"TAGS", a program for the evaluation of test accuracy in the absence of a gold standard

When a perfect reference test (i.e. "gold standard") is not available, it is possible to obtain estimates of test sensitivity and specificity using "latent-class" methods. However, there are few widely available software programs that allow implementation of these procedures. We describe the development of a program (implemented in R and S-Plus software) for this purpose that yields maximum-likelihood estimates of sensitivity, specificity and prevalence. We also have implemented an HTML form, which submits data to a web-based interface to R. The programs can incorporate data obtained from several populations, results of multiple tests, and can account for data obtained from a reference population in which the true status (infected or non-infected) of each individual is known exactly. Two estimation methods are used: a Newton-Raphson procedure and an expectation-maximisation (EM) procedure. The estimation methods assume test independence conditional on the infection status of the individuals and constant test accuracy in each population. A goodness-of-fit statistic and the residuals of pairwise correlation coefficients are calculated to check the validity of these assumptions. Two examples are used to illustrate application and limitations of the programs. The programs are available at www.afssa.fr/interne/tags.htm (Europe) or www.epi.ucdavis.edu/diagnostictests/ (USA).