SEM

Structural equation modeling (SEM) is a statistical approach for examining the

causal relationships and testing the hypotheses between the observed and latent

variables in a research model (Hoyle, 1995). In this study, we propose an extended

version of TAM based on the related literature in order to examine an online learning

community research model. Thus, we use SEM to analyze the data by two procedures,

as shown in Figure 4.1.

Theoretical Development

Model Specification

Model Identification

Sampling and Measurement

Parameter Estimation

Assessment of Fit Model Modification

Discussion and Conclusion Procedure 1.

Model Development

Procedure 2.

Estimation and Evaluation

Figure 4.1 The basic procedures of SEM analysis

Procedure 1. Model Development

In the Model Development stage, we construct a hypothesized model and analyze it

with SEM.

- Step 1. Theoretical Development

Because the SEM model is based on theories, we must consider the development

of related theories, the induction of our research hypothesis, as well as a process of

the theoretical justification and interpretation to propose a hypothesized model.

- Step 2. Model Specification

Model Specification is the most specific step in Procedure 1. The purpose is to

develop specific variables from theories by using SEM to examine and estimate the

parameters.

- Step 3. Model Identification

When developing a model, researchers must clearly identify two types of variables,

namely, exogenous and endogenous variables. Exogenous variables play the role of

independent variables, whereas endogenous variables play the role of dependent

variables. This means that the endogenous variables are predicted by the exogenous

variables. We list the variables below (see Table 4.1).

Table 4.1 The independent variables and dependent variables in our model Independent variable Dependent variable

H1 Online Course Design Perceived Usefulness H2 Online Course Design Perceived Ease of Use H3 Online Course Design Perceived Interaction H4 User Interface Design Perceived Ease of Use H5 User Interface Design Perceived Interaction H6 Previous Online Learning

Intention to Use an Online Learning Community

H9 Perceived Ease of Use Perceived Usefulness H10 Perceived Ease of Use Perceived Interaction

H11 Perceived Usefulness Intention to Use an Online Learning Community

H12 Previous Online Learning Experience

Intention to Use an Online Learning Community

H13 Perceived Interaction Intention to Use an Online Learning Community

Procedure 2. Estimation and Evaluation

After developing the SEM model, researchers must collect data to measure the

model and determine whether the observed data matches the model.

- Step 4. Sampling and Measurement

This stage begins with the collection of samples and measurements. After

processing the observed data, we follow SEM analysis methods to further estimate a

series of parameters. We also use statistical software, such as SPSS and LISREL, to

evaluate the reliability, validity, and correlation coefficient matrix, and test if the

hypotheses between the variables are supported.

- Step 5. Parameter Estimation

Because maximum likelihood estimation is set as default in LISREL software, we

adopt this widely used method to estimate the parameters.

- Step 6. Assessment of Fit

As criteria for the model’s evaluation, we adopted the following indices

recommended by Hoyle & Panter (1995):

(1)χ²/d.f; (2) Goodness-of-fit index (GFI); (3) Adjusted GFI (AGFI); (4) Normed fit

index (NNFI); (5) Non-normed fit index (NNFI); (6) Relative fit index (RFI); (7)

Incremental fix index (IFI); (8) Root mean square residual (RMR); (9) Root mean

square error of approximation (RMSEA); and (10) Critical N.

- Step 7. Model Modification

When the model is tested by SEM, if the results are rejected by the data, i.e., the

model is not a good fit, it is important to find the problematic causal relationships and

improve the model. If the model requires modification, we need to return to step 2 for

model respecification. We also made some modifications so that the entire model

presents a good fit and strong stability.

- Setp 8. Discussion and Conclusion

Based on the results of data analysis, we validate the proposed research model and

hypotheses. Finally, we identify the phenomena that derive from the causal

relationships in practice, and interpret their implications in the real world.

Overall, when we want to examine a research model, it is appropriate to use the

SEM statistical method, which combines factor analysis and path analysis, to test the

model’s fit. Numerous TAM related empirical studies have adopted SEM to validate

research model and hypotheses (e.g. Adams et al., 1992; Arbaugh, 2002; Arbaugh &

Duray, 2002; Gao, 2005; Igbaria, Guimaraes, & Davis, 1995; Landry, Griffeth, &

Hartman, 2006; Lee, Cheung, & Chen, 2005; Liaw, 2007; Liu, Chen, & Sun, 2006;

Ngai, Poon, & Chan, 2007; Ong, Lai, & Wang, 2004; Pan et al., 2005; Pituch & Lee,

2006; Raaij & Schepers, 2006; Selim, 2003; Straub, keil, & Brenner, 1997; Venkatesh,

2001; Yi & Hwang, 2003).

The main advantage of SEM is that it can estimate a measurement and structure

model, and achieve a good model fit after analysis and modification (Ngai, Poon, &

Chan, 2007). In addition, SEM integrates factor analysis, principle components

analysis, discriminant analysis, path analysis, and multiple regression from

first-generation techniques as a comprehensive statistical approach. SEM also

provides multiple criteria to measure a model’s quality and estimate measurement

errors.

To test the model of this research, SEM and LISREL 8.54 (Joreskog & Sorbom,

1993) software was used for validation. We adopt the maximum likelihood method to

estimate the model’s parameters. For the sample size, Boomsma (1987), suggested

that if the maximum likelihood method is used to estimate the parameters, the

smallest sample size should be higher than 200. However, he indicated that the sample

size would have to be smaller than 100 to actually generate incorrect results and

inferences. Thus, the sample of 436 students selected for this research was sufficient.