Which study design answers your question?

Research is a systematic investigation designed to produce new, generalisable knowledge (see the what is research chapter). Research design is the plan that states whom you will study, what you will measure, when, and how you will compare.

1. Quantitative, observational, descriptive: case report, case series, cross-sectional (prevalence) survey, descriptive ecological study. Example: prevalence of anaemia in adolescent girls.
2. Quantitative, observational, analytical: ecological, cross-sectional analytical, case-control, cohort (prospective, retrospective, ambispective), and hybrids (nested case-control, case-cohort). Example: smoking and lung cancer.
3. By time: a single cross-sectional survey, repeated cross-sectional surveys (a new sample each round), and longitudinal or panel studies (the same people measured repeatedly).
4. Quantitative, experimental: randomised controlled trial, field trial, community trial, cluster randomised trial. Without randomisation: quasi-experimental designs. Example: a new dressing compared with standard dressing.
5. Qualitative: phenomenology, grounded theory, ethnography, case study, narrative research. Example: why patients stop tuberculosis treatment.
6. Mixed methods: quantitative and qualitative parts in one study.
7. By purpose (research methods texts): exploratory, descriptive, causal (explanatory).

Why design matters: it decides whether the time order of exposure and outcome is known, how bias and confounding are controlled, which measure of association and sample size formula apply, what the study costs, and whether a conclusion about cause is justified. STROBE item 1(a) asks authors to name the design in the title or abstract.

Role of qualitative research: it explores reasons and experiences, generates hypotheses, helps build questionnaires, and explains quantitative findings.

Literature review: search databases with a written strategy, screen titles and abstracts, record each study in a table (design, population, exposure, outcome, main result, limitations), and group the studies by theme or design before writing. See the literature search chapter.

Before concluding cause: show an association, show that the cause came before the effect, and rule out chance, bias and confounding (see the bias and validity chapter).

Components of a research proposal: title, background and rationale, review of literature, research question and objectives, hypothesis, methods (design, setting, study population, eligibility criteria, sampling, sample size, variables and definitions, data collection tools, intervention if any, statistical analysis plan), ethical considerations and consent, timeline, budget, references, and annexures (consent form, case record form). See the research protocol chapter.

Definition: designs in which the investigator does not assign the exposure or randomise participants, but observes and measures variables as they occur.

1. Descriptive: case report, case series, surveys.
2. Correlational and ecological: the relationship between variables, or between group rates, without assigning either.
3. Cross-sectional analytical: exposure and outcome measured at the same time.
4. Case-control (called causal-comparative or ex post facto in some research methods texts): groups defined by the outcome and compared on past exposure.
5. Cohort: groups defined by exposure and followed for outcomes.

Uses: when the exposure cannot ethically be assigned (smoking, poverty), to generate hypotheses, for rare outcomes, and to measure prevalence. Limits: confounding, selection and information bias, and weaker evidence for cause than an experiment.

Characteristics of a true experiment: manipulation (the investigator gives the intervention), control (a comparison group) and randomisation (chance decides who receives the intervention). Quantitative designs as a whole are classified in the first outline above.

1. Pre-experimental (no comparable control group): one-shot case study, one-group pretest-posttest (knowledge before and after a health talk), static-group comparison.
2. True experimental: pretest-posttest control group (randomised groups measured before and after), Solomon four-group (adds groups without a pretest, to detect an effect of the pretest itself), posttest-only control group.
3. Quasi-experimental: time series (repeated measurements before and after a policy change), nonequivalent control group (a programme in one school, another school for comparison).
4. Clinical trial designs: parallel group, cross-over, factorial, cluster randomised. Framework: superiority, equivalence or non-inferiority (CONSORT 2025 item 9). In a 2 × 2 factorial trial each patient is randomised to drug A or placebo and to drug B or placebo, which answers two questions in one trial if the drugs do not interact (ISIS-2, 1988: streptokinase and aspirin).

Merits: the strongest single-study evidence for cause. Limits: cost, ethical limits, artificial settings, strict eligibility (Campbell and Stanley, 1963).

Definition: an observational study in which exposure and outcome are measured at the same time in a defined population. Its main result is prevalence.

1. State the objective and define the population.
2. Calculate the sample size (single-proportion formula) and choose a sampling method.
3. Define the variables and prepare a validated tool.
4. Collect data at one contact per person.
5. Analyse: prevalence with 95% confidence interval, and associations.

Advantages: quick, cheap, no follow-up, several exposures and outcomes at once, useful for planning services. Limitations: time order unknown, long-lasting cases over-represented, unsuited to rare or short illnesses, non-response. Examples: prevalence of diabetic retinopathy in a diabetes clinic, anaemia survey among schoolchildren.

Repeated cross-sectional surveys, such as successive NFHS rounds, show trends in the population but not change within individuals. A difference between age groups in one survey may be a birth-cohort effect rather than an effect of age.

