Experimental Research Design: Overview, Types, and Examples
An experimental study is one where the investigator allocates the intervention. Whether that allocation is random, whether the upcoming allocation was concealed, and whether anyone was blinded afterwards are three separate questions — and conflating them is the most common way a sound trial is reported as a weak one.
- What separates experimental from non-experimental research, and the three families inside it
- Why randomisation, allocation concealment and blinding are three different safeguards
- The current reporting guidelines — and which ones have been retired
- “Allocation concealment and blinding are the same thing.” They happen at different moments, protect against different biases, and a trial can have one without the other. Concealment operates before allocation and prevents foreknowledge of the next assignment, guarding against selection bias. Blinding operates after allocation and hides who got what, guarding mainly against ascertainment bias. A trial can be impossible to blind — surgery against physiotherapy — and still have impeccable concealment.
What makes a study experimental
The dividing line is allocation. Cochrane’s EPOC group defines a randomised trial as “an experimental study in which people are allocated to different interventions using methods that are random”, and a non-randomised trial as “an experimental study in which people are allocated to different interventions using methods that are not random”.1
Note what both definitions share: someone allocates. That is what makes a study experimental. If nobody assigns the intervention and exposure is simply recorded as it occurs, the study is observational, whatever else it does well.
Random sampling concerns who enters the study, and it governs how far results generalise. Random allocation concerns what happens to them once they are in, and it governs whether the comparison is fair. A trial can randomly allocate a thoroughly unrepresentative sample — most do — and still produce an internally valid estimate. The two are independent properties, and only the second makes a study a randomised trial.
What randomisation buys you
Randomisation is not a formality, and it is not simply a fairer way of dividing people up. It does specific work that no other allocation method does, and the CONSORT explanation and elaboration paper states that work plainly.
The first benefit is the one that matters most. Random assignment, “when properly implemented, eliminates selection bias, balancing both known and unknown prognostic factors”.10 Matching, adjustment and statistical control can only handle factors someone thought to measure. Randomisation is the only mechanism that also balances the factors nobody knew to look for, which is why its loss cannot be repaired in the analysis.
The second is statistical. “Random assignment permits the use of probability theory to express the likelihood that any difference in outcome between intervention groups merely reflects chance.”10 Without it, a p value is answering a question the design did not set up.
The cost of going without is not neutral: “without randomisation, treatment comparisons may be prejudiced, whether consciously or not”.10 Allocation concealment and blinding, covered further down this page, protect that mechanism once it is in place.
Two limits should be stated alongside those benefits, because overclaiming on this point is common. First, balance is not guaranteed: “although proper random assignment prevents selection bias, it does not guarantee that the groups are equivalent at baseline”.10 What it guarantees is the source of any imbalance — “any differences in baseline characteristics are, however, the result of chance rather than bias”.10 Second, internal validity is bought at a price elsewhere. Tight eligibility criteria “do not affect the internal validity of a trial, but they are central to its external validity”.10 A trial can be sound and still not describe your patients.
The three families of design
Experimental research is usually sorted into three families, distinguished by what controls the comparison rather than by subject matter.
| Type | Description | Example in research | Advantages | Limitations |
|---|---|---|---|---|
| Pre-experimental | A single group, or groups formed with neither randomisation nor a concurrent control — typically one group measured before and after | Service evaluations and audits where no comparator exists | Feasible when nothing else is; useful for generating hypotheses and establishing feasibility | Low internal validity. Nothing separates the intervention from secular trend, maturation or regression to the mean. Cochrane EPOC lists uncontrolled before-after studies among designs that are “strongly discouraged”1 |
| True experimental | Participants or clusters allocated to interventions “using methods that are random”1 | Parallel-group, crossover, factorial and cluster-randomised trials | Random allocation makes the groups comparable on everything, measured and unmeasured — the only design that does | Often impossible for ethical, political or logistical reasons;2 expensive; samples are rarely representative |
| Quasi-experimental | Someone allocates the intervention, but not at random — a design that “by definition lacks random assignment”2 | Interrupted time series, controlled before-after, regression discontinuity, difference-in-differences | Works where randomisation is impossible; can evaluate programmes that already happened; avoids withholding a promising treatment2 | Selection bias is the standing threat, and causal conclusions are “less definitive” than those from a well-conducted randomised trial2 |
A note on counting: sources differ on how many “types” there are, because some count the three families above while others count the geometries inside the true experimental family — parallel, crossover, factorial, cluster — as types in their own right. Both are defensible; mixing the two systems is what produces lists that promise four types and describe three. This page keeps them separate: three families here, four geometries below.
