by a chance process that cannot be predicted.
Alternating A, B, A, B is not random.
Assigning by date of birth is not random.
Assigning by hospital record number is not random.
Those methods can look impartial because the rule is mechanical.
But future assignments can often be predicted.
Predictability reopens the human selection pathway.
4. Unpredictability Is More Important Than Visual Chaos
A randomisation sequence does not need to look messy to be valid.
It needs to be generated by an appropriate chance process and protected from prediction.
A computer-generated random sequence can produce runs such as A, A, A, B, A.
That may feel “not random” to a human observer because people expect alternation.
Randomness permits clusters.
Forcing every short block to look perfectly balanced can increase predictability if the block structure is exposed.
5. Sequence Generation Creates the Assignment Order
A random sequence can be generated using validated computer procedures or other suitable chance mechanisms.
The protocol should describe who generated the sequence and how.
This is not bureaucratic detail.
It tells readers whether the assignment process could have been influenced by people responsible for recruitment or treatment delivery.
CONSORT 2025 specifically asks authors to report who generated the sequence, the method, the randomisation type and any restrictions such as stratification or blocking.
6. Allocation Concealment Protects the Sequence From the Recruiter
Imagine a recruiter knows the next participant will be assigned to Treatment A.
A borderline-eligible patient arrives.
The recruiter strongly believes Treatment A is better.
Even with no intention to cheat, the recruiter may interpret eligibility more generously.
Now the treatment group contains a participant partly because the upcoming assignment was known.
Allocation concealment prevents that information from entering the enrolment decision.
Central web systems, telephone systems, pharmacies or carefully prepared sequentially numbered opaque sealed containers can be used to protect assignments, depending on the context.
Random sequence generation makes the future assignment unpredictable. Allocation concealment keeps the future assignment unknowable to the people who could act on it.
7. Allocation Concealment Is Not Blinding
Allocation concealment acts before and until assignment.
Blinding acts after assignment.
Concealment prevents selection bias at enrolment.
Blinding can reduce behaviour or measurement changes caused by knowing which intervention was received.
CONSORT 2025 explicitly warns against confusing them.
A trial can have perfect concealment and no blinding. This is common where interventions are obvious, such as surgery, education or lifestyle programmes.
8. Randomisation Balances Prognostic Factors in Expectation, Not by Guarantee
Random assignment does not promise that every baseline variable will be numerically identical between groups.
By chance, one group can contain more older participants, higher baseline scores or more severe cases.
The design guarantee is different.
The assignment mechanism did not systematically use those prognostic factors to decide treatment. Across repeated randomisations, the groups are balanced in expectation.
Chance imbalance is part of sampling uncertainty, not proof that the randomisation failed.
9. Baseline Significance Tests Are Usually the Wrong Test of Randomisation
Researchers sometimes test whether baseline characteristics differ “significantly” between randomised groups.
But if the sequence was genuinely random and concealed, any baseline differences are already known to be due to chance under the allocation mechanism.
The scientifically useful task is to describe important baseline characteristics, understand chance imbalance and follow the pre-specified analysis plan—not to use p-values as a forensic test of whether a correctly implemented random process was random.
10. Simple Randomisation Is the Purest Form
Simple randomisation assigns treatments according to a fixed allocation probability, such as 1:1, independently for each participant.
Conceptually, it resembles repeated fair coin tosses for two groups, though large trials normally use robust computer systems rather than literal coins.
Its advantage is unpredictability.
Its disadvantage in smaller trials is that group sizes or key baseline factors can become imbalanced by chance.
11. Block Randomisation Protects Group-Size Balance
Blocked randomisation restricts the sequence so that treatment numbers remain reasonably balanced through enrolment.
This is useful when recruitment occurs over time or a trial might stop early.
But fixed, known block sizes can make upcoming assignments predictable near the end of a block.
Randomly varying block sizes and maintaining concealment reduce that risk.
The design trades unrestricted randomness for controlled balance without giving up the central requirement of unpredictability.
12. Stratified Randomisation Protects Balance on Important Baseline Factors
Suppose disease severity strongly predicts outcome.
Researchers may randomise separately within severity strata so treatment groups remain balanced on that factor.
Educational trials might stratify by school, grade or baseline attainment.
Stratification can improve balance on a small number of important variables.
It cannot feasibly stratify on every variable, and unnecessary complexity can create implementation problems.