• Selection bias: non-response, volunteers, a sampling frame that misses some groups.
• Prevalence-incidence bias: people who died or recovered quickly are missed, so long-lasting cases dominate.
• Information bias: recall bias, interviewer bias, measurement error, socially desirable answers.
• Reverse causation: the outcome may have changed the exposure (people with diabetes change their diet after diagnosis).
• Confounding. Details are in the bias and validity chapter.

1. Describe: frequencies, means or medians.
2. Prevalence with 95% confidence interval, allowing for the design effect in cluster samples.
3. Compare groups: chi-square test for proportions, t-test or ANOVA for means, non-parametric tests for skewed data.
4. Measures of association: prevalence ratio and prevalence odds ratio.
5. Multivariable models: logistic regression (odds ratio), Poisson or log-binomial regression (prevalence ratio), linear regression for continuous outcomes.
6. Correlation between two continuous variables.
7. Complex surveys: sampling weights and design-based standard errors. See the choosing a test chapter.

Definition: an observational analytical study that starts with people who have the disease (cases) and people who do not (controls), and compares their past exposure.

1. State the hypothesis about the exposure and the disease.
2. Define cases with clear diagnostic criteria. Prefer new (incident) cases.
3. Choose the source of cases: hospital or population register.
4. Select controls from the population that produced the cases (hospital, neighbourhood or population controls). Consider matching on age and sex, and more than one control per case. Match by pairs or by frequency, and only on confounders (matching on a factor caused by the exposure is overmatching). Efficiency with k controls per case is about k ÷ (k + 1), so more than four controls adds little.
5. Measure exposure the same way in cases and controls (records, structured interview, laboratory tests), with interviewers unaware of case status where possible.
6. Analyse: 2 × 2 table, odds ratio with 95% confidence interval, conditional logistic regression for matched designs, logistic regression for confounders. For pair-matched data the odds ratio is the ratio of the two kinds of discordant pairs, tested with McNemar's test.
7. Interpret: consider chance, bias and confounding before cause.

Advantages: quick and cheap, efficient for rare diseases and long latency, several exposures can be studied, few participants, no follow-up. Disadvantages: recall and selection bias, incidence and relative risk cannot be calculated directly, time order may be uncertain, inefficient for rare exposures, one outcome at a time.

Definitions and control of bias are covered in the bias and validity chapter.

Analytical epidemiology tests hypotheses about the determinants of disease by comparing groups. Its main designs are case-control and cohort studies. Cohort study: people free of the outcome are classified by exposure and followed to compare the incidence of the outcome.

1. Selection of the cohort: the general population or a special group (an occupation, doctors, pregnant women), all free of the outcome and able to be followed. The cohort may be closed (membership fixed at the start), open (people join during follow-up), or an inception cohort (all enrolled at the same early point of the disease, such as diagnosis).
2. Exposure data: measured at the start, with level and duration, and repeated if exposure changes.
3. Comparison group: internal (unexposed members of the same cohort), external (another cohort), or general population rates. When nobody is unexposed, compare levels of exposure.
4. Follow-up: regular examinations, records and death registers, with methods fixed in advance.
5. Outcome ascertainment: the same criteria for exposed and unexposed people, preferably by assessors unaware of exposure.
6. Analysis: incidence in each group, relative risk, attributable risk, population attributable risk, survival analysis for time to event.

Advantages: clear time order, incidence and relative risk, several outcomes of one exposure, suitable for rare exposures, less recall bias. Disadvantages: large numbers and long follow-up, high cost, loss to follow-up, unsuitable for rare diseases, exposure and diagnostic criteria can change during the study, results are delayed.

Methods to overcome them: retrospective or nested designs using existing records, a population that is easy to trace, record linkage with registers, regular contact and reminders, definitions fixed at the start, and a sample size that allows for expected losses. Biases: selection (healthy volunteers, healthy worker effect), loss to follow-up that differs by exposure, outcomes looked for more carefully in the exposed, misclassification of exposure, change in exposure status during follow-up, the Hawthorne effect, unplanned (post hoc) analyses reported as findings, and confounding.

Retrospective (historical) cohort: the investigator uses records made in the past to identify exposed and unexposed people, then traces their outcomes up to the present. The direction is still exposure to outcome. Only the timing lies in the past.

1. Requirements: reliable past records of exposure (employment, hospital or pharmacy records) and a way to trace outcomes (death registers, follow-up records).
2. Teaching example: factory employment records from 20 years ago identify workers exposed to a dust, and death records show which of them died of lung disease.
3. Advantages: faster and cheaper than a prospective cohort, suits long latency, results available soon.
4. Disadvantages: records may be incomplete, lack information on confounders, or use old diagnostic criteria.
5. Ambispective (ambidirectional) cohort: exposure from past records, with follow-up continuing into the future. It is quicker than a prospective cohort and still allows new outcome data to be collected.

