888Discriminating between risk discriminators: OPERA, AUC, and polygenic variance

Abstract

Abstract Focus of Presentation Epidemiological risk estimates are often adjusted for age (and, if necessary, sex) by design or analysis, and for other factors. Consequently, risk estimates do not pertain to the crude risk factor, but to its population residual after adjusting for age, sex and other covariates. For disease risk, the change in Odds PER Adjusted standard deviation (OPERA) estimates the risk gradient on an appropriate scale; log(OPERA) is the natural measure to compare estimates. Findings Under a multiplicative risk model for a normally distributed (adjusted) risk factor, log(OPERA) = the difference in mean between cases and controls. The area under the receiver operating curve (AUC) = □(log(OPERA)/√2), where □ is the standard normal cumulative distribution function. The risk discrimination from combining risk factors can be predicted from their OPERAs. The polygenic standard deviation estimated from pedigree data = log(OPERA). The OPERA for knowing all familial risk factors can be calculated. Conclusions/Implications We present examples from the breast cancer literature where the wrong conclusions can be made by not using the OPERA concept. We give examples of the value of OPERA estimates in predicting the risk discrimination of their combination and demonstrate why the better one predicts the disease the harder it is to predict it better. Key messages OPERA overcomes problems about interpreting risk factors, and combinations of risk factors, in a way not apparent using changes in AUC. OPERA also puts an upper bound on the role of genetic factors in explaining differences in risk across the population.