Quantitative Data Analysis using SEM

3.5 Quantitative Data Analysis using SEM

This study applied the SEM approach for data analysis in the quantitative phase, which followed the analysis process outlined by influential authors of SEM theory and application, such as Barclay, Thompson and Higgins (1995), Chin (1998a) and Gefen, Straub and Boudreau (2000). Use of this approach was based on the following reasons. Firstly, SEM is widely known as a powerful second-generation multivariate analysis technique to test the proposed relations among the variables in a model (Fornell 1982). Secondly, SEM is superior to first generation methods, such as traditional regression and factor analysis, due to the fact that the measurement model is assessed within the context of the theoretical structural model (Fornell 1982).

Thirdly, SEM is more suitable for the mathematical modelling of complex processes to serve both theory and practice (Gefen, Straub and Boudreau 2000).

In addition, use of the Partial Least Square (PLS) was based on the following reasons. Firstly, compared with the covariance-based technique that only deals with reflective observed variables, PLS deals with both formative and reflective observed variables (Hulland, Chow and Lam 1996; Chin 1998a). Formative observed variables can be referred to as indicators that cause the latent variables and the construct to be a function of formative measures (Thompson, Barclay and Higgins 1995).

Meanwhile, reflective observed variables reflect the latent variables as representational of the construct and should be unidimensional and correlated (Gerbing and Anderson 1988). Secondly, PLS is most suitable since the objective of the research is theory building. The primary research objective of this study is the explanation of the model variance for some constructs and theory building.

Furthermore, PLS does not aim to fit a theorized model to the data, but instead examines the strength of both direct and indirect relationships among constructs.

PLS, therefore, is a more suitable technique for this study since it is suitable for assessing the strength of the relationships among the constructs in the model, and not the overall fit of the data with the model (Gefen, Straub and Boudreau 2000).

Thirdly, the reliability and validity of the measures of the theoretical constructs in

58 PLS can be assessed simultaneously, and the relationships among these constructs can be predicted (Barclay, Higgins and Thompson 1995). In particular, the PLS technique can assess all dependent variables and map them on paths to analyse them simultaneously (Gefen, Straub and Boudreau 2000). Fourthly, PLS is also more suitable for application on observable measurement variables (items) that are not well established and are used within a new measurement context (Barclay, Higgins and Thompson 1995). Fifthly, PLS deals with a small sample size and non-normal conditions of latent constructs; which makes it more popular among researchers in recent years (Compeau and Higgins 1995; Chin 1998a).

For the above reasons, the PLS technique was considered to be suitable for analysing the data of the current study. Data from the national survey were analysed using PLS-Graph version 3.0 (www.plsgraph.com), which was developed by Chin (2003).

Two sequential stages are involved in the data analysis using PLS, namely assessment of the measurement model and assessment of the structural model (Barclay, Higgins and Thompson 1995; Hair et al. 1998). Figure 3.2 illustrates the two-step PLS analysis approach.

Figure 3.2 The Two-step PLS Analysis Approach

As can be seen from the figure above, the first step was used for measuring the validity and reliability of the instruments of the study, while the second was used for analysing the data of the study. A more detailed description of these two models is presented below.

3.5.1 Assessment of the Measurement Model

The measurement model focuses on the relationships among the observed variables and constructs in this study: effort expectancy, social influence, facilitating

59 conditions, privacy concerns, trustworthiness, outcome expectancy, motivation, intention and e-Service usage (Igbaria, Guimaraes and Davis 1995). There are two steps to assess the measurement model, namely: (a) convergent validity; and (b) discriminant validity. Convergent validity evaluates the degree to which items of the constructs are really related to the constructs. The convergent validity can be assessed by individual item reliability and internal consistency. Meanwhile, discriminant validity evaluates the degree to which constructs differ from each other.

The measurement model followed the PLS procedure, and can be assessed by examining the item reliability, internal consistency and discriminant validity, as illustrated in Figure 3.2 (Barclay, Higgins and Thompson 1995; Quaddus 2004;

Santosa, Wei and Chan 2005). The following section will discuss, in detail, the procedure of the assessment of the measurement model.

3.5.1.1 Item Reliability of the Questionnaire

Previous researchers have assessed item reliability by examining loadings, or simple correlations, of the measures with their respective constructs. In PLS, item reliability can be assessed by evaluating: (1) the loading score for reflective items, or (2) weight score for the formative items.

Researchers have different opinions on the assessment of an item loading’s strength.

