Spectrum bias

Spectrum bias occurs when the case-mix of patients in a diagnostic accuracy study does not reflect the clinical population in which the test will be used.¹ This includes differences in disease severity, stage, comorbidity, age, sex, and other clinical characteristics. Because test performance often varies across such subgroups, the overall sensitivity and specificity computed from a skewed sample may not generalize.² For example, a test may appear highly sensitive in a study with mostly severe cases but perform poorly in mild or early-stage disease, where the test is often needed most.³

The primary consequence is a biased estimate of sensitivity and specificity, which can mislead clinical decision-making and health policy.⁴ Overestimation of sensitivity may lead to missed diagnoses when the test is applied in a broader population, while overestimation of specificity can result in unnecessary confirmatory testing or treatment. Spectrum bias also affects predictive values, which depend on disease prevalence and case-mix, further complicating interpretation.² In systematic reviews, ignoring spectrum bias can produce misleading pooled estimates of test accuracy.⁵

Researchers can detect spectrum bias by comparing test performance across clinically relevant subgroups, such as disease stage or age, and by reporting results stratified by these factors.³ To mitigate it, studies should enroll a consecutive or random sample of patients representing the full spectrum of disease and non-disease encountered in practice.¹ Statistical methods, such as adjustment for case-mix or use of hierarchical models, can also help.⁵ Reporting guidelines like STARD (Standards for Reporting of Diagnostic Accuracy Studies) encourage transparent description of study populations to allow assessment of spectrum bias.⁶

Spectrum bias is closely related to but distinct from selection bias, which arises from how patients are chosen for the study, whereas spectrum bias concerns the distribution of characteristics within the chosen sample.² A classic example is the evaluation of exercise stress testing for coronary artery disease: early studies in high-prevalence, severe-disease populations reported high sensitivity, but later studies in community settings with milder disease showed much lower sensitivity. Another subtlety is that spectrum bias can affect not only sensitivity and specificity but also likelihood ratios, which are often assumed to be constant across disease severity—an assumption that may fail. In the era of machine learning, spectrum bias can arise from training algorithms on datasets that do not reflect the target population, leading to poor real-world performance.⁵

Sensitivity

The proportion of true positives correctly identified by the test.

Specificity

The proportion of true negatives correctly identified by the test.

Predictive values

The probability that a positive or negative test result is correct, depending on prevalence.

Case-mix

The distribution of patient characteristics in a study population.