Retrospective Phase III Clinical Trial

Rarely, and only under specific circumstances. Regulatory agencies such as the FDA and EMA generally require randomized, controlled prospective trials for pivotal approval. However, for rare diseases, pediatric indications, or conditions where randomization is impossible, well-conducted retrospective studies using external control arms or registry data can contribute as supportive evidence within a broader submission package. The FDA's Real-World Evidence program provides a framework for evaluating such submissions case by case.

Randomization eliminates both measured and unmeasured confounding by ensuring that, on average, all patient characteristics are balanced across treatment arms. Propensity score matching balances only the measured covariates included in the propensity model; unmeasured confounders remain unbalanced. This is why even perfect propensity score matching cannot fully substitute for randomization, and why sensitivity analyses for residual confounding are mandatory in retrospective studies.

Immortal time bias occurs when a period during which patients could not have experienced the outcome is incorrectly assigned to one treatment group. For example, if patients are classified as 'treated' from the date they first received a drug, but the index date is set at a prior hospital admission, the intervening period — during which they survived long enough to reach treatment — inflates the apparent survival benefit. It is a particularly dangerous artifact in retrospective analyses of administrative data.

Choose a retrospective design only when a prospective randomized trial is genuinely infeasible — due to ethical constraints, rarity of the disease, cost, or urgency — and when high-quality historical data with complete covariate and outcome ascertainment are available. If randomization is achievable, the prospective RCT remains the gold standard and should be preferred. Retrospective designs are best positioned as hypothesis-confirming tools that complement, not replace, the prospective evidence base.

Sample size calculations follow the same principles as prospective Phase III trials: the study must be powered to detect the minimum clinically important difference in the primary endpoint at acceptable alpha and beta levels. In practice, the available historical sample is often fixed, so the analysis begins with a power calculation to determine whether the existing dataset is adequate. If power is insufficient, the study should not be interpreted as Phase III-level evidence — a larger data source or a different design should be sought.