13. Minimisation Is Not Pure Simple Randomisation but Can Include Randomness
Minimisation assigns new participants in ways that reduce imbalance across several prognostic factors, usually retaining a random component.
It is useful when studies are small relative to the number of important baseline factors.
The analysis and reporting must reflect the actual allocation method.
The governing principle remains the same: assignment should not become a predictable discretionary choice by investigators.
14. Unequal Allocation Can Still Be Random
A trial does not need a 1:1 ratio.
Participants might be randomised 2:1 or 3:1 for ethical, recruitment, safety, learning or logistical reasons.
Each participant still has a known probability of assignment under an unpredictable process.
Unequal allocation can reduce statistical efficiency for a fixed total sample in simple two-group comparisons, so the reason and sample-size consequences should be planned rather than improvised.
15. Cluster Randomisation Assigns Groups, Not Individuals
Sometimes an intervention naturally operates at group level.
A school adopts a curriculum.
A clinic changes a protocol.
A village receives a public-health intervention.
Randomising individual people inside the same cluster may create contamination because they share teachers, clinicians or environment.
Cluster randomisation assigns whole groups.
The price is correlation: participants within one cluster are more similar than independent participants, reducing effective information and requiring cluster-aware analysis and usually a larger sample.
16. The Unit of Randomisation Must Match the Unit of Analysis
If ten schools are randomised, the trial does not suddenly contain 1,000 independent randomised units because each school has 100 students.
Students provide outcome information, but assignment happened at school level.
Ignoring clustering can severely underestimate uncertainty.
This is the same hierarchical principle described in How Research Variables Work and How Sampling Works.
17. Pair-Matched and Matched Randomisation Use Baseline Similarity Before Chance Assignment
Clusters or participants can be matched into similar pairs or sets using pre-treatment characteristics, then randomised within those sets.
This can improve balance and efficiency when strong prognostic factors are known.
But matching variables and pair structure should be incorporated into the analysis.
Design information should not disappear after data collection.
18. Factorial Randomisation Tests More Than One Intervention Efficiently
In a 2×2 factorial design, participants can be randomised to receive Intervention A or not and Intervention B or not, producing four combinations.
This can efficiently estimate main effects when assumptions about interactions are appropriate and can also investigate interaction between interventions.
Factorial designs show that randomisation is not tied to one binary treatment decision. It is a general assignment architecture for creating controlled contrasts.
19. Crossover Randomisation Lets Participants Receive More Than One Condition
In crossover trials, participants receive interventions in different randomised sequences.
Each person can serve as their own control, increasing efficiency for stable chronic conditions and reversible interventions.
But carryover effects, period effects and irreversible treatment consequences can make crossover inappropriate.
The design must match the biology or behaviour.
20. Randomisation Does Not Make Treatment Delivery Identical
Assignment is only the beginning.
Participants may not adhere.
Teachers may implement a curriculum differently.
Clinicians may add co-interventions.
Students in different groups may share materials.
Randomisation protects assignment.
Implementation needs its own fidelity and contamination controls.
21. Blinding Can Protect What Happens After Randomisation
If participants know their treatment, expectations may alter behaviour.
If clinicians know treatment, co-interventions can differ.
If assessors know treatment, subjective outcome judgement can shift.
Blinding reduces some of these post-assignment information pathways.
It cannot always be implemented. A student knows whether they attend tutoring. A surgeon knows which operation they perform.
When blinding is impossible, objective outcomes, independent assessors and standardised care can reduce remaining bias.
22. Placebos Help Separate Treatment Content From Expectation and Context
In some drug trials, a placebo resembles the intervention without containing the active component.
This can help blind participants and distinguish pharmacological effects from expectation, attention and trial participation effects.
But a placebo comparison answers a particular question.
Sometimes decision-makers need comparison with existing standard treatment rather than placebo.
Randomisation does not choose the scientifically relevant comparator automatically.
23. The Comparator Defines the Effect Being Estimated
“Does Treatment A work?” is incomplete.
Compared with no treatment?
Placebo?
Usual care?
An alternative treatment?
A lower dose?
Different randomised comparisons estimate different causal contrasts.
Excellent randomisation cannot repair an irrelevant comparator.
24. Intention-to-Treat Preserves the Meaning of Random Assignment
In many superiority trials, the primary analysis follows the intention-to-treat principle: participants are analysed according to the group to which they were randomised, regardless of whether they fully adhered to the assigned intervention.
Why?
