How to Select the Right Study Design for Your Original Research Article
A study design is a tool for answering a particular kind of question. The selection problem is therefore narrower than it appears. Most of the apparent choice disappears once the question is stated precisely. What remains is a check on feasibility, which can rule a design out but cannot justify one. There is also a consequence for reporting, since the design determines which checklist the manuscript must satisfy.
- Which question types each family of designs can and cannot answer
- Why feasibility rules designs out without justifying the one that remains
- Which reporting checklist each design commits an author to
- “A randomised trial is the strongest design, so use one where possible.” Strength is relative to a question. A trial answers questions about the effects of interventions. It answers questions about prevalence, prognosis and diagnostic accuracy poorly or not at all.
- “Observational designs carry a lighter ethical burden.” The requirement for independent review attaches to research, not to a design. Any activity containing an element of research is to undergo review.
What this article covers
This article concerns the decision that precedes a protocol: which design will answer the question that has been posed. It does not compare the observational designs in detail, since a companion article does that. Nor does it describe how to execute any particular design. The focus is the selection itself, and the three things that govern it. Those are the question, the constraints, and the reporting obligation that follows from whichever design is chosen.
The order of those three matters more than their content. A design selected from the question can be checked against the constraints afterwards. A design selected from the constraints produces a study that answers a question nobody asked. A design selected from a ranking does the same. The second failure is harder to detect in a finished manuscript. The methods will be internally consistent, and the writing will conceal the substitution.
What each observational design can and cannot measure is covered in comparing observational study designs. The obligations attaching to a design before any form is filed are covered in ethical considerations in research design. Writing the study up once the design has done its work is covered in how to structure a research article.
The question determines the design
Designs are not interchangeable instruments of differing quality. Each produces a particular kind of estimate. A question calling for one kind of estimate cannot be answered by a design that produces another. Stating the question in a form that names the estimate is therefore most of the selection work.
Question types and what each requires
Questions about the effects of an intervention require a design in which the investigator assigns the exposure. Randomised allocation is what permits the comparison groups to be treated as exchangeable. The reporting guideline for such studies is accordingly written around allocation, concealment and blinding.1 The core checklist for these is written principally for the two-group parallel design. Separate extensions cover adaptive, cluster, crossover, factorial, non-inferiority, pragmatic, multi-arm, n-of-1, pilot and within-person trials.1
Questions about how common a condition is, or about associations where assignment is impossible or unethical, require an observational design. Three of these are covered by a single reporting guideline, which addresses cohort, case-control and cross-sectional studies together.3 The guideline's structure records where they differ. Loss to follow-up and follow-up time are items specific to cohort studies. Matching of cases and controls is specific to case-control studies. Accounting for the sampling strategy in analysis is specific to cross-sectional studies.3
Questions about how well a test identifies a condition require neither of the above. Diagnostic accuracy studies compare an index test against a reference standard. They have a reporting guideline of their own, running to thirty essential items.2 That guideline was updated specifically to incorporate newer evidence on the sources of bias and variability peculiar to accuracy estimates.2 A trial design applied to such a question would answer something else.
A question is specified sufficiently to select a design once it names the quantity to be estimated. Prevalence, incidence, a relative effect, a hazard, a sensitivity: each of these points at a family of designs and away from the others. A question that does not name a quantity is not yet ready. No design can be chosen for it.
The hierarchy of evidence and its limits
Rankings of designs are widely reproduced and widely misapplied. A hierarchy places randomised trials above cohort studies, and cohort studies above case-control and cross-sectional designs. The ranking is defensible for one question type and misleading for the rest. It ranks designs by susceptibility to confounding in estimating intervention effects.
Applied to other questions it breaks immediately. A randomised trial cannot estimate prevalence in a population, since it recruits a selected sample under eligibility criteria. It is a poor instrument for rare harms. Such a trial is powered for an efficacy outcome and followed for a trial duration. It cannot establish diagnostic accuracy, which requires an index test and a reference standard rather than two allocated arms. In each case the design further down a hierarchy is the correct one.
