구조방정식 모델에서 항목묶음이 인과 모수의 검정과 적합도 평가에 미치는 영향 (The Effects of Item Parceling on Causal Parameter Testing and Goodness-of-Fit Indices in Structural Equation Modeling)

서지정보

· 발행기관 : 한국마케팅과학회
· 수록지 정보 : Journal of Global Scholars of Marketing Science(마케팅과학연구) / 17권 / 3호 / 133 ~ 153페이지
· 저자명 : 조현철, 강석후

초록

본 연구에서는 3개 일반모델(general models)의 실증분석을 통해 항목묶음(item parceling)이 구성개념 간의 인과관계를 나타내는 모수의 유의성 검정 결과 및 모델의 적합도 평가에 미치는 영향을 검토하였다. 연구 결과에 의하면, 개별항목을 적용한 분석과 비교할 때 항목묶음을 통한 분석을 적용해도 모델 인과모수의 검정통계량은 그다지 변하지 않으므로 유의성 검정 결과에도 변화가 없는 것으로 나타났다. 하지만 전반적 적합도지수의 측면에서는 RMSEA를 제외한 주요 모델 적합도 지수, 즉 GFI, AGFI, CFI 및 NFI의 값들이 상당히 개선되는 경향을 보였다. 주요 모델 적합도 지수들의 값이 높아진 것은 항목묶음을 이용하여 분석을 한 결과가 개별항목을 통한 분석의 결과에 비해 다변량 정규(분포)성의 개선 등으로 인해 높아진 것으로 해석된다. 하지만 항목묶음을 적용함에 있어서 주의해야 할 사항은 적용하기 전에 구성개념의 단일차원성(unidimensionality)을 우선적으로 검토해야 한다는 점이다. 본 연구에서는 항목묶음을 이용하여 분석을 할 경우 실제 구성개념간의 인과적 관계를 나타내는 모수의 유의성 검정과 모델의 적합도 지수들에 어떤 변화가 발생하는 지를 세 개의 일반모델을 대상으로 파악하였다.

영어초록

The purpose of this article is to examine the effects of item parceling on the consistency of significance testing of the causal parameters with regard to the relationship between the relevant constructs. It is also to investigate the effects of item parceling on the goodness-of-fit indices of LISREL's general models. Most researchers use structural equation modeling (SEM) to test their research hypotheses associated with the causal parameters. We investigated three general models of LISREL rather than the frequently used confirmatory factor analytic (CFA) models used by many researchers. The results of the study showed that there was a high level of consistency in the calculated test statics of causal parameters between the item-parceled solutions and the item-level solutions, and that the item-parceled solutions had better goodness-of-fit indices, such as GFI, AGFI, CFI, and NFI, than the solutions at the item level. However, in terms of RMSEA, there was no such tendency. GFIAGFICFINFIRMSEAIndividual Items0.7580.7100.8960.8480.0919Two Items Parceled0.8580.7540.9380.9210.130Table 2. Improvements in GFI Indices in General Model I.
Specifically, in the examination of our general Model I, we had highly consistent test results of causal parameters between the item-parceled solution and the item-level solution. In the item-parceled solution, we achieved improved GFI, AGFI, CFI, and NFI but worse RMSEA. The constructs tested in Model I were self-image congruence, products knowledge, brand dependability, brand emotion, purchase loyalty, and attitude loyalty. CR (composite reliability) and AVE (average variance extracted) for each construct in Model I were as follows: self-image congruence (CR = 0.78, AVE = 0.48), products knowledge (CR = 0.65, AVE = 0.32), brand dependability (CR = 0.90, AVE = 0.70), brand emotion (CR = 0.92, AVE = 0.71), purchase loyalty (CR = 0.84, AVE = 0.64), and attitude loyalty (CR = 0.80, AVE = 0.51). The range of all the indices for standardized coefficients was between 0.592 and 0.957 at a significance level of ps<0.001.
In the case of investigating our general Model II, we had the same consistent test results between the two solutions with regard to the testing of causal parameters. The indices such as GFI, AGFI, CFI, NFI and RMSEA were improved in the model II. The constructs used in model II were organizational control, individual control, job definiteness, emotional commitments, salesperson's performance, and customer orientation. CR (composite reliability) and AVE (average variance extracted) for each construct in model II were as follows: organizational control (CR = 0.75, AVE = 0.53), individual control (CR = 0.77, AVE = 0.47), job definiteness (CR = 0.92, AVE = 0.71), emotional commitments CR = 0.85, AVE = 0.42), salesperson's performance (CR = 0.82, AVE = 0.43), and customer orientation (CR = 0.71, AVE = 0.35). The range of all indices of standardized coefficients was between 0.587 and 0. 941 at the significance level of ps<0.001. However, two items for customer orientation were not statistically significant.
GFIAGFICFINFIRMSEAIndividual Items0.6710.6240.7570.7110.110Two Items Parceled0.8990.8360.9320.9120.0978
Table 4. Improvements in GFI Indices in General Model II
The result of our examination of our general Model II showed highly consistent test results in both solutions with regard to causal parameter tests and the item-parceled solution resulted in improved GFI, AGFI, CFI, NFI, and RMSEA. The constructs involved in model III were congruence between the individual and the organization, leader's support, fairness in compensation, job satisfaction, and organizational citizenship behavior (OCB). CR (composite reliability) and AVE (average variance extracted) for each construct in Model III were as follows: congruence between individuals and the organization (CR = 0.84, AVE = 0.63), leader's support (CR = 0.89, AVE = 0.62), fairness in compensation (CR = 0.88, AVE = 0.64), job satisfaction (CR = 0.78, AVE = 0.54), and OCB (CR = 0.88, AVE = 0.73). The range for all indices of the standardized coefficients were between 0.392 and 0. 907 at a significant level of ps<0.001.
GFIAGFICFINFIRMSEAIndividual Items0.8810.8460.9250.8860.0724Two Items Parceled0.9560.9260.9760.9520.0537Table 6. Improvements in GFI Indices in General Model III

In addition, we showed that better item-parceling can be made if we use AVE (average variance extracted) and SMC (squared multiple correlation) in order to select good indicators. Finally, we noted that with regard to item-parceling, the most important point is that we should confirm the unidimensionality of the indicators in each construct beforehand.

참고자료

· 없음