Because switching people analytically according to what they actually received can reintroduce baseline differences. Adherence is not random; people who comply may differ from those who do not.
Intention-to-treat preserves the original randomized comparison and often estimates the effect of assignment to an intervention strategy under trial conditions.
25. Per-Protocol Analyses Answer a Different Question
A per-protocol analysis focuses on participants who adhered sufficiently to the protocol.
This may seem like a purer measure of treatment efficacy.
But adherence can be related to prognosis, side effects, motivation or response, so simply comparing compliers can sacrifice the protection of randomisation.
Modern causal methods can estimate more specialised adherence effects under additional assumptions.
The key is to recognise that the estimand changed.
26. Missing Outcomes Can Break the Randomised Comparison
Randomisation occurs at baseline.
If outcome data later go missing differently by group and prognosis, the analysed groups may no longer preserve the original comparability.
Complete-case analysis can therefore introduce bias.
Prevention of loss to follow-up, transparent participant-flow reporting and appropriate missing-data methods are part of maintaining randomisation’s value after assignment.
27. Randomisation Does Not Remove Measurement Bias
A perfectly randomised trial can still use a poor outcome measure.
If both groups are measured with an invalid proxy, the treatment contrast may be precisely estimated for the wrong construct.
If assessors measure groups differently, detection bias can distort the difference.
Randomisation protects one causal path.
Measurement still needs its own quality system.
28. Randomisation Does Not Remove Reporting Bias
A trial can randomise perfectly and report only favourable outcomes.
It can switch a secondary outcome to primary after seeing the results.
It can emphasise a significant subgroup while downplaying the primary null result.
Trial registration, protocols, statistical analysis plans and CONSORT reporting standards protect the timeline and visibility of those decisions.
29. Randomisation Does Not Guarantee Generalisability
A trial may enrol healthy volunteers, exclude older adults, use expert centres or deliver an intervention with unusually high fidelity.
The causal effect can be valid for the enrolled study while differing in routine practice.
External validity remains a separate question.
Randomisation tells us why groups differ inside the trial more credibly. It does not automatically tell us where the result travels.
30. Randomisation Does Not Guarantee a Large Enough Study
A tiny trial can be beautifully randomised and too imprecise to answer the question.
Randomisation creates an unbiased assignment mechanism.
Sample size determines how precisely the resulting contrast can usually be estimated under the design.
Statistical power, measurement reliability, outcome variability and clustering all matter.
31. Randomisation Does Not Guarantee the Groups Look Balanced in a Small Trial
Chance can create visibly different groups.
This can feel alarming.
But trying to manually “repair” a randomised sequence after seeing assignments can destroy the randomisation.
Better design choices are made prospectively: stratification, blocking, minimisation or pre-specified covariate adjustment where appropriate.
Do not edit chance after it speaks because the output does not look aesthetically balanced.
32. Covariate Adjustment Can Improve Precision After Randomisation
Adjusting for strongly prognostic baseline variables can improve precision and account for stratification factors when pre-specified appropriately.
This does not mean the trial ceased to be randomised.
Nor does it mean every variable should be added.
Post-randomisation variables require particular care because adjustment can alter the estimand and introduce bias.
33. Randomisation Gives Probability Theory a Design-Based Role
Because assignment was generated by chance, statistical inference can use the known assignment mechanism.
Randomisation inference asks how unusual the observed treatment contrast would be under alternative assignments consistent with the trial design if there were no treatment effect of the specified kind.
Model-based analyses are often used in practice, but the random assignment remains the foundation of the causal contrast.
34. Randomisation Turns Hidden Baseline Causes From Confounders Into Chance Imbalance
In observational treatment data, an unknown prognostic factor may influence who receives treatment and also influence outcome.
That factor confounds the treatment effect.
Under successful randomisation, treatment assignment does not systematically depend on that unknown baseline factor.
The factor can still differ by chance in the realised sample, but it is not structurally tied to treatment assignment.
This is why randomisation can protect against both measured and unmeasured baseline confounding in a way ordinary adjustment cannot guarantee.
35. Natural Experiments Are Not Randomised Trials
Sometimes external events assign exposure in ways that approximate random or quasi-random variation.
Researchers call these natural experiments.
They can produce powerful causal evidence when the assignment mechanism and assumptions are credible.
But they should not be called randomised simply because the researchers did not choose the exposure.
The causal identification strategy needs to be demonstrated, not implied by the word natural.
