Study Designs in Clinical Research

• Cross-sectional study: exposure and outcome measured at one point in time. Yields prevalence; cannot establish temporality, so it generates hypotheses rather than causal claims.
• Case–control study: starts with the outcome and looks backwards for exposure. Efficient for rare diseases and long latency; yields an odds ratio, and cannot yield incidence or relative risk. Vulnerable to recall and selection bias.
• Cohort study: starts with exposure and follows forward for outcome. Yields incidence and relative risk; good for rare exposures; prospective cohorts are costly and vulnerable to loss to follow-up.
• Randomised controlled trial: randomisation balances known and unknown confounders, which no observational design can do. Blinding limits observer and reporting bias. Limited by cost, generalisability and ethics.
• Crossover trial: each participant serves as their own control; watch for carryover effects.
• Meta-analysis pools studies for precision but inherits their flaws and is subject to publication bias.
• Case series and case reports describe without a comparison group and cannot measure association.

Frequency measures

• Prevalence: existing cases ÷ total population at one moment. The natural output of a cross-sectional study; inflated by anything that prolongs disease duration (better treatment that averts death but not disease).
• Incidence: new cases ÷ population at risk per unit time. Requires follow-up, so only cohorts and trials produce it. Conceptually, prevalence ≈ incidence × average disease duration.

Measures of association (2×2 table with a = exposed cases, b = exposed non-cases, c = unexposed cases, d = unexposed non-cases)

• Relative risk (RR) = [a/(a+b)] ÷ [c/(c+d)]. Interpretable only when the denominators are real populations — i.e., cohort or RCT. RR = 1 means no association.
• Odds ratio (OR) = ad/bc. In a case–control study the investigator fixes how many cases and controls to enrol, so risk denominators are artefacts of sampling; the exposure odds ratio survives that sampling and is the only valid effect measure. OR approximates RR when the outcome is uncommon (rule of thumb, roughly under 10%); with a common outcome the OR exaggerates the RR away from 1.
• Attributable risk / risk difference = risk in exposed − risk in unexposed. Absolute, not relative — the basis of ARR, NNT = 1/ARR, and NNH = 1/attributable risk increase.
• Attributable risk percent = (RR − 1)/RR: the fraction of disease among the exposed attributable to the exposure.
• Hazard ratio: ratio of instantaneous event rates from time-to-event (Cox) analysis; handles censoring and unequal follow-up, so it is the standard trial and cohort output.

Analysis and reporting conventions

• Intention-to-treat: analyse participants in the arm to which they were randomised regardless of adherence — preserves randomisation and gives the less biased, usually more conservative estimate; per-protocol analysis reintroduces confounding. Codified in the CONSORT statement for trials.
• Reporting frameworks: STROBE for observational studies, CONSORT for randomised trials, PRISMA 2020 for systematic reviews and meta-analyses, GRADE for rating certainty of evidence; ICMJE requires prospective trial registration (e.g., ClinicalTrials.gov) before enrolment.

Choosing the design

• Rare disease or long latency (mesothelioma, Creutzfeldt–Jakob): case–control.
• Rare or unusual exposure (an occupational solvent, a single drug lot): cohort.
• Burden of disease at a moment, or screening-program planning: cross-sectional, and note that USPSTF recommendations rest on the trial and cohort evidence such surveys cannot supply.
• Definitive efficacy claim: randomised trial, analysed by intention-to-treat per CONSORT, with prospective registration per ICMJE.
• Existing cohort with banked specimens: nested case–control — assays are done only on sampled controls, but exposure still precedes outcome.

• Sampling frame names the design: enrolled by outcome = case–control (odds ratio only); enrolled by exposure = cohort (incidence, RR); everything measured at once = cross-sectional (prevalence). This single question resolves most stems.
• Case–control studies cannot yield incidence, relative risk, or attributable risk — the most frequently tested distractor on the exam.
• OR ≈ RR only when the outcome is rare. With a common outcome the odds ratio overstates the effect; a stem describing a 40% event rate and reporting an OR is signalling exactly this.
• A better treatment that prolongs survival raises prevalence while incidence is unchanged (prevalence ≈ incidence × duration). Do not read rising prevalence as rising risk.
• NNT = 1/ARR, always rounded up, and relative risk reduction can look impressive while the ARR — and thus clinical value — is trivial.
• Intention-to-treat preserves randomisation and is the CONSORT-endorsed primary analysis; per-protocol analysis reintroduces the confounding randomisation was meant to erase. "Analyse only adherent patients" is never the best answer for a primary efficacy endpoint.
• Bias by design: recall bias and Berkson bias (hospital controls) belong to case–control; loss to follow-up and healthy worker effect to cohorts; Neyman (prevalence–incidence) bias to cross-sectional studies, which miss rapidly fatal or rapidly resolving disease; publication bias, detected as funnel-plot asymmetry, to meta-analysis.
• Use medical records rather than interviews, and blind the interviewer, to blunt recall bias.
• Matching cases and controls on a variable eliminates it as a confounder but also makes it unstudiable, and requires matched analysis — a classic single-best-answer trap.
• Only randomisation balances unknown confounders; stratification, matching, restriction, and multivariable regression handle measured confounders only.