CAJ | 학술논문

多重中介模型是指存在多个中介变量的模型。多重中介模型可以分析特定中介效应、总的中介效应和对比中介效应。指出了目前多重中介模型分析普遍存在的问题，包括分析不完整、使用Sobel检验带来的局限。建议通过增加辅助变量的方法进行完整的多重中介效应分析，使用偏差校正的Bootstrap方法进行中介检验。总结出一个多重中介SEM分析流程，并有示例和相应的MPLUS和LISREL程序。随后展望了辅助变量和中介效应检验方法的发展方向。
다중중개모형시지존재다개중개변량적모형。다중중개모형가이분석특정중개효응、총적중개효응화대비중개효응。지출료목전다중중개모형분석보편존재적문제，포괄분석불완정、사용Sobel검험대래적국한。건의통과증가보조변량적방법진행완정적다중중개효응분석，사용편차교정적Bootstrap방법진행중개검험。총결출일개다중중개SEM분석류정，병유시례화상응적MPLUS화LISREL정서。수후전망료보조변량화중개효응검험방법적발전방향。
The analyses of mediation effects are frequently applied to the studies of psychology, education, and other social science dis- ciplines. More than one mediator may be involved when the relationship among more than three variables is concerned. For a model with multiple mediators, there are three kinds of mediation effects ： total mediation effect, specific mediation effect through a specified path, and contrast mediation effects for the comparison of two or more specific effects. Compared with analyzing multiple mediators by building up several separate models with a single mediator, an equivalent model with multiple mediators based on structural equation modeling （SEM） has many advantages. For example, specific mediation effects can be tested in the condition controlling other mediators in the model ; total mediation effect which is the sum of the specific mediation effects can be tested; contrast mediation effects can be calculat- ed to determine the relative magnitudes of the different specific mediation effects. The purpose of the present study is to summarize an effective procedure for analyzing multiple mediators based on structural equation modeling. There are at least three weaknesses frequently found in current empirical studies involving multiple mediation effects. First, not all of the three kinds of mediation effects are considered, resulting in incomplete analyses of multiple mediation effects. Second, the Sobers testing method is dominantly used ; but the test method is based on the normality assumption that is typically violated by any kind of the mediation effects because they include the product of two parameters. Third, the computations of standard errors of multiple mediation effects often require manual calculations. In the present study, we propose a procedure to analyze the model with multiple mediators. The procedure is able to deal with both manifest and latent variables, and overcome all the three weaknesses described above. The first step is to establish a model including multiple mediators based on the theoretical framework in the field. In the second step, some auxiliary （phantom） variables are intro- duced into the model. These auxiliary variables will help researchers to obtain all of the three kinds of mediation effects if the output of SEM software does not provide them directly. In the third step, the bias -corrected percentile Bootstrap method, which can be imple- mented easily by the Mplus and LISREL software, is used to analyze multiple mediation effects. It shows that the corresponding media- tion effect is significant if a confidence interval does not include zero. Of course, the results of Bootstrap SEM analysis are acceptable only when the SEM model is fitted well. We used an example to illustrate how to conduct the proposed procedure by using the Mplus and LISREL software. The Mplus and LISREL program is attached to facilitate the implementation of the bias - corrected percentile Boot- strap method to analyze multiple mediation effects. The programs can be managed easily by empirical researchers. In fact, in addition to the Bootstrap method, the Bayesian method can also be selected to analyze multiple mediation effects, the results of the Bayesian SEM analysis are acceptable only when the SEM model is fitted well and the Markov chain is converged. It is possible for the Bayesian method to improve the power to detect mediation effects by incorporating prior information about the indirect effect.