Whether a factor significantly increases the area under the curve (AUC) of a receiver operating characteristic analysis has become the standard test of its utility. Thus, in many studies, apolipoprotein B and LDL particle number have not increased the AUC significantly beyond that produced by the conventional markers, and guideline groups have concluded on this basis that they should not be added to routine clinical practice. This article demonstrates this conclusion to be invalid.
In conventional analyses, no distinctions have been drawn as to whether a novel predictor is causal or whether it is highly correlated with other markers already included in the risk algorithm. However, correlation among the markers will profoundly affect the incremental effect of a factor on the AUC. This distinction is particularly critical for factors that have been shown to play a causal role in the production of clinical event. Accordingly, the AUC approach is valid to determine the total discriminatory ability of a set of variables but is not appropriate to allocate attributable risk among the members of the set.
For correlated predictors that describe different aspects of the same variable such as non-HDL-C and apoB or LDL-C and LDL particle number, discordance analysis offers a simple valid alternative to capture and compare the independent information contained by each predictor.
aMike Rosenbloom Laboratory for Cardiovascular Research, McGill University Health Centre, Montreal, Quebec, Canada
bDepartment of Biostatistics and Bioinformatics, Duke University, DCRI, Durham, North Carolina, USA
cDepartment of Medicine, Royal Victoria Hospital, McGill University Health Centre, Montreal, Quebec, Canada
Correspondence to Allan D. Sniderman, MD, FRSC, Mike Rosenbloom Laboratory for Cardiovascular Research, Royal Victoria Hospital, McGill University Health Centre, 687 Pine Avenue West, Montreal, Quebec H3A 1A1 Canada. Tel: +514 934 1934; 34637; fax: +514 843 2843; e-mail: firstname.lastname@example.org