Spectrum bias: a quantitative and graphical analysis of the variability of medical diagnostic test performance

The performance of a medical diagnostic test may vary in a subgroup of patients according to the severity and clinical presentation of the disease. This phenomenon is often referred to as the spectrum effect. There is currently no technique to determine in which situations this spectrum effect may lead to a spectrum bias, that is, a distortion of the posterior probability which can potentially affect the clinical decision. We show that spectrum bias, on either a positive or a negative test result, can be expressed as the subgroup-specific likelihood ratio (LR) divided by the LR in the overall population of patients. This assessment of spectrum bias is independent of the pretest probability. We present here the test statistic and its variance and also propose a mode of graphically visualizing the presence of spectrum bias. We applied it to 15 examples from the literature, of which three are discussed in detail. In two examples, there was spectrum bias: (1) diagnosis of urinary tract infection in subgroups of low or high risk patients when using the dipstick test, and (2) diagnosis of left ventricular hypertrophy using electrocardiography in subjects with different body mass index. In a third example, there was no spectrum bias for the diagnosis of coronary heart disease in subgroups of patients defined by age, sex or comorbidities when using exercise electrocardiography. As sensitivity and specificity usually vary in opposite directions from one subgroup to the other, the ratio of LRs tends to remain constant, resulting in no or little spectrum bias. We conclude that spectrum effect is more common than spectrum bias and that clinical decisions are affected only under certain conditions.