Comparison of nonparametric recursive partitioning to parametric discriminant analyses in laboratory differentiation of hypercalcemia

Abstract

Recursive partitioning analysis was compared to logistic, linear and quadratic discriminant analyses in the ability to differentiate hypercalcemic patients with primary hyperparathyroidism from those with malignancy. Stepwise discriminant analysis identified serum albumin as the best single discriminant test. Albumin decision values optimally separating the two hypercalcemic groups were 39.46, 38.54, and 32.25 g/l for the logistic, linear and quadratic discriminant methods, respectively. Recursive partitioning analysis identified carboxy-terminus parathyroid hormone (PTH) as the best discriminant test with an optimal decision value of 8.2 mequiv/l. The discrepancy between the selection of PTH by recursive partitioning analysis and albumin by discriminant techniques was attributed to the nonnormal distribution of PTH. Recursive partitioning analysis using PTH classified 85.4% of the patients correctly. Logistic, linear and quadratic methods, using albumin as the predictor variable, correctly classified 79.6%, 78.6%, and 79.6% of patients, respectively.