Study designs and the questions they answer
Study designs differ in how they assign interventions, select participants and observe exposures and outcomes over time. The question determines which design is appropriate; a hierarchy of evidence provides orientation but cannot replace assessment of an individual study.
Questions and designs
The hypothesis is defined before the most appropriate or feasible design for testing it is selected. Different questions require different research evidence: designs include randomised trials, cohort and case-control studies, cross-sectional surveys, case reports, qualitative studies and systematic reviews.
Randomised trials can estimate intervention effects, while rare or long-term harms may be better investigated through cohort or case-control studies. Disease causation is generally investigated through cohort studies or, for rare conditions, case-control studies, although randomised trials can provide stronger causal inference where applicable. Diagnostic accuracy can be investigated through a cross-sectional comparison with an accepted reference standard; prognosis generally uses longitudinal cohorts, and prevalence uses cross-sectional surveys. Economic questions can arise with any study question, with randomised trials preferred for cost-effectiveness analysis; questions about values, descriptions and perceptions call for qualitative research.
Intervention assignment and timing
In experimental or interventional studies, investigators deliberately assign treatment; assignment may be randomised or nonrandomised. In observational studies, investigators do not assign treatment. Observational designs include cohort, case-control and cross-sectional studies, ecological studies using group-level data, database analyses with or without comparisons, and case series.
Prospective studies begin before outcomes are known; retrospective studies analyse past events whose outcomes are already known. Randomised trials are prospective, while nonrandomised studies may be prospective or retrospective.
Randomised and nonrandomised intervention studies
A randomised controlled trial allocates participants by chance to two or more groups, gives the intervention to the experimental group and an alternative treatment, placebo or no treatment to the control group, then follows participants to compare outcomes.
Quasi-experimental studies evaluate investigator-chosen interventions or policy changes without randomisation, often when a randomised trial is not logistically feasible or ethical. A before–after comparison of outcomes around a policy change can investigate whether it was beneficial, since a policy can have unintended adverse effects. Without randomisation, confounding can undermine causal interpretation.
Nonrandomised comparisons are vulnerable to selection bias when factors influencing treatment choice also affect outcomes: severely ill patients may volunteer for an experimental treatment while less severely ill patients prefer standard care. Retrospective comparisons depend on the availability, accuracy and completeness of records that may not have been collected for research. Comparisons of prospectively assessed patients with historical controls may also involve different selection criteria and treatment protocols.
Cohort studies
A cohort is a defined group whose members may share characteristics such as age, exposure, disease or medication use. Cohort studies follow groups characterised by exposure over time to observe outcomes, usually prospectively but sometimes retrospectively using past exposure measurements. For example, exposed and unexposed groups can be followed after an industrial radiation accident.
Cohorts can investigate possible causes of disease, prognosis, exposures that cannot ethically be assigned and rare exposures. An inception cohort followed over time can estimate overall prognosis well, although it provides weaker estimates of treatment effects.
Prospective cohorts permit planned collection of exposures and baseline characteristics, but follow-up can be long, particularly for infrequent outcomes, and loss to follow-up can be high. Historical cohorts use information already available, but tracing participants may be difficult and record quality or recollection may be poor. For example, records can compare uterine-cancer outcomes in women previously exposed to hormone replacement therapy with outcomes in otherwise similar unexposed women. Ambispective or ambidirectional cohorts combine retrospective and prospective information across past, present and future time points.
Case-control studies
Case-control studies select people with a specified disease or outcome and suitable controls without it, then compare antecedent exposures or risk factors. They can investigate possible disease causes or unwanted treatment effects and are usually retrospective. Controls are selected to resemble cases in relevant characteristics such as age, occupation and education.
The design is useful for rare diseases, for which prospective studies might require infeasibly large enrolment, and may have more power than a cohort study when larger case-control samples are easier to obtain.
Cross-sectional studies
Cross-sectional studies observe a population at one time rather than following it over time; a census is a special form. They can describe disease prevalence, patient characteristics and possible risk factors, or investigate associations. The quick, single-occasion snapshot can assess multiple outcomes and exposures, support patient surveys and multiple prevalence estimates, and aid descriptive analysis, hypothesis generation and resource planning.
The design does not establish the sequence of exposure and outcome, so it cannot distinguish cause from effect. It is unsuitable for measuring incidence or studying infrequent or short-duration conditions. Migration, low response, survival bias and misclassification from recall can affect prevalence estimates. High-mortality conditions may be underrepresented because prevalence can remain low despite high incidence. A single snapshot may also be poorly suited to diseases with a relapsing–remitting course.
