Evaluation of the results of network meta analysis random effect models in different statistical software-a simulation study

Network meta analysis is an analysis method that statistically combines the results of studies conducted with the same purpose and method in the literature in which more than two methods are compared, and provides an overall effect size estimation. In this method, the aim is to create a network between multiple methods at the same time and evaluate the direct and indirect effect size estimations, increase the statistical power with the combined effect size obtained, evaluate the inconsistencies between studies and determine the best method effectiveness. In this thesis, it is aimed to discuss the applicability of network meta-analysis, which has been popular in recent years and has the highest evidence value, under different scenarios and to compare the results obtained in R and SAS programs in accordance with certain scenarios. In this study, in the comparison of three methods (A, B, C); including different network structures (closed and open) a 1000 replicates simulation was carried out under scenarios with different number of studies (balanced (5-5-5;10-10-10 / 5¬ 5; 10-10) or unbalanced (5-5-1; 5-5-10) / 5-1; 5-10)), with different sample sizes (n=30, 50, 100) and with different heterogeneity (I2<0.2 (low) and I2<0.5 (moderate)). For these scenarios, data was derived in the R package program. For the derived data, simulations were run in the R netmeta package and SAS glimmix procedure, and the results of the network meta analysis were compared. As a result of the simulation studies, it has been seen that the R netmeta package and the SAS glimmix procedure give similar results, and make valid estimations regardless of the number of studies, sample size and heterogeneity in open and closed networks in balanced condition, but it is seen that this result changes in scenarios including few studies in one arm in the case of unbalanced in open networks. In the arm with few studies (5-1), it was observed that both programs could not detect the existing difference between the methods in the presence of a small sample size (n=30). It was observed that this situation changed as the sample size increased, and the programs detected the difference between the methods although there was few studies in one arm when the sample size of n=100. Keywords: Network meta analysis; R netmeta; SAS proc glimmix; simulation study.