Chapter 5. Research Methodology
5.7. Data analysis
5.7.1. Quantitative data analysis
As stated above, this study used Smart PLS to examine the research model. PLS is an alternative approach to traditional structural equation modelling, and has been used widely in information systems research. It is a method that is designed to maximize prediction rather than fit (Fornell & Bookstein, 1982). PLS was used to study the validity of the model‟s main components and to determine the relationships among the constructs of the proposed model.
According to Fornell and Bookstein (1982) and Hulland (1999), PLS is usually used as an analytical tool in two sequential steps; these are assessment of the validity of the measurement model and then assessment of the structural model. This two-step
approach was adopted for the testing of the proposed model. In this approach, the fit and construct validity of the proposed measurement model are tested first. Once a
satisfactory measurement model is obtained, the measurement model is “fixed” when the structural model is estimated. Through this technique overall reliability and validity are usually assured (Hair, Anderson, Tatham, & Black, 2006).
5.7.1.1. Measurement (outer) model assessment
The measurement model was assessed in terms of convergent validity and discriminant validity. Convergent validity is shown when each measurement item correlates strongly with its proposed construct (Gefen & Straub, 2005). Convergent validity was assessed using item loadings and their significance, composite reliability, Cronbach alpha, and average variance extracted (AVE). When the criteria established to assess convergent
validity are met then the items are said to be convergent on the proposed latent construct. These criteria are shown in Table 5.13.
Table 5-13 Criteria used for convergent validity
Convergent validity criteria Guideline Source
Item loadings >=0.70 Hulland (1999)
Composite reliability >=0.70 Hair et al. (1995) Average variance extracted >=0.50 Hair et al. (1995) Cronbach alpha coefficient >=0.70 Gefen & Straub (2005) t-value of outer loading >=1.96 Gefen & Straub (2005)
Measurement items which did not load satisfactorily on their constructs (>=0.7), were dropped from the model. The t-values of the outer loadings were also tested to ensure that each item loaded significantly on its latent variable.
Composite reliability was used to assess the internal consistency of the measurement model. Composite reliability is a general measure of reliability that uses the item loadings estimated within the model. Composite reliability should be at least 0.7 to be accepted (Hair et al., 1995).
Cronbach alpha is used to measure the inter-correlation among items in a group indicating to what level the items are measuring a single latent variable. In PLS composite reliability is often used instead of Cronbach alpha when validating the measurement model. Both were included in this analysis. Cronbach alpha can be
interpreted similarly to composite reliability and values of at least 0.7 are considered acceptable (Hair et al., 1995).
AVE reflects the overall amount of variance in the indicators accounted for by the latent construct. AVE should be more than 0.5 to be considered acceptable (Hair et al., 1995).
The measurement model was also assessed in terms of discriminant validity.
Discriminant validity validates that each measurement item correlates weakly with all other constructs except for the one with which it is proposed to be associated.
Discriminant validity in PLS is tested by comparing AVE and inter-construct correlations. This is done in two steps:
Comparing item cross loadings to construct correlations;
Examining the ratio of the square root of the AVE of each construct to the correlations of this construct with all other constructs (Gefen & Straub, 2005).
For satisfactory discriminant validity each item should load more highly on its own construct than on other constructs. In addition, the average variance shared between a construct and its measures should be greater than the variance shared by the construct and any other constructs in the model (Gefen & Straub, 2005).
5.7.1.2. Structural (inner) model assessment
The structural model is tested to evaluate interrelationships of the constructs. In this study the structural model was evaluated on two criteria:
The ability to explain variance in the dependent variables; The significance of the path coefficients.
An estimate of the variance explained the dependent variables is provided by the
squared multiple correlations (R2) of the structural equations of these variables. R2 was used as an estimate of how much of the variability of a dependent variable is explained by the independent variables (Hair et al., 1995).
For the second evaluation criteria, the structural model was evaluated on whether it reflects valid interrelationships by testing the t-values of the proposed relationships (Hair et al., 1995). Smart PLS provides path coefficients that indicate the strength of the relationship between two constructs. The bootstrap procedures with 500 re-sample were used to calculate the significance of these path coefficients. In addition to the significance of the path coefficients, the strengths of the relationships they represent were also of interest. In this study correlations of less than 0.2 were considered weak, correlations between 0.2 to 0.5 were considered to be moderate, and correlations of more than 0.5 were considered to be strong (Cohen, 1988).
5.7.2. Qualitative data analysis
The qualitative data analysis in this study depended on interview data analysis. Interview analysis is the mechanism of processing raw interview data, most likely recorded voice or another format of captured interview data, to produce evidence based interpretations that can be represented in a standard academic report (Silverman, 1993).
It is not only the process of collecting, coding, sorting, and sifting but it also covers the process of noticing, categorizing, contrasting, weighing, and merging results to develop meaning and implications of patterns (Seidel, 1998). Interviews conducted in this study yielded digital voice recorded data. A qualitative data analysis took place which
involved abstracting the data that was related to the main variables and themes of the research. The qualitative analysis in this study adopted the technique described by Dey (1993) and Silverman (1993). This qualitative data was analysed to compliment the quantitative findings using the procedures described below.
The analysis started by examining all the interview transcripts to identify concepts and themes associated with the constructs from the proposed model. Since the interviews were semi-structured and the interview guide was built around the research main constructs, it was easy to identify themes related to the constructs. Synonyms for constructs were identified and used in theme extraction. For example, terms such as “risk threats” and “risk concerns” were used to identify themes around e-privacy risk concerns. This process was carried out to identify all concepts and themes in all transcribed responses and was undertaken using Microsoft Word capabilities such as word searching, using Word‟s search facility and highlighting text in multiple colours.
Data then was categorized according to the identified themes. Because the interview questions were structured directly around the research constructs and their relationships as proposed in the model, it was relatively straightforward to classify responses related to identified themes against each research construct as shown in Appendix D. Microsoft Word was used to present this classification in table format.
Portions of the responses that were not related to the research constructs were classified as either participants‟ background data or general data. The background data about the participants was compiled to describe the backgrounds of the participants in the
interviews. The general data that was not specifically related to the research model was retained for later further analysis.
The fourth step in the process was to further analyse the data relating to the research model. The responses relating to each construct (see Appendix D) were synthesised as shown in Appendix E and Appendix F. These synthesised findings were used to reflect the participants‟ views on the research constructs and their relationships. This step enabled the researcher to map responses together within one consolidated table and examine them in-depth to identify where there was consensus and where opinions varied. Conclusions were drawn based on this.
The final step in the process was to utilize the analysis from previous steps (see Appendix E and Appendix F) to link the synthesised responses to each research hypothesis. Appendix G shows each proposed hypothesis and the support (or lack of support) for it from the interview findings.
5.8.
Summary
This chapter described the research methodology used in the study. This research was conducted using quantitative and qualitative approaches. The quantitative approach was carried out via questionnaire and the qualitative approach was undertaken through semi- structured interviews. The chapter explained the reasons for the selection of these methods and provides details of how the participants were selected. The chapter also explained how the questionnaire and interview questions were designed and the data
analysis approaches used. The next chapter presents the results of the quantitative testing of the research model.