How to deal with dichotomous tests? Application of a rankit ordinal scale model with examples from the Nordic ordinal scale project on screening tests.
The purpose of this investigation was to improve a rankit ordinal model for evaluating and validating dichotomized tests in a prospective Nordic project. The model is based on the assumption that the S-shaped curve of fractions of positive for increasing concentrations can be de-convoluted to a histogram and thereby used to calculate the parameters for a ln-Gaussian distribution. In a Nordic survey, four urine samples with known concentrations of hCG (human chorionic gonadotrophin) and nitrites were distributed to more than 2500 practitioners' offices. The results are presented as parameters (geometric mean and CV) for the components urine-hCG and urine-nitrites, together with fractions of positive for clinical critical values (5 and 40 IU/L for hCG), for which fractions should be below 0.01 and above 0.99, respectively, and 7 micromol/L for nitrites. Furthermore, the concentration intervals of varying fractions of positive from 0.01 to 0.99 are estimated as grey zones. The parameters and grey zones for different kits are compared. No urine-hCG kit fulfilled the low clinical criterion, whereas all fulfilled the high criterion. Seven of the eight nitrites kits had fractions of positive above 0.9 for the company confirmation limit, but varying fractions for the clinically important limit of 7 micromol/L (fractions from 0.06 to 0.83). The present model makes it easy to estimate parameters for the kits, and also to estimate the fractions of measured positives for specified concentrations. It is thus suited for external quality assessment as well as for manufacturers' method validation.