The one-shot case study
The pre-experimental family has a named worst case, and it is worth knowing by name because it is still common. Campbell and Stanley group three designs under the heading “THREE PRE-EXPERIMENTAL DESIGNS”, of which the first is “THE ONE-SHOT CASE STUDY”: a single group receives the intervention, and the outcome is measured once afterwards.11 They observe that “much research in education today conforms to a design in which a single group is studied only once”.11
The objection to it is not that the sample is small or the measurement crude. It is structural, and Campbell and Stanley put it in one sentence: “securing scientific evidence involves making at least one comparison”.11 A single post-intervention measurement supplies no comparison at all — not against the same group beforehand, not against anyone who went without. Whatever the number is, there is nothing it can be contrasted with, so nothing can be attributed to the intervention.
This is why such designs sit outside the experimental family rather than at the bottom of it, and why Cochrane EPOC places uncontrolled before-after studies among the designs that are “strongly discouraged”.1 A one-shot case study can describe what happened to a group. It cannot support a claim about what caused it.
When you are allowed to randomise
Randomisation is a methodological ideal constrained by an ethical one. The Declaration of Helsinki requires that research “must have a scientifically sound and rigorous design and execution”, be “likely to produce reliable, valid, and valuable knowledge and avoid research waste”, and “be based on a thorough knowledge of the scientific literature” (paragraph 21).3
That last clause has a practical consequence often missed: if the existing literature already answers your question, randomising people to find out again is not rigour, it is waste — and the Declaration treats avoiding it as an ethical requirement rather than a courtesy.
The choice of comparator is governed separately. Paragraph 33 requires that a new intervention “must be tested against those of the best proven intervention(s)”, with defined exceptions: “if no proven intervention exists, the use of placebo, or no intervention, is acceptable”, or where placebo is “necessary to determine the efficacy or safety of an intervention” and participants face no additional risk of serious or irreversible harm. The Declaration adds: “Extreme care must be taken to avoid abuse of this option.”3
Randomisation, concealment and blinding
These three are routinely treated as one safeguard. They are not, and a reviewer who knows the difference will notice immediately.
| Safeguard | When it operates | What it does | What it prevents | If it fails |
|---|---|---|---|---|
| Random allocation | At the moment of assignment | Assigns participants to groups by a chance process1 | Systematic differences between groups, including on characteristics nobody measured | The groups may differ at baseline in ways no adjustment can recover |
| Allocation concealment | Before allocation | Hides the upcoming assignment — “concealing assignments until the point of allocation prevents foreknowledge”4 | Selection bias: the rationale for random comparison groups is “to eliminate selection bias”4 | Recruiters can steer favourable patients into the preferred arm, undoing the randomisation |
| Blinding (masking) | After allocation | “Measures taken to conceal group identity after allocation”4 | Primarily ascertainment bias — bias in how outcomes are measured or judged4 | Outcome assessment and participant behaviour can drift with expectation |
The people who must not have foreknowledge before allocation are “investigators, health care providers, and participants”.4 After allocation, some or all of them may be unblinded without the randomisation itself being compromised.
“Double-blind” tells a reader nothing about concealment, and an unblindable trial is not thereby a weak one. Report the allocation sequence, the concealment mechanism and who was blinded as three separate statements. Reporting them as one is how a well-conducted trial acquires an avoidable risk-of-bias rating.