36. Mendelian Randomisation Is Not Experimental Random Assignment by Researchers
Mendelian randomisation uses genetic variants as instrumental variables to study causal effects of modifiable exposures under strong assumptions.
The name refers to the random allocation of alleles at conception, not to a trial investigator assigning people to treatments.
The method can be useful, but its validity depends on instrumental-variable assumptions such as relevance and absence of prohibited pathways from the genetic instrument to outcome.
Names that contain “randomisation” do not erase the need to inspect the actual assignment architecture.
37. Randomised Encouragement Designs Separate Assignment From Uptake
Sometimes researchers cannot randomise the treatment itself but can randomise an encouragement to receive it.
The encouragement changes treatment uptake probabilistically.
Instrumental-variable methods can then estimate particular causal effects under assumptions.
This makes explicit a larger lesson:
Randomise the lever you can control, then be precise about which causal quantity the lever identifies.
38. Adaptive Randomisation Changes Probabilities During the Trial
Some trial designs alter allocation probabilities as information accumulates.
Response-adaptive randomisation may assign future participants with greater probability to treatments performing better so far.
Such designs can have ethical or efficiency motivations but require careful statistical planning because outcomes influence future allocation probabilities.
Adaptive does not mean improvised.
The adaptation rule must be prospectively specified and analysed accordingly.
39. Platform Trials Use Randomisation Inside a Living Research System
Platform trials can evaluate multiple interventions under a shared master protocol, allowing arms to enter or leave over time.
Randomisation remains central, but time trends, shared controls, adaptation and changing eligibility can create sophisticated analysis requirements.
The design demonstrates how far randomisation has evolved beyond the simple two-arm experiment while preserving its core assignment logic.
40. Randomisation in Education Needs Ethical and Practical Restraint
Not every classroom decision should become an experiment.
Randomisation may be appropriate when genuine uncertainty exists between acceptable alternatives and no learner is denied essential support.
But educational interventions involve children, curriculum obligations, parent consent, teacher workload and unequal learner needs.
Ethics and feasibility define what can legitimately be randomised.
A method is not scientifically superior if using it would be unethical.
41. Cluster Randomisation Often Fits Schools Better Than Individual Randomisation
If a teacher is trained in a new method, it may be impossible for that teacher to teach half the class with the old method without contamination.
Randomising teachers, classes or schools can align assignment with the intervention’s natural unit.
But a study with twenty schools has twenty randomised clusters even if thousands of students are observed.
Sample-size and analysis plans need to respect that fact.
42. Randomisation in Small Tuition Groups Has a Different Ethical Geometry
In a three-student tuition group, formal randomised experiments are often neither necessary nor appropriate for ordinary teaching decisions.
The teacher’s first duty is to the learners in front of them, not to producing publishable causal estimates.
Structured observation, repeated measures, alternating acceptable practice formats, delayed checks and carefully bounded single-learner evidence can often support instructional improvement without pretending to be a randomised trial.
Research design should fit both the question and the receiver.
43. Randomisation in Medicine Requires Clinical Equipoise and Ethical Oversight
Random assignment is ethically appropriate only when genuine uncertainty and the risk-benefit situation justify exposing participants to the compared conditions.
Researchers cannot randomise people to known harmful treatment merely because the causal estimate would be clean.
Informed consent, independent ethics review, safety monitoring and stopping rules are part of the trial architecture.
Causal identification does not outrank human welfare.
44. Randomisation in Public Policy Can Be Powerful and Politically Difficult
Governments and organisations sometimes randomise rollout timing, information, incentives or programme access where demand exceeds capacity or implementation already requires phased delivery.
This can produce strong causal evidence.
But public legitimacy matters. Citizens may object to important services being assigned by lottery even when the alternative allocation mechanism is no fairer.
Research design operates inside institutions and ethics, not outside them.
45. A/B Testing Is Randomisation in a Product Environment
Digital platforms routinely randomise users to interface variants.
The mechanics look simple.
But validity still depends on:
- stable assignment;
- appropriate randomisation unit;
- interference among users;
- pre-specified metrics;
- multiple testing;
- novelty effects;
- long-term outcomes;
- and whether the metric represents user value.
A perfectly random A/B test can optimise clicks while harming the receiver if clicks are the wrong outcome.
Randomisation protects causality, not objective selection.
46. Interference Breaks the Simplest “One Person, One Outcome” Assumption
Many causal analyses assume one participant’s treatment does not affect another participant’s outcome.
