بررسی مقایسه ای مدل سازی معادله ساختاری و روش پژوهش های چندگانه رگرسیون:زمینه تجارت الکترونیکی

Structural equation modeling (SEM) is a powerful statistical technique that establishes measurement models and structural models. On the other hand, multiple regression (MR) is considered a sophisticated and well-developed modeling approach to data analysis with a history of more than 100 years. This paper empirically compares SEM and MR by testing a model of commitment in a B-to-C e-commerce travel context, shedding light on applications of these two popular methods in tourism research. The findings indicate that only two significant relationships are justified by MR. In comparison, SEM results reveal more statistically significant relationships after the “best-fitting” measurement model with model D being the “best-fitting” model. The findings support some key empirical limitations of MR as a widely used statistical technique in the tourism research.

Structural equation modeling (SEM) has recently become a popular statistical technique to test theory in a number of academic disciplines (Hair et al., 1998 and Schumacker and Lomax, 2004). It is a method of multivariate statistical analysis capable of measuring the underlying latent constructs identified by factor analysis and assessing the paths of the hypothesized relationships between the constructs (Klem, 2000). Overall, SEM has two main advantages: (1) it allows for the estimation of a series, but independent, multiple regression equations simultaneously, and (2) it has the ability to incorporate latent variables into the analysis and accounts for measurement errors in the estimation process (Hair et al., 1998). In other words, SEM is a statistical technique that establishes measurement models and structural models to address complicated behavioral relationships. SEM is not a new statistical technique (e.g. Jöreskog, 1967 and Jöreskog, 1969); however, its diffusion into the tourism research is relatively recent. For example, Chi and Qu (2008) provided an integrated approach to understanding destination loyalty using SEM. Another study by Gross and Brown (2008) used SEM to examine the relationship between involvement and place attachment in a tourism context. Additionally, He and Song (2009) investigated the mutual relationships among tourists' perceived service quality, value, satisfaction, and intentions to repurchase packaged tour services from travel agents using SEM. Thus these studies adopted the SEM approach because of its ability to address research questions related to complex casual relationships between latent constructs. Hershberger (2003) examined the growth and the development of structural equation modeling from 1994 to 2001. Three conclusions were drawn from his study: (1) SEM has become a pre-eminent multivariate method of data analysis since the number of journals publishing articles using the SEM approach has increased; (2) the total number of SEM articles has also increased; and (3) of all the multivariate techniques, SEM has continued to be the technique that is undergoing the most refinement and extension. SEM can expand the explanatory power and statistical efficiency for model testing with one comprehensive model (Hair et al., 1998). On the other hand, since Pearson (1908) introduced the term multiple regression 100 years ago, this technique has been developed and refined continuously. Commonly used in testing interactions among multiple variables (Evans, 1991), multiple regression is well-recognized for bridging the gap between correlation and analysis of variance in addressing research hypotheses (McNeil, Kelly, & McNeil, 1975). Multiple regression (MR) has become increasingly popular since 1967 (Bashaw & Findley, 1968). Because of its long history, MR has evolved to a sophisticated and versatile tool for various kinds of data analyses, particularly powerful when samples exhibit distinctive characteristics such as censorship, truncation, time series, panel or self-selection and research questions are tailored to address probability related issues. The general model structure involves independent variables and dependent variables, assuming that independent variables cause dependent variables to change and the model error follows a certain known distribution. The model prediction accuracy is usually measured by adjusted R2, which expresses itself as a percentage. The closer the adjusted R2 is to 1, the better the model prediction accuracy is. In tourism research, the linear and probability models of MR are gaining popularity. For example, Uysal and Crompton (1984) used MR to identify factors which exert the most influence on international tourist flows to Turkey; moreover, MR analysis results from Hsu (2000) indicated that respondents' perceptions on “free of crime” and “community amenities and activities” were significant predictors of their support for legalized gaming.

Structural equation modeling is a statistical methodology combining the strength of factor analysis and path analysis. It is carried out by constructing a measurement model and a structural model. The measurement model identifies relations between observed and latent variables. By means of confirmatory factor analysis, the measurement model provides the link between scores on an instrument and the constructs that they are designed to measure. SEM identifies casual relations among the latent variables by specifying that particular latent variables directly or indirectly influence certain other latent variables in the model (Byrne, 2001). In this research paper, the researchers discuss the potential of SEM as a tool to advance the tourism research both statistically and conceptually. By contextually applying and comparing both MR and SEM, the researchers found that the results of hierarchical multiple regression supported H2, that Satisfaction is positively associated with Affective Commitment, and H3, that Trust is positively associated with Affective Commitment. However, H1, that Investment Size is positively associated with Affective Commitment was not supported by MR analysis. It appears that SEM is more straightforward when dealing with both sophisticated relationships and with latent relationships in the empirical model development process. Moreover, the study introduced more statistically significant relationships after the “best-fitting” measurement model. Four additional models were proposed, namely model A, model B, model, C, and model D. The results indicated that model D is the “best-fitting” model. All four of the proposed paths of this model were statistically significant. Specifically, Satisfaction was positively associated with Affective Commitment, Trust was positively associated with Investment Size, Trust was positively associated with Affective Commitment, and Investment Size was positively associated with Affective Commitment. The study of Dixit (2003) that investment size is positively related to trust and the study of Bottazzi et al. (2006) that trust has a significant effect on investment size were both examined in finance contexts. Thus, the findings of both models B and D specifically the significant path coefficients from investment size to trust and from trust to investment size have considerable significance in existing tourism literature. Consequently, the results echo SEM's strength of handling relationships among latent variables which cannot be observed. SEM is most appropriate when the researcher has multiple constructs, each represented by several measured variables, and these constructs are distinguished based on whether they are exogenous or endogenous. One principal difference in SEM is that a construct that acts as an independent variable in one relationship can be the dependent variable in another relationship. This example has supported that SEM is more effective than MR in finding the “best-fitting” model. Although sharing the same mathematic foundation, applications of SEM and MR are highly contextual. Therefore, choice between these two relies on, ultimately, the research question raised and data available. When research questions are raised to address relationships between latent variables in a study, SEM is probably a good choice. However, when censored, truncated, time-series or panel data are involved or research questions are related to probability, MR is likely preferred.