Murdoch University Research Repository

Welcome to the Murdoch University Research Repository

The Murdoch University Research Repository is an open access digital collection of research
created by Murdoch University staff, researchers and postgraduate students.

Learn more

Effect of non-linearity of a predictor on the shape and magnitude of its receiver-operating-characteristic curve in predicting a binary outcome

Ho, K.M. (2017) Effect of non-linearity of a predictor on the shape and magnitude of its receiver-operating-characteristic curve in predicting a binary outcome. Scientific Reports, 7 (1).

[img]
Preview
PDF - Published Version
Available under License Creative Commons Attribution.

Download (1MB) | Preview
Free to read: https://doi.org/10.1038/s41598-017-10408-9
*No subscription required

Abstract

Area under a receiver-operating-characteristic (AUROC) curve is widely used in medicine to summarize the ability of a continuous predictive marker to predict a binary outcome. This study illustrated how a U-shaped or inverted U-shaped continuous predictor would affect the shape and magnitude of its AUROC curve in predicting a binary outcome by comparing the ROC curves of the worst first 24-hour arterial pH values of 9549 consecutive critically ill patients in predicting hospital mortality before and after centering the predictor by its mean or median. A simulation dataset with an inverted U-shaped predictor was used to assess how this would affect the shape and magnitude of the AUROC curve. An asymmetrical U-shaped relationship between pH and hospital mortality, resulting in an inverse-sigmoidal ROC curve, was observed. The AUROC substantially increased after centering the predictor by its mean (0.611 vs 0.722, difference = 0.111, 95% confidence interval [CI] 0.087-0.135), and was further improved after centering by its median (0.611 vs 0.745, difference = 0.133, 95%CI 0.110-0.157). A sigmoidal-shaped ROC curve was observed for an inverted U-shaped predictor. In summary, a non-linear predictor can result in a biphasic-shaped ROC curve; and centering the predictor can reduce its bias towards null predictive ability.

Item Type: Journal Article
Murdoch Affiliation(s): School of Veterinary and Life Sciences
Copyright: © 2017 The Author
URI: http://researchrepository.murdoch.edu.au/id/eprint/38538
Item Control Page Item Control Page

Downloads

Downloads per month over past year