Estimation and comparison of covariate adjusted semi-parametric and nonparametric ROC curves by bayesian and classical methods

Aim: The comparison of AROC methods in which the covariate effect is also evaluated in the ROC analysis, which is used to determine the effectiveness of biomarkers that have an important place in medical decisions and diagnostic tests, is included in the scope of the thesis. In this framework, semi-parametric and non-parametric classical methods and Bayesian semi-parametric and non-parametric methods, in which correction is made according to the covariate, are the methods determined within the scope of the thesis study, and it is aimed to estimate these methods and make performance comparisons. Methods: In the evaluation of the methods, different distribution scenarios were created for the biomarker associated with the covariate with a uniform distribution, and for each of these scenarios, the criterion AUC values, which are expected to converge, were determined. AUC and MSE were taken into account in the performance comparisons of the models. In addition, in order to evaluate the performance of the methods in different sample sizes for patient and healthy groups, data were derived as n1 = n0 = (25, 50, 100, 200, 500) each time in the simulation study, and 1000-times simulation studies for each combination were designed. Results: When the findings obtained from the simulation studies were examined, it was seen that all AROC methods, especially semi-parametric methods, had similar performance in data sets with normal distribution. In the skewed distribution scenarios, it has been observed that non-parametric methods, especially Kernel, have higher performance. Conclusion: As a result, in real life, it is a rare situation that biomarker values show normal distribution in both patients and healthy populations. For non-normally distributed data sets, it is suggested to use the Kernel method as the adjusted AROC model. Keywords: AUC, covariate, adjusted ROC, simulation, PSA