The methodological guidance frames the requirement without reference to any ranking. Research must have a scientifically sound and rigorous design, and the protocol must describe and justify that design.5 Justification is relative to the objective rather than to a general ordering. The ethical guidelines make the same point from the other direction, holding that scientific value means a study is capable of producing “reliable, valid information” for the question it addresses.6
Figures: choosing a study design
Figure 1 links a research question to its study design and to the reporting checklist that follows from it. Figure 2 sets out the order in which the design decision is made, including where feasibility is assessed. Authors can use them to justify the chosen design in the methods section.
The right-hand column is a consequence rather than a choice. Selecting the design selects the checklist with it.
What to look for. Each row reads in one direction only. A question admits a design family, and that family carries its checklist. The reverse order is the error this diagram is drawn to make visible. It runs from a familiar checklist to a design, and then to whatever question that design can answer. Four rows are shown here, and the library of guidelines covers a good many more study types than these.
Schematic. The guidelines named are the principal ones for each family, and most of them have extensions for particular variants.
Feasibility enters after the design has been chosen. Failing it sends the author back to the question, not down the hierarchy.
What to look for. The dashed return path is the part most often missing in practice. The required design sometimes turns out to be unaffordable or impermissible. The common response is to retain the question and substitute a weaker design for it. The correct response is to change the question to one the available design can answer. The manuscript should then say so.
Schematic. Step 3 covers access to participants, time, funding, instrumentation and whether the exposure may ethically be assigned.
Constraints that are not justifications
Three considerations are routinely presented as reasons for choosing a design, and none of them is one. Each of the three can rule a design out. None can make the remaining design appropriate for a question it does not fit.
Feasibility
Access to participants, available time, funding and instrumentation are genuine limits, and a design that exceeds them cannot be executed. The logical error lies in what follows. Infeasibility of the correct design does not make an incorrect design correct. The second design still produces the wrong kind of estimate.
The honest response is to change the question rather than the design. The narrower question that the available design can answer should then be stated. A cross-sectional survey conducted because a cohort could not be funded is a legitimate study of prevalence. It is not a weakened study of incidence, and presenting it as one misdescribes what was done.
Available data
An existing dataset is the most common driver of design selection and the least defensible. A registry, a chart archive or a trial's unused variables will support some analyses and not others. Which analyses they support was determined by how the data were collected, not by what anyone now wishes to ask.
The ethical guidelines identify the limiting case directly. Some studies are designed to promote a product rather than to produce knowledge. Such studies, sometimes called seeding trials, are held to violate the requirement for social value.6 The same logic reaches a study designed around a dataset rather than around a question. A study whose objective was reverse-engineered from available variables is not producing the knowledge that justified collecting them.
Ethical constraint as a real constraint
The third consideration differs from the first two because it is binding rather than practical. Where an exposure cannot ethically be assigned, the allocated design is unavailable, and no amount of funding changes that. Research is permissible only where the importance of the objective outweighs the risks and burdens. Risks must also be assessed in advance and minimised.5
This constraint is also the one that most often explains why an observational design is correct rather than second best. Questions about the effects of smoking, of occupational exposure or of a behaviour cannot be answered by assignment. The observational design is the only instrument available. Describing such a study as a compromise misstates its standing. It is the appropriate design for a question that forecloses the alternative.
A constraint tells an author which designs are unavailable. The question tells an author which design is required. Where those two sets do not intersect, the study cannot be done as conceived. The question must be narrowed until they do.
The reporting guideline follows the design
Selecting a design has a consequence that authors frequently discover late. Each design family has a reporting guideline written for it, and journals increasingly require the matching checklist at submission. Knowing which checklist applies while the protocol is being written is considerably easier than reconstructing the required information afterwards.
