Modeling individual and relative accuracy of screening tools in geriatric oncology

Classification probabilities reflect to what degree a screening test represents the true disease state and include true positive (TPF) and false positive fractions (FPF). With two tests, one can compare TPF and FPF using relative probabilities which offer advantages in terms of interpretation and statistical modeling. Our objective was to highlight how individual and relative TPF and FPF can be easily estimated and compared within a regression modeling framework. This allows the modeling of tests’ accuracy while adjusting for multiple covariates, and thus provides valuable information in addition to the crude TPF and FPF. We illustrate our purpose with the G8 and VES-13 screening tests aimed at identifying elderly cancer patients in need for a comprehensive geriatric assessment (CGA). Prospective cohort with a paired design. TPF and FPF of each test, as well as relative TPF and FPF were modeled using log–linear models. G8 detected patients in need for CGA better than VES-13 at the expense of misclassifying a large number of normal patients. Both tests had better TPF with older age and poorer performance status (PS), and for all cancer subtypes compared with prostate cancer. Effect of age and PS on TPF was more pronounced with VES-13. Age affected FPF, but not differentially. Regression modeling helps provide a thorough assessment of the accuracy of diagnostic tests and should be used more frequently. In the context of screening, we encourage the use of G8 as failing to identify patients in need of a CGA might be more problematic than over-detection. Moreover, although we identified variables associated with the sensitivity of these tests, this association was less pronounced for the G8.