According to Carmines and Zeller (1979), a rule of thumb is to accept items with loadings equal to or greater than 0.707, which therefore explains at least 50% of the variance in a construct (Nunnally 1978). Hair et al. (1998) recommend three types of significance levels for item loadings; (i) item loadings greater than 0.3 are considered significant, (ii) item loadings greater than 0.4 are considered more significant, and (iii) loading in excess of 0.5 are considered very significant. In addition, Igbaria, Guimaraes and Davis (1995) suggested that an item loading equal to or greater than 0.4 was an acceptable reliability limit. The results of the item reliability analysis can be seen in Chapter 6.

3.5.1.2 Internal Consistency

Internal consistency was used to establish the convergent validity and to assure that there were correlations among the items for a construct. Many quantitative

60 researchers have used Cronbach’s alpha as a measure of internal consistency. Fornell and Larcker (1981) suggested two types of measurement for evaluating internal consistency, namely: composite reliability (CR) and average variance extracted (AVE). These two types will be discussed separately.

Firstly, CR is used to measure internal consistency. The value of CR can be calculated using the following formula (Fornell and Larcker 1981; Barclay, Higgins and Thompson 1995; Chin 1998b).

(Σλ

)

(Σλ

)

+ Σ

Var( ε

)

where

λ

ε

λ

CR is considered superior to Cronbach’s alpha since it practices the item loading estimation within the casual model (Fornell and Larcker 1981). Nunally and Bernstein (1994) suggested that an alpha of 0.7 indicates applicable internal consistency and could be set as a benchmark to establish convergent validity of the measurement model (Barclay, Higgins and Thompson 1995).

Secondly, AVE was used to measure the reliability coefficients of each construct.

This measurement reflects the amount of variance of the construct explained by its items (Fornell and Larcker 1981). The following formula was used to calculate AVE.

Σλ

Σλ

+ Σ

Var( ε

)

where

λ

ε

λ

Although AVE is not a measure of convergent validity, Fornell and Larcker (1981) suggested that AVE should be at least 0.5 for a construct to achieve adequate reliability. The results of the internal consistency analysis can be seen in Chapter 6.

61 3.5.1.3 Discriminant Validity

Discriminant validity was used to assess the degree of variance among items and constructs in the model (Barclay, Higgins and Thompson 1995). Discriminant validity assessment is essential to assure that there are no overlaps among different constructs. The work of Fornell and Larcker (1981) suggests that discriminant validity is considered adequate when AVE for one’s construct is greater than constructs shared variance.

According to Barclay, Higgins and Thompson (1995), discriminant validity can be established by comparing the square root of the AVE to the inter-construct correlations. For each construct, the square root of the AVE for that construct should be larger than the variance shared between one construct and another in the model.

Moreover, the diagonal value (the square root of the AVE) should be greater than the off-diagonal value (the correlation between constructs in the corresponding column and rows) in the correlation matrix (Hulland 1999) as illustrated in Table 6.18 in Chapter 6.

3.5.2 Assessment of the Structural Model

Structural model assessment was conducted in order to analyse the relationships among the constructs as hypothesised in the final research model. To establish the structural model assessment, the amount of variance was explained and the statistical significance was tested, based on three components of the information, namely: (i) amount of variance or R square (R²); (ii) path coefficient (β); and (iii) statistical significance of t-value, as illustrated in Figure 3.2. In PLS, the structural model was assessed by using the bootstrapping method. This method was used to calculate the statistical significance of the t-values and path coefficients as well as R square values. The bootstrapping procedure is a non-parametric approach (Chin 1998b). The reason for using a non-parametric approach was due to the fact that the data are not assumed to be multivariately normal in PLS (Barclay, Higgins and Thompson 1995;

Hair, Ringle and Sarstedt 2010).

62 3.5.2.1 Amount of Variance Explained

Amount of variance explained (R²) determines the explanatory power of the components of the model by indicating the amount of variance in the construct, which is explained by its corresponding independent constructs. In the PLS analysis, interpreting the values of R² is the same as that of the R² value produced by multiple regression analysis (Barclay, Higgins and Thompson 1995). The R² value is derived from the bootstrapping procedure, as mentioned above. Thus, the predictive power of the research model can be assessed by obtaining the R² values (Barclay, Higgins and Thompson 1995; Santosa, Wei and Chan 2005).

3.5.2.2 Path Coefficient and Statistical Significance

The next test was to evaluate the relationships among the constructs, as predicted in the hypotheses based on the final research model. More specifically, it was a statistical analysis conducted by evaluating the path coefficient (β) and t-value. The β and t-value were produced from the bootstrapping procedure of the PLS.