In infectious disease, education, social networks and marketplaces, that assumption can fail.
Vaccinating one person can protect others.
Teaching one student a strategy can spread it to classmates.
A marketplace experiment can change prices experienced by users in the control group.
Cluster or network randomisation and specialised estimands may be needed when treatment spills over.
47. Contamination Shrinks the Contrast Between Groups
If control participants gain access to the intervention, the assigned groups become more similar.
This often reduces the observed effect of assignment and can make an effective intervention look weaker.
Cluster randomisation, geographic separation or clear implementation protocols can reduce contamination where feasible.
Again, the assignment mechanism alone cannot control what happens after people enter the world.
48. Randomisation and Preregistration Protect Different Timelines
Randomisation protects how treatment is assigned.
Preregistration protects when hypotheses, outcomes and analytic decisions were specified.
A randomised trial can still search dozens of outcomes after seeing the data.
A preregistered observational study can still suffer unmeasured confounding.
These are complementary controls over different failure pathways.
49. Registered Reports Move Review Before Results
A randomised design can still be weak if the question is trivial, outcome invalid or sample inadequate.
Registered Reports allow editors and reviewers to evaluate the importance of the question and quality of the design before the results are known.
This reduces the chance that publication depends primarily on whether randomisation produced an exciting outcome.
The design principle is elegant:
Judge the experiment before knowing which side wins.
50. The Hostile Test: A Random List Left Open on the Desk
A statistician generates a perfect random sequence.
The recruiter keeps the list beside the consent forms.
Every upcoming assignment is visible.
The trial is randomised on paper.
The enrolment process can still be subverted.
This is why sequence generation without allocation concealment is an incomplete randomisation system.
51. The Second Hostile Test: Alternation Called Random
A clinic assigns the first patient to A, second to B, third to A and so on.
The rule feels fair.
The next assignment is always predictable.
If staff can delay or accelerate enrolment, assignment can become associated with patient characteristics.
Mechanical is not the same as random.
52. The Third Hostile Test: Randomisation Followed by Excluding the Inconvenient Participants
A trial randomises 500 participants.
After outcomes are known, investigators exclude many non-adherent participants from one arm and compare only those who followed treatment perfectly.
The original randomised groups have been replaced by selected subgroups.
The causal protection of the initial assignment no longer transfers automatically to the new comparison.
53. The Fourth Hostile Test: Randomised but Measured Differently
Participants are randomised perfectly.
The treatment group receives weekly assessment.
The control group is assessed only at the end.
Repeated assessment itself changes behaviour or makes more events detectable.
The outcome pathway is no longer equivalent.
Random assignment cannot rescue differential measurement after assignment.
54. The Fifth Hostile Test: Randomised to the Wrong Question
A trial flawlessly randomises two teaching apps.
The primary outcome is number of app logins.
App A wins decisively.
But the educational question was which app improves durable learning.
Randomisation established the causal effect on logins.
It did not validate logins as the educational outcome.
Design quality cannot substitute for construct validity.
55. Primary School: Randomisation Begins as “Do Not Choose the Groups Yourself”
A child wants to compare two fertilisers.
If they deliberately give the new fertiliser to the healthiest plants, the result is unfair before treatment begins.
Drawing plant numbers from a hat can create a simple chance assignment.
The child learns the foundation:
Do not let what I already know about the units decide who gets which condition.
56. Secondary School: Separate Selection, Assignment and Blinding
- How were participants selected?
- How were selected participants assigned to groups?
- Could recruiters predict the next assignment?
- Who knew the assigned condition afterward?
- Could that knowledge change treatment or measurement?
These questions teach more research literacy than memorising “randomised controlled trial” as the top of a pyramid.
57. JC and University: Randomisation Becomes a Full Allocation System
At higher levels, learners should reconstruct:
- unit of randomisation;
- allocation ratio;
- sequence-generation method;
- blocking or stratification;
- allocation concealment;
- who implemented assignment;
- blinding after assignment;
- analysis population;
- missing-data handling;
- and the exact treatment effect being estimated.
The word “randomised” becomes the beginning of the audit, not the end.
58. Where Randomisation Fits in the eduKateSG “How Works” Landscape
- How Scientific Research Works — the whole investigation route.
- How Hypotheses Work — the causal claims randomisation can test.
- How Research Variables Work — exposures, outcomes and confounders.
- How Research Bias Works — selection and confounding pathways randomisation helps block.
