STATISTICAL ANALYSIS

The selection of data analysis techniques depends on the type of research questions that the study aims to answer. In this study, all data collected from the measurement instruments was analysed by a number of different quantitative techniques. These techniques included item analyses and PLS/SEM. The objective of data analysis is to test the structural model. A short explanation of the different quantitative techniques and programmes used in this study is provided below.

3.9.2 Computer package

Item analysis and PLS analysis methods were used to analyse the collected data. Statistica version 12 was used to perform the item analyses, which provided the reliabilities of the items and constructs. SmartPLS version 3 (M. Kidd, personal communication, 12 February 2016; Ringle, Wende & Becker, 2014) was used to test the relationships between the different variables, which aimed to provide the path coefficients between the variables and to estimate the PLS model.

3.9.3 Item analysis

A variety of scales can be used to test the latent variables. By using item analysis, the understanding of the validity and reliability of tests can be increased. A close examination of individual tests is critical when attempting to understand why some tests show specific levels of reliability and validity and others not (Tabachnick & Fidell, 2013).

Each item of a measurement instrument measures a specific aspect of an individual. Consequently, it is necessary that each measurement instrument include items that measure the actual latent variable or the dimensions of the latent variable that are supposed to be measured. Each variable carries a specific constitutive definition, and each item that is used to measure a specific variable must be in line with the constitutive definition of the variable. The items in each instrument have been developed to indicate the participants’ standing on the specific latent variable and act as stimuli by aiming to elicit the participants’ responses in terms of the behaviour of the underlying constructs. The item responses therefore record the behaviour that

underlies the construct and consequently make the behaviour ‘observable’ in the form of data (Little, 2013).

However, items can be poor at eliciting a response when they are insensitive, inconsistent or portray a poor interpretation of the construct (Theron, 2014). A process called item analysis can be used to identify poor items through item statistics by determining the quality and internal consistency reliability of the items of the respective scales. The literature suggests that reliabilities (Cronbach’s alphas) of .70 or higher are sufficient (Little, 2013).

Depending on the results of the item analysis and the nature of the poor items (if such items are present), a decision should be made whether to transform or delete the items from the instrument or respective scale(s) (Theron, 2014). If the overall reliability of an instrument or subscale shows significant improvement after the selected items have been deleted, they are excluded from subsequent analyses. Cronbach’s alphas and average inter-item correlations for each total scale were used for this purpose.

3.9.4 Partial least squares structural equation modelling analysis

The researcher made use of PLS modelling, which is a soft modelling approach and utilises partial least squares, in contrast to the hard modelling approach of SEM, which uses maximum likelihood (Monecke & Leisch, 2012). The motivation for using PLS modelling is its exploration and prediction value, as PLS path modelling is recommended at an early stage of theoretical development involving testing and validating exploratory models. PLS path modelling has another advantage in that it is suitable for prediction-orientated research. Consequently, PLS modelling can assist researchers in focusing on the explanation of endogenous constructs (Monecke & Leisch, 2012). Moreover, seeing that the PLS approach is distribution free, the data is not required to be normally distributed (Chin, 1998). This method could therefore easily accommodate both reflective and formative scales.

PLS models include two sets of linear equations, namely the outer model and the inner model. The outer model includes an analysis of the relationships between latent variables and their observed or manifest variables, whereas the inner model is aimed at analysing the relationships between unobserved or latent variables (Hair, Ringle &

Sarstedt, 2011). The outer model in PLS can be compared to the measurement model used in SEM, while the inner model in PLS can be compared to the structural model used in SEM.

However, before commencing with the PLS model estimation, a series of analyses needed to take place (Hair et al., 2011). Firstly, the reliability of the latent variables was evaluated. This was done by looking at the composite reliabilities, average variance extracted (AVE) and R-squares. If the coefficients exceeded .70, they were regarded as satisfactory (Hair et al., 2011). After the systematic evaluation of the reliabilities of the latent variables, the PLS estimates revealed the reliability and validity of the measurement model (i.e. the outer model) according to certain criteria associated with the measurement model.

Secondly, the structural model (i.e. the inner model) estimates needed to be evaluated once the calculated latent variable scores showed evidence of sufficient reliability and validity (Chin, 1998). The structural model (i.e. the inner model) relates latent variables to each other. In order to assess the significance of main effects and interaction effects, a bootstrapping sampling procedure was performed. After bootstrapping, the accuracy of the path estimates to the true effects was assessed. Thirdly, moderating effects were analysed by using PLS path modelling. This process includes two steps: the iterative process, which is characterised by latent variable scores estimated for each latent variable, which are then entered as dependent and independent variables into one or more regressions, and the testing of moderating effects in multiple regression through PLS path modelling. When the researcher mentions moderating effects, it should be read in the context of PLS path modelling (i.e. the moderating relationships within the structural model). The researcher was interested in the moderating effect of latent variables on the direct relationships between latent variables (Hair et al., 2011).