The scale of the catalogue
The size of the guideline library surprises most authors. The EQUATOR Network's searchable collection lists 709 guidelines. They are filtered by study type across seventeen categories, by clinical area across roughly fifty specialties, and by section of the report.4 Those categories include experimental studies, observational studies, diagnostic and prognostic studies, systematic reviews, qualitative research, quality improvement and economic evaluations.4
Most of that number consists of extensions rather than base guidelines. The library lists extensions for paediatric trials and for equity analyses. It also lists them for trials of artificial intelligence interventions, for social and psychological interventions, and for early-phase dose-finding trials.4 Similar families exist for observational studies, diagnostic accuracy, prediction models, scoping reviews, case reports and quality improvement projects.4 The practical consequence is that a specialised design almost certainly has a specialised checklist. Searching for it is worth the few minutes it takes.
Randomised trials
The guideline for randomised trials was revised in 2025, and the revision is not optional. Its authors state that “The CONSORT 2025 statement supersedes the CONSORT 2010 statement, which should no longer be used”.1 The updated checklist carries thirty items in all.1
Seven items are new, and several of them bear on decisions taken at the design stage rather than the writing stage. They cover data sharing, authors' conflicts of interest, and patient and public involvement. They also cover eligibility criteria for sites and interventionists, how harms were assessed, the analysis population and missing data, and intervention delivery and concomitant care.1 An author who reads the checklist only at submission will find a problem. Patient and public involvement, in particular, cannot be added retrospectively.
Three items were substantially revised, covering access to the protocol and statistical analysis plan, protocol changes, and outcome reporting.1 One item was removed, since generalisability was folded into the item on trial limitations.1 The revision also integrated content from the harms, outcomes and non-pharmacological treatment extensions, and from the intervention description guideline.1
Observational studies
Cohort, case-control and cross-sectional studies share one checklist of twenty-two numbered items, expanded to thirty-six rows by sub-items.3 Design-specific guidance sits inside the item descriptions rather than in separate lists.3 Four sub-items apply to a single design each. Identifying which of the three applies is therefore part of using the checklist correctly.3
Diagnostic accuracy and other designs
Diagnostic accuracy studies have their own guideline of thirty essential items. It was updated in 2015 to reflect newer evidence on sources of bias and variability.2 Its authors describe the problem it addresses in strong terms, calling incomplete reporting “a major source of avoidable waste in biomedical research”.2 The same framing applies across the catalogue, and it is the reason journals have adopted these checklists at all.
Designs outside these three families are covered too. A guideline exists for prediction models, for qualitative studies, for case reports, for quality improvement projects and for economic evaluations.4 Searching the library by study type takes less time than drafting a methods section twice.
The limits of a reporting guideline
Reporting guidelines are frequently mistaken for design guidance, and the confusion runs in a direction that damages studies. A checklist describes what a finished report must contain. It does not describe how the study should have been conducted. Nor can it be used in place of a design decision.
The scope the guidelines claim for themselves
The point is stated explicitly alongside the observational checklist. Reporting guidelines do not prescribe how research should be designed or conducted; they help authors describe transparently what they did, why they did it, and what they found.3 The checklist exists to demonstrate that a manuscript follows the guideline. It is intended for use with the full guidance rather than on its own.3
That boundary has a corollary worth stating plainly. A study with an inappropriate design can satisfy its checklist completely. Every item can be reported, every sub-item addressed, and the resulting manuscript will still answer the wrong question. Checklist compliance is evidence about the report, not about the study.
Where the design obligation actually sits
The obligation to design the study well sits in the ethical instruments rather than the reporting ones. Research must have a scientifically sound and rigorous design and execution. The protocol must clearly describe and justify the design and methods.5 The ethical guidelines make the connection to the design explicit, holding that a study “cannot be socially valuable without appropriate and rigorous research methods”.6
Journal editors apply a parallel standard at the other end. Approval by a review body does not prevent editors from forming their own judgement about whether research was conducted appropriately.7 A completed checklist is therefore not a defence against a methodological objection, any more than committee approval is.
Using the checklist early
The productive use of a reporting guideline is at protocol stage. There it functions as a list of decisions that will have to be defensible later. Several of the newer trial items illustrate this directly. Patient and public involvement, the method of assessing harms and the analysis population are all design commitments. The checklist will later ask about each of them.1
Read at that point, a checklist prompts decisions. Read at submission, it prompts explanations for decisions that were never made. The difference between those two experiences is largely a matter of when the author first opened the document.
