Study Designs

Observational vs experimental

The dividing line. What separates the two families is whether the investigator controls the exposure. Observational studies (case-control, cohort, cross-sectional, ecological) watch exposures and outcomes without intervening. Experimental studies, chiefly randomized controlled trials (RCTs), assign an intervention.

Naming the design from a stem. Ask two questions. Did the investigator assign the exposure? If not, where did the study start and in which direction did it look?

Observational designs

Case-control. The study starts with diseased cases and disease-free controls and looks back at prior exposures. It is retrospective and reports an odds ratio (OR = ad/bc from the 2 × 2 table). It cannot measure incidence because the investigator chose how many cases and controls to include. It is efficient for rare diseases and can examine many exposures for one outcome. Its weak points are recall bias and selection bias in choosing controls. When the disease is rare, the odds ratio approximates the relative risk.

Cohort. The study follows exposed and unexposed groups forward and counts new cases, so it measures incidence directly and reports a relative risk. Cohorts are usually prospective. A retrospective (historical) cohort uses past records to define exposure and then traces outcomes. They suit rare exposures and can track several outcomes of one exposure. They are slow and costly for rare diseases and vulnerable to loss to follow-up.

Cross-sectional. This design measures exposure and disease at a single point in time and yields prevalence. Because both are assessed at once, it cannot establish temporality, and therefore not causation. It is good for estimating disease burden and generating hypotheses.

Ecological. This design correlates exposure and outcome using population-level aggregate data rather than individuals. Its main pitfall is the ecological fallacy: assuming a group-level association holds for individuals. An example is concluding that chocolate boosts individual achievement because countries that eat more chocolate have more Nobel laureates.

Case reports and case series. These describe one or a few patients without a comparison group. They can flag a new disease or adverse effect and generate hypotheses, but they cannot test associations.

Experimental designs

Randomized controlled trial. Subjects are assigned at random to treatment or to control (placebo or standard care), ideally with blinding. Randomization balances known and unknown confounders, which gives the RCT the highest internal validity of any primary study and makes it the gold standard. Double-blinding keeps both subjects and investigators unaware of assignments, which limits bias in reporting and assessing outcomes. Intention-to-treat analysis keeps everyone in the group they were randomized to, even if they stop treatment, and so preserves the benefit of randomization.

Crossover trial. Each subject receives both the treatment and the control in sequence, separated by a washout period, so every patient serves as their own control. This removes between-subject variability and needs fewer participants. The order of treatments and any carryover effects must be controlled. The design suits stable, chronic conditions and treatments with short-lived effects.

The washout period stops the first treatment from carrying over into the second period.

Measuring disease frequency

Prevalence ≈ incidence × duration (P = I × D). Chronic, long-lasting diseases such as diabetes accumulate cases, so their prevalence is much higher than their incidence. A treatment that improves survival without curing raises prevalence while incidence is unchanged. A rapid cure or a quick death lowers prevalence. Preventing new cases lowers incidence first, and prevalence falls later.

Evidence hierarchy, meta-analysis and trial phases

Trade-offs. Higher levels are less prone to bias but harder, slower and costlier to conduct.

Meta-analysis. A meta-analysis statistically combines the results of several studies, which increases power and precision. It is limited by heterogeneity between studies and by publication bias. A funnel plot helps detect publication bias, and asymmetry suggests that small negative studies are missing.

Garbage in, garbage out: pooling biased trials produces an estimate that is precise but wrong.