Alternative approaches for the evaluation of biomarkers in mediation analysis

Investigating Alternative Approaches to the Evaluation of Biomarkers in Mediation Analysis Biomarkers provide important information in the diagnosis and prognosis of a disease. In the evaluation of biomarkers, the t-test or ROC analysis is often used. These methods only evaluate the relationship between two variables (dependent and independent). However, the relationship between the dependent variable and the independent variable may depend on a third variable. Examining how a third variable affects this relationship is important for the evaluation of biomarkers. Mediation analysis is used to test the effect of this third variable called mediator variable. In the literature, Baron-Kenny (BK) and Sobel methods are the most frequently used methods to test the effect of mediating variables. However, these methods have limitations. The BK method makes higher Type 1 and Type 2 errors when evaluating the mediating variable effect. The Sobel method, on the other hand, violates the normality assumption and neglects the possible correlation structure. In this thesis, it is aimed to examine the performance of BK and Sobel methods and to develop a decision mechanism that can be an alternative to these methods by considering the limitations of these methods. This decision mechanism is named as the proposed method. In addition, the proposed decision mechanism is compared with BK, Sobel, Aroian, Bootstrap and ML methods proposed by McGuigan and Langholtz, which are used in the literature to test the mediation effect. The proposed method is evaluated using simulation study and real data. In the simulation study, the effect of the mediating variable was evaluated for different combinations of proportion mediated (0.03-0.57) and sample sizes (25-1100) for the case where the dependent variable is continuous and binary. When the results of the simulation study are analyzed, the performance of the Proposed method is higher than the other methods at small sample sizes and when the effect size is medium and large when the dependent variable is continuous and binary. When the results of the real data are analyzed, the proposed method outperformed the other methods at small sample sizes. As a result, it is recommended to use the proposed method as an alternative when the dependent variable is continuous and binary in small sample sizes and when the effect size is medium and large. Key words: Baron-Kenny, Sobel, Proposed method, Mediator, Mediation analysis, Biomarker