Ethical review applies to every design
A belief circulates that observational work carries a lighter ethical burden, and that review is principally a requirement for trials. The instruments do not support that belief. The requirement attaches to research as an activity. The design determines what the review will examine, not whether review is needed.
The requirement as stated
Before research begins, the protocol must be submitted for consideration, comment, guidance and approval to a research ethics committee.5 The committee must be independent and transparent, and must retain the power to monitor, amend or halt an ongoing study.5 Nothing in that provision is conditioned on the design.
The foundational report resolves ambiguous cases in the same direction. Where research and practice occur together, it holds that “if there is any element of research in an activity, that activity should undergo review”.8 The test is the presence of a research element, not the presence of an intervention or an allocation.
Variation in the content of review
What does vary is the content of the review. An allocated design raises questions about the comparator, about stopping rules and about post-trial provisions. An observational design using existing records raises questions about consent waivers, about privacy and about the governance of secondary data. A qualitative design raises questions about identifiability in quotation. The questions differ, while the obligation does not.
One consideration cuts across all of them. Groups and individuals must be selected for scientific reasons. Groups unlikely to benefit should not bear a disproportionate share of the risks and burdens.6 That requirement bears on eligibility criteria, which are settled while the protocol is being written rather than by the committee afterwards.
Registration, whatever the design
A further obligation is widely assumed to apply only to trials. Research must be registered in a publicly accessible database before the first participant is recruited.5 The reporting duty attached to it is equally general. Negative and inconclusive results must be published or otherwise made public alongside positive ones.5
Registration before recruitment is also what makes a design decision auditable. The outcomes named in the register are the ones the study undertook to report. The design named there is the one it undertook to run. Divergence from either becomes visible to a reader. That is the mechanism by which the choices discussed here can be checked at all.
Common mistakes
| Fault | Stage | What it costs | The remedy |
|---|---|---|---|
| The design is chosen before the question names a quantity | Question | The study produces an estimate nobody asked for | State the quantity to be estimated, then select the family that yields it |
| A design is picked from a hierarchy of evidence | Selection | A ranking built for intervention effects is applied to other questions | Match the design to the question type, not to a general ordering |
| A weaker design substituted when the correct one is infeasible | Selection | The manuscript answers a different question than it claims | Narrow the question until the available design can answer it |
| The objective reverse-engineered from an existing dataset | Selection | Analyses limited by collection decisions nobody is revisiting | Start from the question, and treat the dataset as a feasibility check |
| An observational design described as a compromise | Writing | The correct design for an unassignable exposure is misrepresented | State why assignment was impossible or impermissible |
| CONSORT 2010 cited for a randomised trial | Reporting | A superseded guideline, missing seven items now required | Use CONSORT 2025, which its authors say replaces the 2010 version |
| The checklist consulted at submission rather than at protocol stage | Reporting | Items such as public involvement cannot be added retrospectively | Open the checklist while the protocol is being drafted |
| Checklist compliance treated as methodological adequacy | Reporting | A badly designed study passes its checklist completely | Treat the guideline as a reporting standard, not a design standard |
| Ethics review assumed unnecessary for observational work | Approval | A requirement that attaches to research, not to a design, is missed | Submit any activity containing an element of research for review |
| Registration treated as a requirement for trials alone | Registration | The design and outcomes become unauditable after the fact | Register before the first participant, whatever the design |
The sixth row will apply to a large proportion of manuscripts currently in preparation. The revision is recent, and the superseded version remains widely cited. The third row is the one that does the most damage. The substitution it describes leaves no trace in the finished methods section. A reader sees a competently executed study and has no way to know that it replaced a different one.
