ABSTRACT

Cluster memberships associated with the mediation effect are often changed due to the temporal distance between the cause-and-effect variables in longitudinal data. Nevertheless, current practices in multilevel mediation analysis mostly assume a purely hierarchical data structure. A Monte Carlo simulation study is conducted to examine the consequence of ignoring the changes in cluster memberships in multilevel mediation analysis. Results show that the proposed method, Cross-Classified Multilevel Mediation Model (CC-M3), outperforms the conventional multilevel model with substantially smaller relative biases in parameter estimates (about 50% less) and a more consistent and higher coverage rate. Findings of this simulation study inform the empirical researchers that the changes in cluster-membership needs to be appropriately taken into consideration in mediation analysis. We demonstrate the use of CC-M3 in the applied example.

摘要

由于纵向数据中因果变量之间的时间距离，与中介效应相关的集群成员资格通常会发生变化。然而，当前多级中介分析的实践大多假设纯粹的分层数据结构。进行蒙特卡罗模拟研究以检查在多级中介分析中忽略集群成员变化的后果。结果表明，所提出的方法，交叉分类多级中介模型（CC-M3），优于传统的多级模型，参数估计的相对偏差明显更小（约少 50%），覆盖率更一致且更高。该模拟研究的结果告诉实证研究人员，在中介分析中需要适当考虑集群成员的变化。我们在应用示例中演示了 CC-M3 的使用。