Weak associations between prognostic biomarkers and disease in preliminary studies illustrates the breach between statistical significance and diagnostic discrimination.
Clinica chimica acta; international journal of clinical chemistry, 411 (7-8):
467-73 (April 2010)5593<m:linebreak></m:linebreak>CI: 2010; JID: 1302422; 0 (Biological Markers); RF: 64; 2009/11/19 received; 2010/01/08 revised; 2010/01/11 accepted; 2010/01/18 aheadofprint; ppublish;<m:linebreak></m:linebreak>Proves diagnòstiques.DOI: 10.1016/j.cca.2010.01.015
Abstract
INTRODUCTION: Prior to actual well-designed outcome studies, statistical significance does not imply diagnostic usefulness. Sensitivity/specificity graphs have been a mainstay in evaluating potential diagnostic usefulness. Unfortunately, many recent studies have been expressed as relative risk (RR) or odds ratios (OR) based on statistical significance that produces misleading constructs, especially when the relationship between the biomarker and disease is weak. The purpose of this review is to discuss the meaning of the diagnostic relationships when the association between a biomarker and a disease is weak such that it shows poor discrimination by sensitivity/specificity plots. METHODS: Review of the literature, illustrative example and calculation of positive predictive values. RESULTS: Information from the literature showing the equivalency between OR and sensitivity/specificity plots and defining weak diagnostic discrimination. Demonstration of how confounders can alter between study statistical significance when there is a weak diagnostic relationship. CONCLUSIONS: RR or OR should be expressed as diagnostic sensitivity/specificity plots. Usually predictive values will identify whether or not there is potential diagnostic value that is economically reasonable. Very weak diagnostic relationships produce very low positive predictive values. Great caution should be exercised when considering weak associations for diagnostic use, especially if the biomarker is not a modifiable risk factor.