Selecting a design is mostly a matter of stating the question well enough that the design becomes obvious. A question that names the quantity to be estimated points at one family and away from the rest. Hierarchies of evidence are unhelpful here, since they rank designs for intervention effects alone. Feasibility and available data can rule designs out, and neither can justify the design that remains. Where the required design is unavailable, the question should be narrowed rather than the design weakened. Each design family carries a reporting guideline, and the trial guideline was replaced in 2025. The checklist is most useful at protocol stage rather than at submission. Ethical review and prospective registration attach to research rather than to any particular design.
How to apply this to your work
- State the question until it names a quantity, such as a prevalence, an effect or a sensitivity.
- Select the design family that produces that quantity, without reference to any ranking.
- Test feasibility afterwards, treating it as a filter rather than as a reason.
- Narrow the question if the design proves unavailable, and say in the manuscript that this happened.
- Identify the matching reporting guideline before the protocol is finished.
- Check the guideline's current version, since the trial guideline was superseded in 2025.
- Submit for ethical review whatever the design, and register before the first participant.
- Justify eligibility criteria on scientific grounds, and record the justification.
Common questions
Is a randomised trial always the strongest design?
Only for questions about the effects of interventions, which is what the ranking behind that claim was built for. A trial cannot estimate prevalence in a population, is poorly suited to rare harms, and cannot establish diagnostic accuracy, which requires an index test and a reference standard. The instruments require a design that is scientifically sound for its objective. They do not require a design high in a general ordering.5
Which reporting guideline applies to my design?
The library of guidelines is searchable by study type. It lists 709 guidelines across seventeen study-type categories. These include experimental, observational, diagnostic and prognostic, systematic review, qualitative, quality improvement and economic evaluation.4 Randomised trials fall under CONSORT 2025,1 cohort, case-control and cross-sectional studies under STROBE,3 and diagnostic accuracy studies under STARD 2015.2
Can CONSORT 2010 still be used?
No. Its successors state the position directly: “The CONSORT 2025 statement supersedes the CONSORT 2010 statement, which should no longer be used”.1 The updated checklist has thirty items, seven of them new. The new items cover data sharing, conflicts of interest, patient and public involvement, eligibility of sites and interventionists, assessment of harms, analysis population and missing data, and intervention delivery.1
Does one checklist cover all three observational designs?
Yes, with design-specific sub-items inside it. The checklist carries twenty-two numbered items across thirty-six rows, covering cohort, case-control and cross-sectional studies.3 Four sub-items apply to one design each. Loss to follow-up and follow-up time belong to cohort studies, matching to case-control studies, and the sampling strategy to cross-sectional studies.3
What should be done when the required design is not feasible?
Narrow the question rather than weaken the design. A design that cannot produce the required quantity does not answer the original question less well. What it answers is a different question entirely. Stating the narrower question honestly is the better course. The alternative presents a substituted design as a compromised version of the intended one, and leaves no trace in the methods section.
Does an observational study need ethics committee approval?
Yes. The protocol must be submitted for consideration, comment, guidance and approval before research begins. The provision is not conditioned on design.5 The foundational report resolves ambiguous cases the same way, holding that “if there is any element of research in an activity, that activity should undergo review”.8 What changes with the design is what the review examines.
Can a reporting guideline tell me how to design the study?
No, and the guidelines say so. Reporting guidelines do not prescribe how research should be designed or conducted; they help authors describe transparently what they did and found.3 A poorly designed study can satisfy its checklist completely. Compliance is therefore evidence about the report rather than about the study.
Does prospective registration apply to designs other than trials?
The requirement as stated is general. Research must be registered in a publicly accessible database before the first participant is recruited.5 The accompanying reporting duty is equally general, since negative and inconclusive results must be published or otherwise made publicly available.5
Declared interest: Pubrica provides research and medical writing support commercially. Nothing on this page should be read as suggesting that a manuscript can rescue an unsuitable design. A design that does not fit its question cannot be written around. Where the design and the question have come apart, the useful work is to restate the question. That conversation belongs before the data are collected rather than after.
