类别变量的中介效应分析

Mediation Analysis of Categorical Variables

在心理学和其他社科研究领域,研究者能熟练地进行连续变量的中介效应分析,但面对自变量、中介变量或(和)因变量为类别变量的中介效应分析,研究者往往束手无策。在阐述类别自变量中介分析方法的基础上,我们建议使用整体和相对中介相结合的类别自变量中介分析方法,并给出了分析流程。以二分因变量为例,讨论了中介变量或(和)因变量为类别变量的中介分析方法的发展过程(即尺度统一的过程),建议通过检验Za×Zb的显著性来判断中介效应的显著性。用二个实际例子演示如何进行类别变量的中介效应分析。最后展望了类别变量的中介效应分析研究的拓展方向。

In the research of psychology and other social science disciplines, researchers often do not know how to analyze the mediation effect when the independent, mediator or/and dependent variables are categorical, even if they can skillfully conduct a mediation analysis of continuous variables. The conventional mediation analysis, transforming multi-categorical variables into dichotomous or continuous variables or using analysis of variance （ANOVA）, might be lacking in efficiency when the independent variables are multi-categorical. A procedure is proposed and recommended to use the method integrating relative mediation with Omnibus mediation to analyze the mediation effect when the independent variables are multi- categorical. The first step is to implement an Omnibus mediation analysis. If the Omnibus mediation effect is not significantly different from zero, the k-1 relative mediation effects are zero, where k is the number of the categories. Otherwise, go to the second step. In the second step, a relative mediation analysis is used to find out if each relative mediation effect is significant. If there is no relative mediation effect which is significantly different from zero, the mediation analysis is done. Otherwise, go to the third step. In the third step, the results with relative direct effects are reported. An example is given to illustrate how to conduct the proposed procedure by using the SPSS software. Then, the evolution of the mediation analysis method with categorical mediator or dependent variable is discussed, and the scale unified process is the focus. In early years, the product of coefficients （ab） obtained from the logistic regression was used to analyze the mediation effect when mediator or dependent variable were categorical. Later, ab^std was adopted to analyze the categorical mediation effect. Recently, Za × Zb was used to analyze the mediation effect. We suggest that Za × Zb is preferred to analyze the mediation effect when mediator or dependent variables are categorical. In addition, we emphasized that asymmetric interval was used to test the significance of Za × Zb. We used an example to illustrate how to conduct the proposed procedure by using the SPSS software. The directions for future study on categorical mediation are discussed at the end of the paper. In fact, in addition to dummy coding, sequential coding and contrast coding are an alternative to code multi-categorical independent variables, and the results of the mediation effect tests with these three coding methods are equivalent, but the coding method will influence the implication of the relative indirect, direct, and total effects. Furthermore, the method integrating relative mediation with Omnibus mediation to analyze the mediation effect of multi-categorical independent variables could be expanded to more complicated mediation models, such as single-step or parallel multiple mediator models in which there is more than one mediator. Finally, it is a promising direction to analyze the mediation effect with binary dependent variables in Structural Equation Modeling.