Case reports and series
Individual case reports and series can signal rare treatment harms or benefits, new diseases or conditions, or rare disease manifestations. They may reveal unexpected adverse events, clues to pathogenesis, unexpected associations, unique therapies or anatomical variations, including information missed by larger studies. Such findings generally need formal verification through stronger epidemiological studies and do not constitute systematically derived evidence. For rare harms, a systematic review of case reports provides more reliable evidence than a haphazard selection.
Design and measures of association
A cohort that starts with people free of the outcome can estimate its risk and compare risks between exposure groups. For example, currently healthy people can be enrolled, their exercise habits recorded over several years and subsequent heart attacks counted.
Case-control sampling instead starts with a chosen number of cases and controls; the total population at risk is not known, so the comparison uses odds of exposure rather than directly estimating disease risks. Cross-sectional comparisons use prevalence ratios or prevalence odds ratios because exposures and outcomes are measured at one time.
Qualitative and mixed-methods research
Qualitative studies investigate social, emotional and experiential phenomena in healthcare, including issues underlying clinical decisions. Approaches include case studies of individuals or groups; ethnography, studying people in their surroundings; phenomenology, exploring lived experience; and ethnomethodology, examining how people make sense of experiences. Methods include focus groups, in-depth unstructured or semi-structured individual interviews, participant observation and direct observation. Common criteria for scientific rigour are credibility, transferability, dependability, confirmability and reflexivity. Mixed-methods research combines quantitative and qualitative methods in one study.
Reviews and synthesis
A systematic review identifies, appraises and synthesises primary research to answer a focused question, following a strict protocol intended to include relevant research and standardise appraisal and synthesis. A reproducible database search and predefined eligibility criteria distinguish systematic selection from a convenient collection of papers. Searches commonly span databases such as Web of Science, Embase and PubMed, and systematic reviews of randomised trials often inform guidelines and evidence-based medicine.
Meta-analysis mathematically combines results from two or more primary studies conducted under similar conditions, potentially increasing the precision of a review's overall result. It may estimate a pooled measure such as an odds ratio or risk ratio. Meta-analyses virtually always include a systematic literature review, but systematic reviews need not include meta-analysis. The quality of a systematic review depends on how effectively its scientific methods minimise bias and error. An umbrella review brings together evidence from previously published reviews.
Hierarchies of evidence
Evidence hierarchies differ with the question: treatment effects, diagnosis, prognosis, disease frequency, causation, harms and economic value need different evidence. Hierarchies can help locate evidence, but study quality also depends on conduct, size and relevance, and novel or hybrid designs may not fit the ranking. A review of small, poor-quality trials may rank above a large, well-conducted multicentre trial despite these limitations. Studies should not be interpreted in isolation, and risk–benefit assessment should draw on different kinds of research.
A commonly used treatment-evidence ranking runs from randomised controlled or comparative trials, through nonrandomised controlled or comparative studies and uncontrolled or noncomparative studies, to expert opinion and case reports. Some systems place meta-analyses of randomised trials in the highest category; others omit them because they are not primary evidence.
One example hierarchy for intervention effectiveness orders evidence as follows.
- Systematic review with homogeneous randomised controlled trials.
- Individual high-quality randomised controlled trial with narrow confidence intervals.
- Systematic review with homogeneous cohort studies.
- Individual high-quality cohort study or poor-quality randomised trial, for example with less than 80% follow-up.
- Systematic review with homogeneous case-control studies.
- Individual high-quality case-control study.
- Case series and poor-quality cohort or case-control studies.
- Expert opinion without explicit critical appraisal.
An example prognosis hierarchy instead prioritises inception cohorts.
- Systematic review with homogeneous inception cohorts, validated in different populations.
- Individual inception cohort with at least 80% follow-up.
- All-or-none case series.
- Systematic review with homogeneous retrospective cohorts or untreated control groups from randomised trials.
- Retrospective cohort or follow-up of untreated controls in a randomised trial.
- Outcomes research.
- Case series and poor-quality prognostic cohorts.
- Expert opinion without explicit critical appraisal.
A general illustrative pyramid orders designs from meta-analysis and systematic review through randomised controlled trials, quasi-experimental studies, prospective cohorts, retrospective cohorts, case-control studies and cross-sectional studies to case reports. Another illustrative pyramid separates systematic reviews or meta-analyses, double-blind randomised trials, other randomised trials, nonrandomised trials, cohorts, case-control studies, case series and cross-sectional studies, case reports, editorials and reviews, and animal, in-vitro and cell models. Designs with greater validity can be more expensive, time-consuming and difficult to manage.