References and sources
- Hopewell S, Chan A, Collins GS, et al. CONSORT 2025 statement: updated guideline for reporting randomized trials. JAMA. 2025;333(22):1998–2005. doi:10.1001/jama.2025.4347 read 21 October 2026 at the journal.View the full text
- Bossuyt PM, Reitsma JB, Bruns DE, Gatsonis CA, Glasziou PP, Irwig L, et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. Clinical Chemistry. 2015;61(12):1446–1452. doi:10.1373/clinchem.2015.246280 read 21 October 2026 from the UMC Utrecht research portal; the statement was published simultaneously in several journals, and this is the version consulted.View the record
- STROBE checklist: items that should be addressed in reports of cohort, case-control and cross-sectional studies. EQUATOR Network. read 21 October 2026, in full.View the checklist
- EQUATOR Network. Library of reporting guidelines, searchable by study type, clinical area and section of report. read 21 October 2026; the listing runs to roughly 111,000 characters and the portion returned covered the first 356 of the 709 entries, so the extensions named here are drawn from that portion.View the library
- World Medical Association. WMA Declaration of Helsinki: ethical principles for medical research involving human participants. Adopted June 1964; current revision adopted by the 75th WMA General Assembly, Helsinki, October 2024. read 21 October 2026, in full.View the Declaration
- Council for International Organizations of Medical Sciences, in collaboration with the World Health Organization. International ethical guidelines for health-related research involving humans. Fourth edition. Geneva: CIOMS; 2016. read 21 October 2026; the guidelines cited here are 1 and 3.View the full text
- International Committee of Medical Journal Editors. Recommendations for the Conduct, Reporting, Editing, and Publication of Scholarly Work in Medical Journals: protection of research participants. read 21 October 2026, in full.View the section
- National Commission for the Protection of Human Subjects of Biomedical and Behavioral Research. The Belmont Report: ethical principles and guidelines for the protection of human subjects of research. Washington: Department of Health, Education, and Welfare; 18 April 1979. read 21 October 2026, in full.View the full text
One further source could not be read. The publisher's page for the STROBE statement itself declined the request, so the checklist hosted by the EQUATOR Network is cited above in its place. Nothing attributed here to that checklist depends on the statement's accompanying article.
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.
Work through these before the design is settled.
THE QUESTION [ ] The question names the quantity to be estimated [ ] That quantity is a prevalence, an effect, a hazard, an accuracy or similar [ ] The population, exposure or test, and outcome are each specified [ ] The question has not been adjusted to suit an available dataset THE DESIGN [ ] The design family follows from the quantity, not from a hierarchy [ ] No ranking of designs was used to make the selection [ ] If the exposure is assigned, the allocation method is specified [ ] If the exposure is not assigned, the reason is recorded [ ] For a diagnostic question, index test and reference standard both named [ ] Any specialised variant of the design is identified by name FEASIBILITY, AS A FILTER [ ] Access to the required participants confirmed [ ] Time frame compatible with the measurement required [ ] Funding and instrumentation sufficient for the design, not for a substitute [ ] Assignment of the exposure is ethically permissible [ ] If the design fails any of these, the QUESTION was narrowed, not the design [ ] Any such narrowing is stated in the manuscript REPORTING OBLIGATION [ ] The matching reporting guideline identified before the protocol is finished [ ] The current version confirmed, not an earlier one [ ] For a randomised trial, CONSORT 2025 rather than CONSORT 2010 [ ] Any relevant extension for the specific variant located [ ] Checklist items that are design commitments addressed now, not at submission [ ] Patient and public involvement considered at this stage if applicable [ ] Method of assessing harms decided at this stage [ ] Analysis population and handling of missing data defined ETHICS AND REGISTRATION [ ] Protocol submitted for independent review, whatever the design [ ] Eligibility criteria justified on scientific grounds [ ] No group unlikely to benefit bears a disproportionate burden [ ] Study registered publicly before the first participant [ ] Outcomes named in the registration record [ ] A plan exists to report negative and inconclusive results
We help you state a question precisely enough to select a design for it. We check the design against the question it has to answer, and identify the reporting guideline it commits you to.