- How Research Validity Works — the internal-validity gain and external-validity boundary.
- How Sampling Works — who enters before assignment.
- How Counterfactuals Work — the missing alternative randomised groups approximate.
- How Probability Works — the mathematical language enabled by chance assignment.
Randomisation owns one highly specific scientific job: separating treatment assignment from baseline prognosis strongly enough that the treatment contrast can carry more causal weight.
59. What This Article Does Not Claim
- Randomisation is not the same as random sampling.
- Alternation or assignment by date is not genuine randomisation merely because it looks impartial.
- A random sequence is insufficient if allocation is not concealed from recruiters.
- Allocation concealment is not the same as blinding.
- Randomisation balances baseline prognostic factors in expectation, not perfectly in every realised trial.
- Randomisation does not prevent attrition, contamination, non-adherence, measurement bias or selective reporting.
- Randomisation does not guarantee adequate power.
- Randomisation does not guarantee external validity.
- Per-protocol comparison does not automatically preserve randomisation’s causal protection.
- Randomised evidence should still be interpreted relative to the actual comparator, population, implementation, outcome and estimand.
60. A Compact Randomisation Audit
- What is the unit being randomised?
- When does eligibility become final?
- Who generates the allocation sequence?
- What chance mechanism is used?
- What allocation ratio is planned?
- Is randomisation simple, blocked, stratified, clustered or otherwise restricted?
- Can anyone enrolling participants predict the next assignment?
- How is the sequence concealed?
- Who implements allocation?
- Who is blinded after assignment?
- Can groups contaminate one another?
- How well is the intervention implemented?
- Who is lost to follow-up?
- Are outcomes measured equivalently?
- Are participants analysed according to assigned groups for the primary estimand?
- Does the analysis respect clustering or stratification?
- Were outcomes and analyses pre-specified?
- How precise is the treatment effect?
- What causal question does the comparator actually answer?
- Where can the result generalise beyond the enrolled trial?
61. Frequently Asked Questions
What is randomisation in research?
Randomisation is the assignment of enrolled study units to intervention conditions through an unpredictable chance process with known assignment probabilities, ideally protected by allocation concealment so upcoming assignments cannot influence enrolment.
Why is randomisation important?
It prevents treatment assignment from being systematically determined by baseline prognostic factors and, when properly implemented, strengthens causal comparison by balancing known and unknown baseline factors in expectation.
What is allocation concealment?
Allocation concealment prevents the people enrolling participants from knowing the upcoming treatment assignment before enrolment is complete, reducing selection bias at trial entry.
What is the difference between allocation concealment and blinding?
Allocation concealment acts before and until assignment to protect who enters each group. Blinding acts after assignment to reduce behaviour or outcome measurement being influenced by knowledge of treatment.
Does randomisation guarantee equal groups?
No. Chance imbalance can occur in any realised sample, especially small trials. Randomisation guarantees the assignment mechanism, not identical baseline values.
Does a randomised trial automatically prove causation?
Randomisation strongly improves causal identification for the assigned contrast, but interpretation still depends on implementation, follow-up, measurement, analysis, reporting and the exact estimand. No single design feature makes every conclusion valid.
62. Authoritative Research Corridor
- CONSORT 2025 Explanation and Elaboration — random sequence generation, restricted randomisation, allocation concealment and blinding
- CONSORT 2025 Statement — updated reporting requirements for randomised trials
- Cochrane Handbook Chapter 8 — bias arising from the randomisation process
- National Academies — Scientific Methods and Knowledge
- Comparative Effectiveness Research Using Randomised Trials and Observational Studies
Final Thought: Randomisation Is a Door That Humans Agree Not to Hold Open
Without randomisation, a researcher choosing who receives which treatment carries enormous responsibility.
Perhaps they choose fairly.
Perhaps they do not notice that the sickest patients receive one treatment.
Perhaps the most motivated students enter one programme.
Perhaps the most promising cases are routed toward the new method.
Randomisation removes that choice from the person.
Then allocation concealment removes knowledge of the next assignment from the recruiter.
The door closes.
Not every bias disappears.
Not every participant behaves as assigned.
Not every measurement is valid.
Not every result generalises.
But one of the most dangerous causal pathways—baseline prognosis influencing treatment assignment—has been deliberately severed.
Randomisation is powerful because it does not ask investigators to be unbiased. It designs one crucial decision so their preferences cannot decide the comparison.