True experimental designs in practice
Within the randomised family, four geometries cover most trials. CONSORT 2025 notes that while it “provides guidance for reporting all randomised trials”, it “focuses on the most common type, the two-group parallel design”.5
| Design | Structure | Best suited to | Advantages | Limitations |
|---|---|---|---|---|
| Parallel group | Each participant is allocated to one arm and stays there | Most questions; the default | Simple to analyse and report; no carry-over to worry about | Needs the largest sample, since every comparison is between people |
| Crossover | Each participant receives both interventions in a randomised order, separated by a washout period | Stable, chronic conditions with short-acting treatments | Each participant is their own control, so far fewer are needed | Useless where the condition changes or the effect persists; carry-over and period effects must be addressed |
| Factorial | Two or more interventions randomised simultaneously, commonly 2×2 | Answering two questions in one trial, or testing an interaction | Efficient — two trials for close to the price of one | Assumes the interventions do not interact unless the trial is powered to detect it |
| Cluster randomised | Whole groups — clinics, schools, wards — are randomised rather than individuals | Interventions delivered to providers or settings, where contamination is otherwise unavoidable | Prevents spillover between arms; EPOC recommends randomising providers or clusters where the intervention targets them1 | Needs more participants for the same power; EPOC recommends at least two intervention and two control sites1 |
What the results look like
Three figures carry most of the reporting burden in a trial report. The schematics below show what each is for and what a reader checks first.
The flow diagram, from eligibility through to analysis — the first thing most readers and reviewers look at.
What to look for. Whether the numbers reconcile down the page, and whether the number analysed matches the number randomised. A gap between them is where intention-to-treat quietly becomes something else, and it is why reasons for exclusion belong in the boxes rather than the discussion.5
Schematic. Illustrative structure, not data from any study.
Parallel, crossover and factorial, as they would be drawn in a protocol.
What to look for. In a crossover, whether a washout long enough to clear the first treatment was built in and whether period effects were tested. In a factorial, whether the trial was powered for the interaction or merely assumes that none exists.
Schematic. Illustrative structure, not data from any study.
The primary outcome reported as an estimate and an interval, against the line of no effect.
What to look for. The width of the interval, not only which side of the line it falls. A narrow interval clear of the line is a precise finding; a wide one crossing the line is an inconclusive result, which is not the same as evidence of no effect. Report the estimate and interval for every prespecified outcome, including the ones that disappointed.5
Schematic. Illustrative shapes, not data from any study.
Where the outcome is how long until something happens, with numbers at risk beneath.
What to look for. The numbers at risk. Curves separate most dramatically at the right-hand end, which is exactly where fewest participants remain and the estimate is least stable. A curve drawn without its at-risk row invites a reader to over-read its tail.
Schematic. Illustrative shapes, not data from any study.
How trials are appraised
Randomised trials are assessed with RoB 2, the revised Cochrane risk-of-bias tool. It works through five domains:6
- “Bias arising from the randomisation process”
- “Bias due to deviations from intended interventions”
- “Bias due to missing outcome data”
- “Bias in measurement of the outcome”
- “Bias in selection of the reported result”
Each domain is rated low risk, some concerns or high risk, based on answers to signalling questions; an algorithm proposes a rating that reviewers may override with reasons. The overall judgement generally follows the worst domain, and a study with “some concerns” across several domains may be rated high risk overall.6
The first domain is where allocation concealment is judged, which is the practical reason the distinction in the previous section matters. A non-randomised study is appraised with a different instrument entirely — ROBINS-I, which covers seven domains including confounding and selection of participants.7 Using the wrong tool for the design is a common and visible error.
Protocol and reporting
Two guidelines bracket a trial, and both were replaced in 2025. Anything citing the previous versions is out of date.
| Stage | Guideline | Items | Status |
|---|---|---|---|
| Protocol, before the trial | SPIRIT 20258 | “34 minimum items to address in a trial protocol”8 | “Supersedes the SPIRIT 2013 statement, which should no longer be used or cited”8 |
| Report, after the trial | CONSORT 20255 | “A 30-item checklist of essential items”5 | “Supersedes the CONSORT 2010 statement, which should no longer be used”5 |
CONSORT 2025 “added seven new checklist items, revised three items, deleted one item”, and restructured the checklist “with a new section on open science”. The deleted item on generalisability is now folded into trial limitations at item 30.5 SPIRIT 2025 likewise adds an open science section, more emphasis on the assessment of harms, and a new item on how patients and the public will be involved in trial design, conduct and reporting.8
SPIRIT exists because protocol quality varies: “despite the central role of protocols, there is substantial variation in the completeness of protocol content”, and gaps “can lead to avoidable protocol amendments, inconsistent or poor trial conduct, and lack of transparency”.8
Extensions exist for cluster, crossover, pilot and many other designs. The CONSORT 2025 authors say they will work with extension leaders to align them, and that until then readers should “use the existing version of the relevant CONSORT extension(s)”.5 So the main statement moves to 2025 while a given extension may legitimately still be on its older edition.
