Statistical methods in epidemiology. VIII. On the use of likelihood ratios for diagnostic testing with an application to general practice.

Abstract

An array of statistical methods is available for 2 x 2 contingency table analysis. These include odds ratios, reliability coefficients and the Chi-squared test. This paper continues the '2 x 2' theme but extends it to look at diagnostic testing. The setting is primary care, the disease is urological cancer. Several statistical indices are discussed to measure diagnostic efficiency; these include sensitivity, specificity and the likelihood ratio. The latter can be converted into a flexible 'weight of evidence', which is equivalent to the naive Bayes' discriminant. The utility of the naive Bayes' discriminant is illustrated in a primary care setting. The naive Bayes' discriminant has proved to be the most durable of statistical discriminants in medical diagnosis. The reason for its durability is ease of use. Despite a concern about independence, it remains a popular choice as a diagnostic aid.