組織研究中的多層次調節式中介效果：以組織創新氣氛、組織承諾與工作滿意的實證研究為例

Multilevel Moderated Mediation of Organizational Study: An Empirical Analysis of Organizational Innovation Climate, Organizational Commitment and Job Satisfaction

在組織研究中，由於個體資料鑲嵌於組織之中，組織對個體的影響則涉及脈絡效果、跨層級交互作用與多層次中介及調節效果的檢驗。傳統Baron 和Kenny(1986)的中介效果檢測，並無法適用於多層次資料結構的分析。本文目的是在Krull 和MacKinnon(1991, 2001)的多層次中介效果模式的設定下，延伸到跨層級交互作用，同樣利用Baron 與Kenny 檢測中介效果的觀念，提出檢驗多層次調節中介效果（3M）的(2-1-1)程序，文中除了分析原理的論證，並佐以實證範例來說明驗證方法。實證資料來自於24 家企業664 名員工的組織創新氣氛、組織承諾（認同與工具承諾）與員工滿意度，以HLM 與Mplus 進行3M 模型的檢測。結果發現，員工的認同承諾會完全中介組織創新氣氛對員工滿意度的影響；此外，組織創新氣氛會調節員工的工具承諾對員工滿意度的影響。實證分析的結果顯示，多層次資料結構的中介與調節作用必須以3M 分析程序來逐步檢驗。最後，本文對於3M 方法的特性與限制進行詳細討論。

In the field of organization research, the common major concerns of researchers are the effects of mediation and moderation. For the single-level dataset, the analytical procedure for the mediation effect proposed by Baron and Kenny (1986) is still taken as the standard procedure. But in the circumstance wherein the research data have a multilevel structure, the organization-level variables may have a top-down influence on the individual-level variables. A particular analytical paradigm is needed to be developed to deal with the nested or clustered data, such as the data wherein employees are nested within departments, and the departments are nested within organizations. In a multilevel data structure, the conventional Baron and Kenny procedure for the mediation effects becomes inappropriate due to the violations of the assumptions of independence and homogeneity. Neglecting the correlated property of the observations such as employees within the organization experiencing the shared organizational climates and cultures thus may result into misleading conclusions. Similarly, the results from the traditional ordinary least squares estimators (OLS) are also incorrect because the standard errors would be underestimated. This article introduces a multilevel modeling method of analysis for the mediation and moderation effects, which is critical and necessary in the response to the complexity in the clustered or nested data. In multilevel models, the withingroup homogeneity and the between-group heterogeneity of errors from the clustered or nested data are taken into account by adding a random intercept term to the regression equation. The estimates of the standard errors generated by the multilevel model are unbiased. In addition, the examination of contextual effects and the cross-level interaction can also be integrated into the multilevel model. The direct effect of organizational-level variables on the individual-level outcome variables (i.e., contextual effects) as well as the moderation effects of organizational-level variables on the individual-level prediction relationship can be empirically examined in the proposed Multilevel Moderated Mediation (3M) paradigm. The conceptualization of the proposed 3M modeling is by integrating the ideas of multilevel mediation proposed by Krull and MacKinnon in 1991 and 2001, and together with the analysis of cross-level interaction; a central concept of multilevel regression paradigm. The analytical procedure for a 2-1-1 model provides a clear and detail framework in the demonstrated empirical data analysis. While, the notation of 2-1-1 is defined as follows: The first number defines the level of the independent variable wherein it belongs to. Number “1” represents the individual level and “2” for the organizational level. The second number defines the level of the location of the mediating variables, and the final number defines the level of the outcome variable. The standard steps for examining a 3M model are, firstly, calculating the intraclass correlation coefficients (ICC); secondly, executing the “intercepts-as-outcome model”; thirdly, examining the “random effect ANCOVA model and random coefficient model”; followed by the procedures of testing of the “mediation effect”; and finally exploring the “moderated effect”. Step one is used to evaluate the magnitude of between-group variability, which has to be strong enough for impacting the other estimators of OLS in the model. If the ICC is small, the influences of multilevel data structure could be ignored, and a traditional regression analysis is appropriate. The purpose of step two is to examine the direct effect of the organizational-level independent variable on individual-level outcome. Because the analytic units of analysis for organizational level and individual level are different, the multilevel modeling is applied herein. Through the intercepts as outcome model, the direct effect between organizational independent variable and individual outcome variable could be examined. Step three is use to examine the influence of the individual level mediator variable on the outcome variables in the same level. While step four takes the organizational level independent variable and individual level mediator variable into account simultaneously in one equation. By examining the significance of the partial coefficients of the organizational level independent variable and individual level mediator variable and comparing the partial coefficient of the organizational level independent variable and the direct effect in step two, we can test if the multilevel mediation is hold. In the step four, we examine the two different models about the slopes of the individual level mediator variable: the fixed effect and the random effect model. If the variance component of the slopes is significant in the random model, it is necessary to test the moderation of organizational level independent variable on the relationships between the individual level mediator variable on the individual level outcome variables. Step five is the cross-level interaction procedure to test the moderated role of the organizational level independent variable. In addition to the above formal five steps, we also suggest to proceed to the Sobel test, in order to assess the indirect effect of the moderator variable. 150The purpose of the illustration on analytic procedures is that the multiple mediators with the hypothesized 3M relationships of the independent, moderators, and outcome variable were empirically analyzed in the present paper. Partial data selected from a previous study (Chiou, Kau, & Liou, 2001) consisted of 664 employees sampled from 24 different organizations. The organizational level independent variable, individual-level mediators, and individual-level outcome variable are the perceived organizational climate, employee’s identification commitment and instrument commitment toward the organization, and job satisfaction, separately. It was hypothesized that the employee’s identification commitment and instrument commitment have a significant mediation effect on the prediction relationship of organizational climate on employee’s job satisfaction. It is also examined that how the organizational climate moderates the relationships between organizational commitment and job satisfaction. Data was analyzed by HLM (Raudenbush, Bryk, Cheong, & Congdon, 2004) and Mplus (Muthen & Muthen, 2004) both for the purpose of comparison among software. Results revealed that the individual-level identification commitment fully mediated the relationship between organizational climate and employee’s job satisfaction; at the same time, it is identified that the macro-level organizational climate moderated the influence of individual-level instrument commitment on job satisfaction. Sobel tests (Sobel, 1982) were applied to examine the significance of the mediation effect. The usage of the proposed 3M procedure was also demonstrated. Implications and technical limitations of the 3M paradigm are discussed in the final section, while the program code for HLM and Mplus are presented on the appendix.