Comparing different statistical methods for evaluating diagnostic effectiveness of clinical tests: respiratory distress syndrome as a model.

Abstract

Phospholipids in specimens of amniotic fluid from 346 patients were quantified and the results evaluated in light of the clinical outcome. Fifty-eight neonates had respiratory distress syndrome. We used this data base to compare different statistical methods for evaluating test effectiveness and diagnostic discrimination. Dichotomizing quantitative tests into binary tests with arbitrary cutoff values was inadequate for comparing test effectiveness. Subgrouping the data into deciles and calculating the incidence of respiratory distress syndrome for each decile avoided the problems of the preceding approach and was easy to calculate and comprehend; however, this method lacked statistical power. Relative operating characteristic curves yielded more statistical power, but results were more difficult to calculate and were not intuitively obvious to most workers in the laboratory. A modified cumulative frequency plot, combining elements of both decile subgrouping and relative operating characteristic curves, was easily calculated and intuitively obvious. These plots, like relative operating characteristic curves, provided an index for quantifying test effectiveness. When used in combination with standard cumulative frequency curves, they also provided direct diagnostic information on disease probability for any value of the clinical assay.