本研究关注因变量为分类数据的情境下，中介效应模型的使用以及不同分析方法的比较。分别通过在单一水平和多水平固定中介效应模型下，分类因变量分析方法与违背变量类型要求的连续型因变量分析方法的比较，说明了当因变量为分类数据时，采用正确的处理分类数据分析方法的必要性。同时，本研究考察了样本量（或者样本量组合）、中介效应大小以及分类因变量所含类别数等因素对中介效应估计的影响。研究结果如下： 单一水平中介效应模型中，当因变量为分类变量而误将其作为连续变量来分析，与使用正确的分类数据分析方相比，其结果明显存在中介效应估计值精度偏低、中介效应置信区间对真值覆盖比例偏低、低估标准误等问题。 在中介效应较大或者样本量较小时，需要考虑分类数据分析方法的估计精度和统计检验力是否较低的问题。分类数据分析方法中，中介效应置信区间对真值覆盖比例在各种条件下始终稳定在95%左右。无论是分类数据分析方法还是连续数据分析方法，系数差异法的结果总存在估计精度偏低、第Ⅰ类错误率较大等问题。随着因变量类别数的增加，连续数据分析方法的中介效应估计偏差变小、中介效应置信区间对真值覆盖比例和标准误随之增大，但是在各个指标上分类数据分析方法的结果还是更优一些。 在多水平固定中介效应模型中，当因变量为分类变量时，采用分类数据分析的标准化方法和采用连续数据分析的结果均相差无几，而分类数据模型中未标准化方法的结果仅在估计精度表现为很不稳定。尽管如此，与分类数据分析方法相比，连续数据分析方法的残差ICC估计精度较低。但在类别数较大时可以考虑使用连续数据分析方法。 当跨级相关（ICC）和组间样本量较小时，需要考虑中介效应的估计精度是不是偏低以及标准误是不是过大的问题。随着因变量类别数的增加，多水平固定中介效应的估计精度升高，中介效应置信区间对真值的覆盖比例和统计检验力随之增大，估计标准误随之减小。但是在较小的组样本量中，各路径系数的收敛难度增加。

The study focus on the sensible use of categorical data analysis rather than continuous data analysis for exploring single mediation and fixed multilevel mediation model when outcome variable is binary or ordinal. Sample size, true mediation level and the number of categories for outcome variable was taken into account for comparison of the two analysis method. The results as follows: First, in single mediation model for categorical outcome variable, there was always low estimation precision, poor statistical test and SE underestimated when forced to use continuous data analysis. On the contrary, estimation results from logistic mediation model always performed better than the continuous data analysis methods. When small mediation value or small sample size appear, low estimation precision and statistical power should be examined. Difference of coefficient always performed not well whatever method was used. As the number of categories increased, although absolute estimation bias, RMSE and estimated standard error of mediation effect derived from continuous data analysis showed decreases, the statistical power increased, obviously categorical data analysis had its advantage among these comparative indicators. Secondly, in the fixed multilevel mediation model, estimation results from the two analysis method made little difference when the outcome variable is a categorical, but unstandardized mediation effect in the categorical data model work less unstably only in estimation accuracy. Moreover, estimation accuracy of error intraclass correlation (ICC) from categorical analysis was better than its outcomes from continual analysis. If low ICC or small group size happen in practice, low estimation accuracy and SE would like happen. When categories of outcome variable increased, estimation accuracy, coverage rate in true value CI and statistical power increased. All of though, the coverage difficulty also increased.