Retrospective Diagnostic Accuracy Study

In a prospective design, participants are enrolled before the index test is applied, both the index test and reference standard are planned in advance, and the sequence of testing is controlled. In a retrospective design, both test results and reference standard outcomes come from data that already exist. The prospective design offers better control over spectrum of disease, standardization of test conditions, and verification completeness, but is slower and more expensive. The retrospective design is faster and cheaper but requires careful attention to spectrum bias and verification bias.

Missing reference standard results must be reported transparently, not silently excluded. Describe how many participants are missing, why the reference standard was not applied, and whether the missingness is likely to be related to the index test result or disease status. Where feasible, perform a sensitivity analysis (e.g., assuming all missing cases are true positives or all true negatives) to bound the possible impact on accuracy estimates. A high proportion of missing reference standard results is a serious threat to validity.

QUADAS-2 is the standard tool. It assesses four domains — patient selection, index test, reference standard, and flow and timing — and separately rates risk of bias and applicability concerns for each. For retrospective studies, particular scrutiny is warranted in the patient selection domain (was the sample representative of the intended clinical population?) and the flow and timing domain (did all patients receive the reference standard, and was it applied independently?).

A case-control selection — deliberately sampling confirmed disease cases and known-negative controls from a registry — is sometimes used for feasibility but inflates estimates of sensitivity and specificity relative to a consecutive or random sample. If you use this approach, report disease prevalence in the source population and apply predictive value calculations with caution; the artificially extreme spectrum means the estimates may not translate to routine clinical practice.

Sample size should be calculated based on the expected sensitivity and specificity, the desired precision of confidence intervals, and the disease prevalence in the accessible archive. A common approach targets a 95% confidence interval half-width of no more than 0.05–0.10 for the primary accuracy measure. For rare diseases, a minimum of 30–50 cases is often cited as a lower bound to avoid very wide intervals, though larger samples are strongly preferred. Report the sample size justification explicitly in the methods.