The influence of referral patterns on the characteristics of diagnostic tests

Abstract

The discrimination of a diagnostic test--characterized by sensitivity, specificity, likelihood ratio and ROC curve--may be influenced by referral patterns of general practitioners. Symptoms and test results in particular will affect the probability of referral, while the degree of development of the pathological process directly influences the probability of positive test results. Using numerical examples, we analyse and discuss a few specific situations: (1) referral depends only on symptoms; (2) referral depends both on symptoms and on test results; (3) referral depends only on test results. In the first situation, test characteristics and predictive values are invariant over the strata of symptomatology, while in the third situation the predictive values are unchanged. If there is a positive relationship between positive test results and referral probability, overall sensitivity will increase while specificity and likelihood ratio will decrease. A general representation is given for the evaluation of the direction of change of the likelihood ratio as a function of referral probabilities. The shape of receiver-operating characteristic curves is less sensitive to bias, but at the level of specific cut-off points considerable changes may occur.