• No results found

CHAPTER 4 RESEARCH METHODOLOGY

4.6 QUANTITATIVE APPROACH USING SURVEY METHOD

4.6.6 Data Analyses

4.6.6.2 Statistical Techniques (Part II)

The main objective of this research was to generate a model of the adoption of innovation that best described the use of BIDSA by ERP user organisations. In order to achieve this main objective, Structural Equation Modelling (SEM) analysis was considered as the best method.

150

Structural Equation Modelling (SEM) has become as an important tool (technique) for data analysis in academic research (Anderson & Gerbing 1988; Breckler 1990; Byrne 2001; Hair et al. 2006; Holmes-Smith, Cunningham & Coote 2006; Jöreskog & Sörbom 1996; Kline 2005). In addition, prior researchers also applied SEM as an integral tool in various research areas (e.g. management, IS) such as studies of behaviour (Homburg & Giering 2001), IT development (Koufteros & Marcoulides 2006), and IT systems (Byrd & Turner 2000;

Etezadi-Amoli & Farhoomand 1996). The models generated by using multivariate technique, particularly SEM are both substantively meaningful and statistically well-fitting (Holmes- Smith, Cunningham & Coote 2006; Jöreskog, K 1993).

When compared to other multivariate techniques, four significant benefits of SEM (Byrne 2001, 2006) are described as:

 SEM takes a confirmatory approach rather than an exploratory approach to the data analysis.

 SEM can provide explicit estimates of error variance parameters.

 SEM procedure can incorporate both unobserved (e.g. latent) and observed variables.  SEM methodology has important features (e.g. modelling multivariate relations) for

estimating point and/or interval indirect effects.

The primary purpose of SEM is to examine the pattern of a series of inter-related dependence relationships simultaneously between a set of latent constructs, each measured by one or more observed variables (Hair et al. 2006; Schumacker & Lomax 1996). Hence, the SEM

technique in this study was used to achieve the objectives as follows:

1) To examine a series of interrelated relationships simultaneously between the analysed dimensions (referred to as non-measurable latent constructs), represented by multiple

151 variables (referred to as measurable manifest variables) or indicators of the latent constructs.

2) To confirm the theoretical relationships in the models between the latent constructs, and the latent constructs and their indicators, as well as to assess their statistical significance.

Thus, it is an appropriate method for use in this study of information system practises (factors affecting the BIDSA adoption).

In this study, SEM was used as “a collection of statistical techniques that allow a set of relationships between one or more independent variables, either continuous or discrete, and one or more dependent variables, either continuous or discrete, to be examined” (Tabachnick & Fidell 2001, p. 653). Hair et al. (2006) mentioned that this technique combines aspects of multiple regression and factor analysis to estimate a series of interrelated dependence

relationships simultaneously. Moreover, SEM integrates other techniques (e.g. recursive path analysis, ANOVA, analysis of covariance) (Holmes-Smith 2000). In addition, SEM is also know as path analysis with latent variables and is currently a regularly used approach for representing dependency relations in multivariate data in social sciences (McDonald & Ringo 2002). In other words, SEM represents a model of relationships among variables using a confirmatory approach to the analysis of a structural theory (Byrne 2006). In addition, this conveys two important aspects of the procedure: 1) the causal processes under study are represented by a series of structural equations; and 2) these structural relations can be modelled pictorially to enable a clearer conceptualisation of the theory under study (Byrne 2006).

152 SEM is based on the assumption of causal relationships where a change in one variable (X1) is supposed to result in a change in another variable (Y1), in which (Y1) affects (X1)

(Shammount 2008). Not only does SEM aim to analyse latent constructs, in particularly the analysis of causal links between latent constructs, but also it is efficient for other types of analyses including estimating variance and covariance, test hypotheses, conventional linear regression, and confirmatory factor analysis (Jöreskog & Sörbom 1996). According to Anderson & Gerbing (1988), SEM is a confirmatory method that could provide a

comprehensive means for assessing and modifying theoretical models. SEM can generate a statistical test of the goodness-of-fit for the confirmatory factor solution using confirmatory factor analysis (CFA) (Kline 2005).

Arbuckle’s (2005) structural equation modelling software AMOS15 version 7.0 (Analysis of Moment Structures) was used to explore statistical relationships among the items of each factor and between the factors of independent (e.g. benefit, complexity, compatibility, top management support, organisational readiness, absorptive capacity, internal need, competitive intensity, and vendor selection) and dependent variables (e.g. the adoption of BIDSA). AMOS 7.0 computing program (Arbuckle 2005; Arbuckle & Wothe 1999) linked to SPSS was used to conduct SEM analysis. As a result, it becomes the most appropriate widely and easily used package. AMOS can fit multiple models into single analysis. Thus, the study can specify, estimate, assess, and present the appropriate model in a causal path diagram to show hypothesised relationships among variables.

