In this section, confirmatory factor analysis was performed to test whether the data fits the hypothesised measurement model whether the measures of a latent variable are consistent with the nature of observed variable. The SPSS® AMOS (Arbuckle, 2013) was used to calculate the formation of the causal relationship among the concepts that comprise the hypothetical model, and to analyse the level of influence among the causal relationships. This study confirms the structural model by verifying its appropriateness from the results of the covariance structural analysis. The maximum likelihood method is used for calculating the covariance matrices between two variables in a measurement model. Several fit indices are produced and moderated. AMOS allows for the use of modification indices to generate the expected model fit, when adequate fit was not achieved. Two methods could have been used to moderate the fit. The first method involved deleting the path that showed a low causal relationship, and the second method involved an additional causal relationship. The second method was chosen by establishing an additional causal relationship to the measurement model.
Generally, it is recognised that to support the model fit, a consensus among the following fit indices is sought and compared with threshold suggested by (Hair et al., 2010). Absolute fit indices measure how well the model is specified by the observed data. That includes χ2:df ratios on the order of 3:1, SRMR is below 0.09, and RMSEA is below 0.08. Incremental fit indices measure how well the estimated model fits relative to some alternative baseline model. A commonly-accepted rule of thumb for the incremental fit indices of CFI, IFI, NFI, and NNFI is to be above 0.9.
The order factor consists of items resulting from exploratory factor analysis that are loaded above 0.6 and are extracted from constructs within related factors. To assess whether all 1st order constructs reflected the 2nd order factors, the 2nd order confirmatory factor analysis was conducted for five 2nd order factors by using extracted 1st order constructs. The results in Table 4-5 indicate that all the higher order measurement models have an acceptable fit. In this research, a measurement model is developed with five 2nd order factors and one 1st order.
84
Table 4-5: Fit indices of measurement model
GOF measure χ 2 df χ2/df SR MS RM SEA Lower bound Upper
bound CFI IFI NFI NNFI
Threshold < 3 < 0.09 < 0.08 > 0.90 > 0.90 > 0.90 > 0.90 External change 65.624 40 1.641 0.051 0.056 0.030 0.080 0.965 0.965 0.916 0.951 Dynamic capabilities 208.706 92 2.269 0.047 0.079 0.065 0.099 0.949 0.950 0.913 0.934 ICT adoption 192.333 104 1.849 0.087 0.065 0.050 0.079 0.965 0.966 0.928 0.954 VE 321.518 204 1.576 0.051 0.053 0.042 0.064 0.961 0.962 0.902 0.952 Agility in the supply chain 219.880 125 1.759 0.043 0.061 0.047 0.074 0.970 0.970 0.934 0.954
Reliability and validity tests are conducted to measure the consistency and accuracy of the measurement models for confirmatory analysis. Composite reliability is a measure of the internal consistency of the latent variables and the suggested threshold is above 0.7. The accuracy of the actual measuring variable was estimated via construct validity test. Two kinds of construct validity tests were executed, including convergent validity and discriminant validity. To assess which indicators of a specific construct “converge” or share a high proportion of variance in common, the average variance extracted (AVE) (Fornell and Larcker, 1981) measure was utilised. Convergent validity was assessed by the AVE above 0.5 and composite reliability greater then AVE. Also as suggested by Hair et al. (2010), factors loading should be statistically significant and estimated 0.5 or higher, and ideally 0.7 or higher. Discriminant validity was assessed through identifying which construct is truly distinct from any other constructs. Discriminant validity is supported when the AVE is higher than the squared correlation between two constructs (Fornell and Larcker, 1981). The squared correlation was represented by the Average Shared Variance (ASV). Another measure is the square root of AVE value belonging to each construct needs to be higher than any correlation among any pair of constructs.
Table 4-6 summarises reliability and validity analysis for all constructs. Hence higher order factors contain more abstract level, 1st order constructs were tested as building block for the measurement model. The values of composite reliability for all constructs are above 0.7, except construct of the changes in customer requirements. Although composite reliability value of changes in customer requirements constructs is 0.685 (slightly below the threshold), this is within in accepted range (Hair et al., 2010). Reliability between 0.6 and 0.7 may be acceptable, provided that other indicators of a model’s construct validity are good, the result represents sufficient reliability for all the constructs. However, some AVE values are slightly below 0.5, all the composite reliability values are greater than the AVE. The result indicates that the measurement model has satisfactory convergent validity. The ASV values for all constructs are
85
lower than the AVE values in the same constructs. Also, square root of AVE is greater than the correlation among pair of other latent variables scores, with respect to its corresponding row and column value. The result indicates no construct shares more variance with another construct than with its own indicators, thus exhibiting sufficient levels of the discriminant validity.
The multi-colinearity occurs when any single independent variable is highly correlates with a set of other variables (Hair et al., 2010). Thus, the multi-colinearity test was conducted, because several inter-construct correlations in Table 3.5 were higher than the benchmark value of 0.60. The rule of thumb to judge the existence of multi-colinearity is if variance inflation factors (VIFs) are greater than 10 or if tolerance values are less than 0.10 (Kline, 2011). This study showed that the highest VIF was 2.480 and the lowest tolerance value was 0.403. Thus, multicollinearity did not appear to be a significant problem in this dataset.