Common mistakes in designing an experimental study
Most of these are decided long before analysis, and none of them are recoverable afterwards. They are listed in roughly the order they tend to occur.
| Mistake | Stage | What it costs | What to do instead |
|---|---|---|---|
| Running a study the literature has already settled | Before design | Research waste, and an ethical failure rather than merely a practical one | The Declaration of Helsinki requires a design “based on a thorough knowledge of the scientific literature” that is “likely to produce reliable, valid, and valuable knowledge and avoid research waste”3 |
| Choosing the comparator for statistical convenience | Design | An unethical trial, and one many journals will not publish | Test against “the best proven intervention(s)”; placebo only where no proven intervention exists or paragraph 33’s narrow exception genuinely applies3 |
| Reporting blinding and treating allocation concealment as covered | Design and write-up | An avoidable risk-of-bias downgrade on a trial that may have been concealed perfectly well | Report the sequence, the concealment mechanism and who was blinded as three separate statements4 |
| Crossover with no washout, or an inadequate one | Design | Carry-over contaminates the second period, and the within-person comparison — the whole point of the design — is lost | Use crossover only for stable conditions with treatments that wash out, and test for period effects |
| A factorial trial powered only for the main effects | Sample size | An interaction that exists cannot be detected, and the two “independent” answers may both be wrong | Either power the trial for the interaction, or state the no-interaction assumption as an assumption |
| A controlled before-after study with one intervention site | Design | Exclusion from synthesis regardless of execution — EPOC excludes studies with only one intervention or control site1 | At least two intervention and two control sites, measured contemporaneously by identical methods1 |
| Randomising clusters but analysing individuals | Analysis | “Confidence intervals will be artificially narrow and P values will be artificially small”9 — a false positive produced by arithmetic | Account for clustering with a multilevel model or generalised estimating equations9 |
| Changing the primary outcome after seeing the data | Analysis | Triggers the RoB 2 domain “bias in selection of the reported result”6 | Prespecify one primary outcome with one definition in the protocol, and report it whichever way it falls8 |
| Analysing completers only | Analysis | Triggers “bias due to missing outcome data”6, and the flow diagram will show the gap5 | Analyse as randomised, account for missing data, and report losses with reasons |
| Appraising or reporting against the wrong instrument | Write-up | Reviewers notice immediately, and it signals the design was not understood | RoB 2 for randomised trials,6 ROBINS-I for non-randomised studies7 |
| Writing to a retired guideline | Write-up | A checklist that no longer matches what editors expect | CONSORT 2025 and SPIRIT 2025; the 2010 and 2013 versions have been superseded and should no longer be used58 |
It is the one on this list that produces a confident wrong answer rather than a weak one. When whole clinics or schools are randomised, people inside a cluster respond similarly, so treating them as independent individuals creates a “unit-of-analysis error” — the unit of analysis differs from the unit of allocation.9 The inflation is captured by the design effect, approximately 1 + (M − 1) × ICC, where M is the average cluster size and ICC the intracluster correlation. Intracluster correlations are often below 0.05, but “even small values can have a substantial impact on confidence interval widths” when clusters are large.9 A trial of 40 clinics with 50 patients each and an ICC of 0.02 has an effective sample size close to half its headline number.
Advantages and limitations
The advantage of a true experiment is singular and large: random allocation balances the groups on everything, including the characteristics nobody thought to measure. No amount of statistical adjustment in a non-randomised study reproduces that, because adjustment can only touch what was recorded.
