Abstract: Over the past 30 years, most efforts to test for mediation have been based on cross-sectional data, which may not to get causal inference. A possible solution for this could be to collect longitudinal data and perform a longitudinal mediation analysis. There are three causal arrows in a simple mediation model for analyzing a system of causality. If there is at least one causal arrow where the effect arises sometime after the cause, a longitudinal mediation design will be necessary for effectively observing the causation. There are three types of longitudinal mediation analysis approaches: 1) Cross-lagged panel model (CLPM); 2) Multilevel mediation model (MLM); 3) Latent growth mediation model (LGM). There are four types of the development of longitudinal mediation analysis. First, time-varying effect of mediation effect was tested. Continuous time models (CTM) would illustrate how mediating effects vary as a function of lag. Multilevel time-varying coefficient model (MTVCM) can capture direct and indirect effects over time. Second, individuals-varying effect of mediation effect was investigated. Random-effects Cross-lagged panel model (RE-CLPM) and Multilevel autoregressive mediation model (MAMM) should be adopted to analyze longitudinal mediation. Third, integration between different longitudinal mediation models, the outstanding performance is that is the integration of CPLM and MLM into MAMM. Fourth, the method testing mediation analysis was compared. Bayesian method should be adopted in mediation analysis of MAMM and MTVCM. Bootstrap method should be adopted in mediation analysis of LGM. Monte Carlo method should be adopted in mediation analysis of RE-CLPM. At the present study, we propose a procedure to analyze longitudinal mediation analysis. The first step is to decide whether it is necessary to make a causal inference. If the aim of research is making a causal inference, go to the second step. Otherwise, go to the third step. In the second step, we decide whether it is necessary to test time-varying effect of mediation effect. If the aim of research is testing time-varying effect of mediation effect, CTM should be adopted to analyze longitudinal mediation. Otherwise, go to the fourth step. The third step is to decide whether it is need to test time-varying effect of mediation effect. If the aim of research is testing time-varying effect of mediation effect, MTVCM should be adopted to analyze longitudinal mediation. Otherwise, LGM or MLM should be adopted to analyze longitudinal mediation. The fourth step is to decide the model would fit by running a RE-CLPM model and CLPM. If AIC and BIC indictors of RE-CLPM are smaller than indictors of CLPM, RE-CLPM should be adopted to analyze longitudinal mediation. Otherwise, go to the fifth step. The fifth step is to decide whether it is necessary to investigate individuals-varying effect of mediation effect. If the aim of research is investigating individuals-varying effect of mediation effect, MAMM should be adopted to analyze longitudinal mediation. Otherwise, CLPM should be adopted to analyze longitudinal mediation.

Key words: longitudinal data, mediation effect, cross-lagged panel model, multilevel model, latent growth model

摘要： 目前中介效应检验主要是基于截面数据，但许多时候截面数据的中介分析不适合进行因果推断，因而需要收集历时性的纵向数据，进行纵向数据的中介分析。评介了基于交叉滞后面板模型、多层线性模型和潜变量增长模型的纵向数据的中介分析方法及其四个发展。第一，中介效应随时间变化，如连续时间模型、多层时变系数模型。第二，中介效应随个体变化，如随机效应的交叉滞后面板模型和多层自回归模型。第三，中介模型的整合，如交叉滞后面板模型与多层线性模型整合为多层自回归模型。第四，中介检验方法的发展，建议使用Monte Carlo、Bootstrap和贝叶斯法进行纵向数据的中介分析。总结出一个纵向数据的中介分析流程并给出相应的Mplus程序。随后展望了纵向数据的中介分析的拓展方向。

关键词: 纵向数据, 中介效应, 交叉滞后面板模型, 多层线性模型, 潜变量增长模型