Dr. Evans raises an important point regarding the potential of reference test bias to masquerade as spectrum effect. Apparent spectrum effect may be observed under three conditions: 1) Test performance may actually differ across subgroups; 2) test performance may appear to differ across subgroups because of reference test bias and variation in prevalence; and 3) test performance may actually differ across subgroups but the observed effect may be exaggerated (or potentially reduced) by reference test bias and variation in prevalence. The magnitude of the bias in the last two circumstances depends on the prevalence in the subgroups, the performance of the new test and the reference test in the subgroups, and the presence of conditional dependence between the two tests. If the reference test is imperfect, spectrum effects may be present for both the new test and the reference test.

In our clinical example, we did not attempt to account for reference test bias. Using ligase chain reaction assay as our reference standard for chlamydial infection, we found that the measured prevalences in the sexually transmitted disease and family planning clinic settings were 11.0% and 7.8%, respectively. By age group, the measured prevalence was 12.5% among women 24 years of age or younger and 3.3% among women older than 24 years of age. Ligase chain reaction assay is not a perfect reference test; its estimated sensitivity is approximately 90%, and its estimated specificity is approximately 99.0% (1).

For the clinician, one key to discriminating spectrum effect from reference test bias is the biological plausibility of the reported spectrum effect. For example, in many infectious diseases, asymptomatic infections may have lower organism burden and, consequently, tests may have lower sensitivity. For the clinical researcher, the potential magnitude of the reference test bias could be addressed in sensitivity analyses by using algebraic adjustment for the reference test bias under different conditions. Alternatively, methods such as latent class analysis that do not require a reference standard may be used to assess test performance in subgroups.

Dr. Scobbo's point that correct diagnosis is an essential precursor to correct treatment is also well taken. The precise diagnosis of a disease should enhance treatment efficacy, although in some circumstances syndromic management is sufficient to treat the condition effectively. However, we concur with Dr. Scobbo's conclusion that applying a rigid hierarchy that universally prioritizes either treatment or diagnosis would subvert both the art and the science of medicine.