The limitations are equally concrete:
- Feasibility. Randomisation is frequently ruled out by “ethical, political or logistical constraints”, and is always impossible for an evaluation designed after the programme ran.2
- Generalisability. Trial samples are selected, so internal validity is bought at some cost to external validity — which is why CONSORT 2025 moved generalisability into the trial limitations item rather than treating it as a separate claim.5
- Blinding is not always available. Many interventions cannot be masked, and the honest response is to report that and protect outcome assessment, not to imply blinding that did not happen.4
- Attrition. Missing outcome data is a RoB 2 domain in its own right, and it accumulates with follow-up length.6
- Ethics can forbid the cleanest comparison. A placebo arm is not available whenever a proven intervention exists, except in the defined circumstances of paragraph 33.3
Take-home message
What makes a study experimental is that someone allocates the intervention; what makes it a true experiment is that the allocation is random. Treat randomisation, allocation concealment and blinding as three separate things to design and three separate things to report — and write the protocol to SPIRIT 2025 and the paper to CONSORT 2025, because the 2013 and 2010 versions have been retired.
How to apply this to your work
- Check whether you are allocating at all. If nobody assigns the intervention, the study is observational and no amount of design language changes that.1
- Establish that the question is still open. The Declaration of Helsinki requires a thorough knowledge of the literature and the avoidance of research waste before you enrol anyone.3
- Settle the comparator on ethical grounds first. Best proven intervention unless paragraph 33’s exceptions genuinely apply.3
- Choose the geometry from the condition. Crossover only where the condition is stable and the effect washes out; cluster where contamination between individuals is otherwise certain.
- Specify the concealment mechanism, not just the sequence. They are separately reportable and separately appraised.4
- Write the protocol against SPIRIT 2025 (34 items) before recruiting.8
- Write the report against CONSORT 2025 (30 items), and check whether your design has an extension still on its earlier edition.5
- Read the five RoB 2 domains while the trial is still designable.6
Common questions
What is the difference between experimental and observational research?
Allocation. In an experimental study an investigator assigns the intervention — EPOC defines both randomised and non-randomised trials as “an experimental study in which people are allocated to different interventions”, differing only in whether the method is random.1 In an observational study nobody allocates; exposure is recorded as it occurs.
Is allocation concealment the same as blinding?
No. Concealment happens before allocation and prevents foreknowledge of the upcoming assignment, guarding against selection bias; blinding happens after allocation and consists of “measures taken to conceal group identity after allocation”, guarding primarily against ascertainment bias.4 A trial that cannot be blinded can still be perfectly concealed.
Should I still use CONSORT 2010?
No. CONSORT 2025 “supersedes the CONSORT 2010 statement, which should no longer be used”.5 The same applies to protocols: SPIRIT 2025 supersedes SPIRIT 2013, “which should no longer be used or cited”.8 Individual CONSORT extensions may still be on earlier editions, and the 2025 authors say to use the existing version of the relevant extension until it is updated.5
How many types of experimental design are there?
It depends what is being counted. There are three families — pre-experimental, true experimental and quasi-experimental — and within the true experimental family there are four common geometries: parallel, crossover, factorial and cluster randomised. Counts of “four types” usually come from mixing the two systems together.
Why is a pre-experimental design considered weak?
Because nothing in the design separates the intervention from everything else that happened at the same time — secular trend, maturation, regression to the mean. Cochrane EPOC lists uncontrolled before-after studies among designs that are “strongly discouraged”.1 It has low internal validity by construction, which is a property of the design rather than of how carefully it was run.
Can a randomised trial give the wrong answer?
Yes, and RoB 2 exists because of the specific ways it can. Bias can arise from the randomisation process itself, from deviations from intended interventions, from missing outcome data, from how the outcome was measured, and from which results were reported.6 Randomisation protects the comparison at baseline; it does not protect everything that happens afterwards.
Declared interest: Pubrica provides study design, statistical analysis and manuscript support commercially. Nothing on this page should be read as evidence that any design choice, or paid assistance with one, improves the likelihood of publication — no claim of that kind is made here.
