Comparison of bias evaluation methods in meta analysis

Meta analysis is a method that combines the results of different studies conducted by different researchers at different places and times on the same subject. Statistical tests that provide objective evaluation of funnel plots, which provide a subjective evaluation to determine whether the information obtained as a result of meta analysis is biased, offers the opportunity to evaluate the reliability of meta analysis results. Examining the performance of statistical tests to detect bias provides information about the time and manner of use of the relevant statistical tests, and the research conditions in which they are appropriate to be applied. In this thesis, the performances of the Begg, Egger, Thompson, Schwarzer and Harbord tests in the literature, which are used to determine the bias in the funnel plot asymmetry evaluation in the meta analysis of binary data, were examined in terms of Type-Ⅰ error rates and powers. How the performances of the tests in terms of Type-Ⅰ error protection and powers were affected in cases of different number of studies, different sample sizes, different degrees of bias and different disease-outcome ratios were examined with simulation scenarios. As a result of the simulation studies, Schwarzer and Harbord tests have low statistical power to define funnel plot asymmetry, and Begg, Egger and Thompson tests have inflated Type-Ⅰ error rates when there is no funnel plot asymmetry. In the given conditions, there is no best test for detecting bias that can be used for all types of data between tests. Since the performance of tests varies with the number of studies included in the meta analysis, levels of bias, disease-outcome ratios, and sample size, these variables need to be taken into account when choosing methods for detecting bias. As a result of the simulation scenarios discussed in this thesis, it can be said that the Schwarzer test shows the best performance in terms of preserving the Type-Ⅰ error probability value if the number of studies included in the meta-analysis and the sample size is low, and the Harbord test if it is more. As a result of the simulation scenarios, when the performance of the tests in terms of detecting bias was evaluated, the Egger test showed the best performance.