Clinical validation of biomarkers for predicting risk

A useful biomarker should improve clinical management in an economically reasonable way. This should be determined from well-designed outcome studies that show clinical management can be altered on the basis of the biomarker. It is important not to confuse results from testing prior to outcome with outcome studies.

This chapter reviews statistical tests used to evaluate studies performed prior to final outcome studies and criteria that assess whether or not a biomarker should be considered for outcome studies at each step. I review how relative risk and odds ratios are related to receiver operator characteristic (ROC) plot analysis. Other statistical techniques such as reclassification and the Hosmer Lemeshow test that have been suggested for evaluation of diagnostic usefulness are considered. Weaknesses of each technique are discussed.

I consider ROC analysis to be a mainstay against which other statistical tests of diagnostic performance should be compared. The importance of expressing data in terms of predictive values is emphasized. Tests showing weak diagnostic associations with a disease are difficult to evaluate for outcome study application, because there is usually great difference in between-study variance so that the true relationship between the biomarker, its diagnostic ability, and predictive capability are unclear.