References and sources
- Cochrane Effective Practice and Organisation of Care (EPOC). What study designs can be considered for inclusion in an EPOC review and what should they be called? EPOC Resources for review authors; 2017. read 20 October 2026.View on Cochrane EPOC
- White H, Sabarwal S. Quasi-Experimental Design and Methods. Methodological Briefs: Impact Evaluation No. 8. Florence: UNICEF Office of Research; 2014. read 20 October 2026.View on BetterEvaluation
- World Medical Association. WMA Declaration of Helsinki — Ethical Principles for Medical Research Involving Human Participants. Adopted June 1964; last amended by the 75th WMA General Assembly, Helsinki, October 2024. read 20 October 2026.View on the World Medical Association
- Schulz KF, Chalmers I, Altman DG, Grimes DA, Moher D, Hayes RJ. ‘Allocation concealment’: the evolution and adoption of a methodological term. J R Soc Med. 2018;111:216–224. read 20 October 2026, via the James Lind Library.View on the James Lind Library
- Hopewell S, Chan A-W, Collins GS, Hróbjartsson A, Moher D, Schulz KF, et al. CONSORT 2025 statement: updated guideline for reporting randomised trials. PLoS Med. 2025;22(4):e1004587. doi:10.1371/journal.pmed.1004587 read 20 October 2026.View on PLOS Medicine
- Sterne JAC, Savović J, Page MJ, Elbers RG, Blencowe NS, Boutron I, et al. RoB 2: a revised tool for assessing risk of bias in randomised trials. BMJ. 2019;366:l4898. doi:10.1136/bmj.l4898 read 20 October 2026.View the article PDF
- Sterne JAC, Hernán MA, Reeves BC, Savović J, Berkman ND, Viswanathan M, et al. ROBINS-I: a tool for assessing risk of bias in non-randomised studies of interventions. BMJ. 2016;355:i4919. doi:10.1136/bmj.i4919 read 20 October 2026, via the National Collaborating Centre for Methods and Tools registry entry.View on NCCMT
- Chan A-W, Boutron I, Hopewell S, Moher D, Schulz KF, Collins GS, et al. SPIRIT 2025 statement: updated guideline for protocols of randomised trials. PLoS Med. 2025;22(4):e1004589. doi:10.1371/journal.pmed.1004589 read 20 October 2026.View on PLOS Medicine
- Higgins JPT, Eldridge S, Li T. Chapter 23: Including variants on randomized trials. In: Cochrane Handbook for Systematic Reviews of Interventions, version 6.5. Cochrane; 2024. Chapter last updated October 2019. read 20 October 2026.View on Cochrane
- Moher D, Hopewell S, Schulz KF, Montori V, Gøtzsche PC, Devereaux PJ, et al. CONSORT 2010 explanation and elaboration: updated guidelines for reporting parallel group randomised trials. Journal of Clinical Epidemiology. 2010;63:e1–e37. doi:10.1016/j.jclinepi.2010.03.004 read 20 October 2026 from an institutional repository copy.View the full text
- Campbell DT, Stanley JC. Experimental and Quasi-Experimental Designs for Research. Boston: Houghton Mifflin; 1963. Three pre-experimental designs, p. 6. read 20 October 2026.View the full text
How to cite this article
Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.
This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.
Settle these before the first participant is enrolled.
[ ] Someone allocates the intervention — otherwise it is observational [ ] The literature confirms the question is still open [ ] Comparator justified: best proven intervention, or a stated exception [ ] Ethics approval and trial registration in place [ ] Geometry chosen from the condition, not from convenience [ ] Allocation sequence generation specified [ ] Allocation concealment mechanism specified — separately [ ] Who is blinded after allocation stated, or stated as not possible [ ] Primary outcome prespecified, with one definition [ ] Sample size justified, including the interaction if factorial [ ] Crossover: washout long enough; period effects to be tested [ ] Cluster: at least 2 intervention and 2 control sites; clustering in the analysis [ ] Protocol written to SPIRIT 2025 (34 items) [ ] Report planned against CONSORT 2025 (30 items) [ ] Five RoB 2 domains reviewed while the design can still change
We help you match the trial geometry to the condition, specify allocation, concealment and blinding as three separate decisions, and build the protocol against SPIRIT 2025 so the eventual CONSORT report writes itself.