Yaşam Çözümlemesinde Alıcı İşlem Karakteristiği Eğrileri

Abstract

The structure of life data includes life times and status indicating whether the event occurred. The basic assumption in survival analysis is that the event of interest will occur if the follow-up period is long. This assumption is often unrealistic since we are interested in the time until the event of interest occurs. Receiver operating characteristic (ROC) curves are a way to evaluate predicted performance. For test results defined on a continuous scale, ROC curves are the standard summary of accuracy. ROC curves are used to obtain sensitivity and specificity by using a continuous variable for the state variable. ROC curves used in the analysis of survival data determine how well they discriminate between those who experience the event and those who do not, allowing to choose the correct cut-off value. Since time is involved in survival analysis and the situation may change, classical ROC curves do not give accurate results. For this reason, new methods have been developed by researching time-dependent ROC curves and ROC curve estimates denoted by ROC(t) have been proposed. Among the ROC curve estimator methods used for survival analysis in this study, under the title of cumulative sensitivity and dynamic specificity (CD): Heagerty et al.'s (2000) Kaplan-Meier estimator (CD1), Heagerty et al.'s (2000) nearest neighbor estimator (CD2), Chambless and Diao's (2006) Kaplan-Meier-like estimator (CD3), Chambless and Diao's (2006) alternative estimator (CD4), inverse probability of intercept weighting (CD5), conditional inverse probability of stopping weighting (CD6), weighted AUC(t) (CD7), Viallon and Latouche (2011) estimator (CD8); Cox regression (ID1), weighted average ranking (ID2), fractional polynomial (ID3) under the title of incident sensitivity and dynamic specificity (ID); under the title of incident sensitivity and static specificity (IS); marginal regression modeling approach (IS1), extended Cox regression (IS2) and finally Naive estimator are introduced. To demonstrate the applicability of these estimators, an application was made on lung cancer data, which is literature data, and cervical cancer data, which is real data. In both data sets, CD1 and CD2 definitions gave similar results. It has been observed that, except for the CD4 definition, the predictive performance of the marker decreases as life expectancy increases in other definitions, while the predictive performance of the marker increases over time in the CD4 definition.