However, as the SEM technique and other statistical methods are alike, some assumptions need to be met before conducting SEM. For instance, the sample size plays an important role

15

AMOS is an acronym for “Analysis of Moment Structures” or the analysis of mean and covariance structures. AMOS computes parameter estimates so that the resulting implied moments are closet in terms of discrepancy function to the sample moments (Arbuckle 2005).

153 in the estimation and interpretation of SEM results (Hair et al. 1995). Some authors stated that sample sizes as small as 50 could provide valid results (Anderson & Gerbing 1984; Hair et al. 1998). Hair et al. (2006) argued that there is no correct sample size and suggested that sample sizes in the range of 150-400 are recommended. Boomsma (1983) suggested that the estimation of SEM by using maximum likelihood methods can be used when the sample size was at least 200. Recommended minimum samples of 100-150 could ensure stable maximum likelihood estimation (MLE) solution (Hair et al. 2006). However, the sample size of this thesis is 150, which is considered appropriate for applying the SEM technique.

In order to perform SEM, a two-stage approach is recommended by (Anderson & Gerbing (1988) rather than a single-stage approach. By using this two-stage approach, Kline (2005) stated that the typical problem of not being able to localise the source of poor model fit associated with the single-stage approach is overcome. According to Hair et al. (1998), to avoid any interaction between the measurement and structural models, the two-sage approach offers on accurate representation of the interaction of the reliability of the items of each construct.

Thus, in this thesis, the two-stage approach was adopted to conduct the analysis. That is analysing the causal relationships in the structural model requires performing the

measurement model first, due to the latter representing a condition that must be satisfied as a matter of logical necessity (Anderson & Gerbing 1988). The two-stage structural model used in this thesis comprises of 1) measurement model (e.g. assessing unidimensionality and examining reliability and validity); and 2) structural model (e.g. testing hypotheses).

154 The first stage of analysis was conducted by specifying the causal relationships between the observed variables (items) and the underlying theoretical constructs (composite and latent variables). At this stage, this was to verify the unidimensionality of the composite and latent constructs. Unidimensionality has been defined as “an assumption underlying the calculation of reliability and is demonstrated when the indicators of the construct have acceptable fit on a single-factor (one-dimensional) model” (Hair et al. 1998).

However, Anderson & Gerbing (1988) argued that unidimensional measurement models are more generally useful because of models offering more precise tests of the convergent and discriminant validity of factor measurement. Thus, the purpose of this stage is to ensure that a set of items empirically measures a single dimension. In accordance with Anderson &

Gerbing (1988), Dunn, Seaker & Waller (1994), and Hair et al. (1998), unidimensionalilty assessment was conducted prior to testing the reliability and validity of each construct.

In assessing unidimensionality, confirmatory factor analysis (CFA) is a better method for use in this research where hypotheses about the grounded theoretical models exist (Bollen 1989), as is the case in the thesis. Kline (2005) also suggested that the factor structure identified in CFA turns out to have best fit when evaluated with CFA. CFA is considered a more powerful (Anderson & Gerbing 1988; Hair et al. 2006) and more flexible (Dunn, Seaker & Waller 1994) technique than others in term of assessment.

Therefore, CFA was used in this thesis. The underlying constructs of relational links (e.g. technology, organisation, and environment characteristics) have already been demonstrated empirically to be valid in the literature. This was to determine whether the number of factors and loadings of measured indicators (items) had conformed to what was expected, based on

155 re-established research and theory. Items that loaded weakly on the hypothesised factors were removed from the scale, thus resulting in a unidimensional scale (Dunn, Seaker & Waller 1994). A factor loading of 0.50 and above on a specified factor has been considered

acceptable (Hair et al. 2006). Thus, this standardisation is used as the cut off value within this thesis.

Once the step of undimensionality of constructs is achieved, reliability and validity of these constructs is demonstrated in the following step (see Chapter 6 for further discussion of reliability and validity). For this purpose, CFA using maximum likelihood estimate was performed (Anderson & Gerbing 1988; Kline 2005). Then, the paths or causal relationships between the underlying theoretical latent constructs were specified in the structure model (the second stage). Further details about these two stages are discussed in the following chapter.

The results and also the assessment of data used in the SEM analysis of the relationship between the independent variable and dependent variables are presented in Chapter 5 where discussed are the assessment of normality, outlier, and multicolinearity. Then, Chapter 6 provides the analysis of structural models.