Nested case-control study: a case-control study carried out inside a defined cohort.

1. Assemble a cohort and store baseline samples or data.
2. Follow the cohort and identify each case as it arises.
3. For each case, select controls from cohort members still free of the disease at that time, usually matched on age, sex and follow-up time.
4. Measure the exposure (for example a biomarker in stored serum) only in cases and controls.
5. Analyse with conditional logistic regression.

Advantages: exposure measured before disease (no recall bias), cases and controls from the same population, much cheaper than measuring the whole cohort. Limitations: needs an existing cohort, and fewer participants give less precision than the full cohort.

See the systematic review and meta-analysis chapter and the clinical trials chapter.

Definition: a randomised trial in which each participant receives two or more treatments in sequence, in a random order, separated by a washout period.

1. AB/BA design: one group receives A then B, the other B then A.
2. Washout: a gap long enough for the effect of the first treatment to disappear.
3. Carry-over effect: the first treatment still acts during the second period.
4. Period effect: the condition changes between periods whatever the treatment.
5. Analysis: within-patient (paired) comparison, with checks for period and carry-over effects.

Suitable for: stable chronic conditions (asthma, hypertension, chronic pain), treatments with rapid onset and short-lasting effect, and bioequivalence studies. Advantages: each patient is their own control, variation between patients is removed, fewer patients are needed. Disadvantages: carry-over, longer participation for each patient, a dropout loses data from both periods, unsuitable for curable or progressive diseases and for outcomes such as death.

1. By phase: phase I (safety and dose), phase II (activity and further safety), phase III (comparison with standard treatment), phase IV (after marketing).
2. By design: parallel group, cross-over, factorial (two or more interventions in all combinations, as in ISIS-2), cluster, adaptive.
3. By framework: superiority, equivalence, non-inferiority.
4. By purpose: treatment, prevention, screening, supportive care, diagnosis.
5. By control: placebo, active treatment, historical control (not randomised).
6. By blinding: open, single-blind, double-blind.

Why oncology needs randomised trials: differences between treatments are often modest, so bias in observational comparisons can be as large as the effect. In routine practice, fitter patients tend to receive more intensive treatment (confounding by indication). Randomisation balances prognostic factors such as stage and performance status, and survival outcomes need concurrent controls.

Impact: randomised trials provide the evidence for standard protocols and guidelines, test whether less intensive treatment is acceptable (non-inferiority trials), and measure the survival benefit and toxicity of new drugs. Challenges: cost, slow recruitment, strict eligibility, and ethical questions with placebo controls. See the clinical trials chapter.

1. Phenomenology: the lived experience of a phenomenon (living with HIV).
2. Grounded theory: builds a theory from the data (how patients decide to stop treatment).
3. Ethnography: the culture and practices of a group, by prolonged observation (care practices in a labour ward).
4. Case study: detailed study of one case or setting (a district programme).
5. Narrative research: individual accounts told in sequence.
6. Others: descriptive qualitative studies, participatory action research.

Tools: in-depth interviews, focus group discussions, key informant interviews, participant and non-participant observation, document review, and participatory methods such as free listing. Software: NVivo, ATLAS.ti and MAXQDA (commercial), QDA Miner Lite and Taguette (free). Software organises and codes the data. The researcher still interprets it. See the qualitative research chapter.

1. Research question: prevalence needs a cross-sectional design, a question about cause needs an analytical or experimental design.
2. Frequency of the disease: a rare disease favours a case-control study.
3. Frequency of the exposure: a rare exposure favours a cohort study.
4. Latency: a long latency favours a case-control or retrospective cohort study.
5. Whether the exposure can ethically be assigned: if it can, a trial is possible.
6. Need for incidence and time order: cohort study.
7. Existing data: registers, records or stored samples allow retrospective cohort or nested designs.
8. Resources: time, money, staff, laboratory facilities.
9. Access to participants and the likelihood of complete follow-up.
10. Ethics, and the strength of evidence needed for the decision the study will inform.

A good proposal has a clear question, specific objectives, a design that matches the objectives, a justified sample size, defined variables, a statistical analysis plan, ethics, a timeline and a budget. The full list of components is in the outline on descriptive and causal studies above.

Design: a community-based analytical cross-sectional study, with exposure measured around each patient's home. An ecological comparison of wards is a quicker first step, but it cannot show the link in individuals.

1. Objective: to assess the association between the number of fast-food outlets near home and diabetes control.
2. Participants: adults with known diabetes, chosen by probability sampling from health centre registers.
3. Exposure: number of fast-food outlets within a fixed distance of home, mapped with GPS, plus each person's own fast-food intake.
4. Outcome: HbA1c, classified as controlled or not by a cut-off stated in the protocol.
5. Confounders: age, duration of diabetes, treatment, adherence, physical activity, income.
6. Analysis: prevalence of poor control, chi-square test, multilevel logistic regression (people within neighbourhoods).

