A Comparative Study Of Five Mediation Analysis Methods For Survival Data

With the improvement of cancer diagnosis rate,more and more people focus on the risk factors and prognostic factors of cancer.Although the researchers can find different influence factors of cancer prognosis,but how these factors affect the prognosis of cancer is not clear.As a way to explore the causal mechanisms between exposure and cancer prognosis,Mediation analysis has being more and more popular.There are five methods can be used to do mediation analysis,including VanderWeele method,Baron and Kenny approach,Imai method,Sobel method and Inverse Weight method.The objectives of this study is to evaluate the five mediation analysis methods for survival data and provide suggestions for how to build models and choose the method.Method of this research is through performing statistical simulations under different scenario to evaluate the methods type ? error,power,degree of accuracy,accuracy and analysis efficiency.The simulation results show that when there is not interaction existing between exposure(X)and mediator(M),IW method has the lower power,degree of accuracy and accuracy than other methods.Baron-Kenny and Imai method is far lower speed than the VanderWeele and Sobel method.Although the principle of Baron-Kenny and Sobel method is simple and easy to understand,but the two methods have some limitations,which is they can not handle interaction between exposure and mediator and nonlinear exposure.VanderWeele method can deal with these problems,so it's application is wider.We can obtain indirect effect through effect minus the direct effect according to IW method,so IW method can not be used when there was interaction as also.Therefore,when there was interaction between exposure and mediator,VanderWeele and Imai method are prior.When there was interaction between exposed and mediator,the power of VanderWeele method and Imai method both can be decreasing when the ratio of mediation,sample size and total effect were decreasing,while censor was increasing.Power of VanderWeele method was affected by rate of interaction(X:M,%),and power of Imai method was decreasing when(X:M,%)was increasing.Whatever in any parameters cases,power of VanderWeele method is not bad than Imai method.The estimated accuracy of VanderWeele method was little affected by sample size,the total effect,correlation coefficient(X:M,%),mediation interaction(M%)and rate of censor(cen%).What's more,its estimation accuracy was much higher than the IW method.Furthermore,VanderWeele method can achieve the estimated value of direct effect,indirect effect and total effect while its analysis efficiency is high,so VanderWeele method is a recommended method.Considering the type I error,power,degree of accuracy,precision and test efficiency,we recommend VanderWeele method to analysis survival data.