Cohort Study — Longitudinal Observational Design

In a prospective cohort study, investigators recruit participants and then follow them forward in time, collecting exposure and outcome data as events occur. In a retrospective (historical) cohort study, both exposure and outcome have already happened; the investigator reconstructs the cohort from existing records. The analytic logic is identical, but retrospective designs are faster and cheaper while being more susceptible to information bias from incomplete historical records.

Sample size depends on the expected incidence of the outcome, the size of the effect you wish to detect, the duration of follow-up, and the desired statistical power. For common outcomes a few hundred participants may suffice; for rare outcomes with modest effect sizes, tens of thousands of participants followed for decades may be necessary. Power calculations should be performed before recruitment begins.

In a randomised trial the investigator controls exposure assignment, which eliminates confounding by design and allows causal conclusions. In a cohort study exposure is determined by natural circumstances, so unmeasured confounders may distort the exposure-outcome association. Cohort studies are observational and therefore occupy a lower position in the hierarchy of evidence for causal questions, but they are feasible when randomisation would be unethical or impractical.

The primary effect measure is the relative risk (risk ratio) when cumulative incidence is used, or the incidence rate ratio when person-time denominators are used. Cohort studies can also estimate the risk difference (attributable risk), which is directly interpretable for public health impact. Odds ratios from logistic regression are sometimes reported for convenience but approximate the risk ratio only when the outcome is rare.

First, minimise loss through active retention strategies. Report the proportion lost and compare baseline characteristics of completers versus non-completers. Analyse using intention-to-follow-up principles, and conduct sensitivity analyses (e.g., best-case and worst-case scenario imputations) to assess the robustness of findings. Inverse probability weighting can correct for informative censoring when predictors of dropout are measured.