Advantages: feasible within a thesis period, one contact per person, uses individual data. Limitations: time order unknown (people may change diet after diagnosis), outlet numbers may reflect income and other neighbourhood factors, and diet recall is inaccurate. A cohort study would be needed to show that the exposure comes before worsening control.

Design: a hospital-based analytical cross-sectional study. Reasons: each patient is seen once, no follow-up is needed, recruitment and analysis fit within a year, and it gives both prevalence and associated factors.

1. Objective (example): to estimate the prevalence of uncontrolled hypertension, and its associated factors, among adults on treatment attending a family medicine clinic.
2. Methodology: setting and study period, inclusion and exclusion criteria, sample size by the single-proportion formula, consecutive sampling, a structured proforma, blood pressure measured by a standard method, definitions fixed in advance, ethics approval and consent.
3. Analysis: prevalence with 95% confidence interval, chi-square test for associations, logistic regression for independent factors (adjusted odds ratios). See the sample size chapter.

Design: a cluster randomised controlled trial with schools as clusters. If schools cannot be randomised, a quasi-experimental nonequivalent control group design, measuring both groups before and after.

1. Title, background and rationale: burden of adolescent obesity and gaps in existing evidence.
2. Objective and hypothesis: the programme reduces BMI-for-age z-score at 12 months compared with usual school activities.
3. Setting and participants: urban schools, students in a stated age range, parental consent and student assent.
4. Randomisation: schools allocated at random to intervention or control, with allocation concealed until schools are enrolled.
5. Intervention: nutrition education, daily physical activity, changes to the school canteen, sessions for parents. Control: usual curriculum, with the programme offered after the study.
6. Outcomes: primary, change in BMI-for-age z-score. Secondary: waist circumference, diet and activity scores, screen time.
7. Sample size: calculated for the primary outcome, multiplied by the design effect for clustering, and increased for dropouts.
8. Data collection: at baseline, 6 and 12 months, by trained staff unaware of allocation where possible.
9. Analysis: intention-to-treat, mixed-effects models that allow for clustering.
10. Ethics: ethics committee approval, consent and assent, avoiding stigma for children with obesity, confidentiality, registration with CTRI.
11. Timeline, budget and dissemination. Report with CONSORT 2025, or TREND if not randomised.

Design: mixed methods. An analytical cross-sectional survey measures how many patients do not take treatment as prescribed and the associated factors. In-depth interviews with a purposive sub-sample then explain the reasons.

1. Objectives: estimate the prevalence of non-compliance, identify associated factors, and explore patients' reasons.
2. Participants: adults with type 2 diabetes, or people living with HIV on antiretroviral therapy, attending a clinic for at least a stated period.
3. Sampling and size: systematic random sampling from the clinic register, sample size by the single-proportion formula. Interviews continue until they add no new themes.
4. Measurement: a validated adherence questionnaire, pill counts or pharmacy refill records, with HbA1c or viral load as supporting measures. Factors: cost, side effects, beliefs, depression, number of pills, distance, family support.
5. Statistical analysis plan: descriptive statistics, prevalence with 95% confidence interval, chi-square tests, multivariable logistic regression with adjusted odds ratios. Interviews transcribed, coded and analysed thematically.
6. Ethical issues (HIV): ethics committee approval, written informed consent, confidentiality of HIV status, interviews in a private room, no names on data forms, coded and secured data, freedom to refuse without any effect on care, and referral for counselling when needed.

Dengue seroprevalence in a city: a community-based cross-sectional serosurvey.

1. Objective: estimate the proportion of residents with antibodies to dengue virus, by age and ward, and the associated factors.
2. Sampling: multistage cluster sampling: wards, then census blocks, then households, then individuals across age groups.
3. Sample size: single-proportion formula using the expected seroprevalence and precision, multiplied by the design effect, increased for non-response.
4. Laboratory: a dengue IgG antibody test of known sensitivity and specificity, with samples transported and stored to a standard protocol.
5. Analysis: seroprevalence with 95% confidence interval adjusted for the survey design (and for test sensitivity and specificity where appropriate), logistic regression for associated factors.
6. Ethics: consent and assent for children, a small blood volume, return of results, referral.

Long COVID in a state: a cohort design is preferred. Adults with a confirmed COVID-19 infection are identified from state records, with an uninfected comparison group, and both are followed for persistent symptoms using a published case definition stated in the protocol. If time is short, a cross-sectional survey of people with a past confirmed infection, chosen by multistage sampling, estimates prevalence, but it cannot give incidence and it has recall bias. The proposal also covers objectives, sample size, validated symptom and function questionnaires, analysis, ethics, timeline and budget.