Understanding consumer online shopping behaviour from the perspective of transaction costs

4.9.6 Model Modification 
Another aspect of the SEM techniques is the need to recognize that an initial specified model 
may not provide a good fit of the data; re-specification and re-estimation of the model is often 
necessary to achieve a better fit of the model (Anderson and Gerbing 1988). Based on the 
two-stage model-building approach developed by Anderson and Gerbing (1988), this study 
adjusted and fitted the measurement model to the data (through CFA) prior to testing and 
modifying the structural component of the model. 
As suggested by Baumgartner and Homburg (1996), if an 
a priori 
measurement model 
provides a poor fit to the data, improvements to the model fit can be made by utilizing 
modification indices provided by the SEM software program. Modification indices suggest 
ways that the model might be altered by allowing the error covariance of corresponding 
parameters to become free or by allowing indicators to load more than one factor 
(multidimensional factors). A better fitting measurement model might then result (Bollen 
1989, Byrne 2009). However, because of reasons related to the study
’
s theoretical 
justification, no modifications were made that allowed indicators to load on multiple factors. 

---

203
Importantly, according to Kline (2011), if the modified measurement model is plausible 
(good fit), then the following patterns should also be seen by the researcher: (i) convergent 
validity is demonstrated by all indicators specified to measure a common underlying latent 
factor having relatively high standardized loadings on that factor, and (ii) discriminant 
validity is demonstrated by estimated correlations between latent factors not being 
excessively high (e.g., > 0.85). 
Importantly, to assess how well the latent constructs are measured by their indicators in CFA, 
SEM scholars (Anderson and Gerbing 1988, Hair
 et al.
2006) urge the researcher to report 
the composite reliability (CR) and the average variance extracted (AVE). The CR measures 
the internal consistency reliability of a summated scale (> 0.7 indicating internal reliability), 
and the AVE measures the amount of variance captured by a construct in relation to the 
variance due to measurement error (> 0.5 indicating convergent validity) (Hair
 et al.
2006). 
However, as AMOS does not output the CR and AVE directly, they were thus calculated by 
hand in this study based on the following formulas provided by Anderson and Gerbing (1988). 
After the modified measurement model is assessed, a structural model is estimated. The fit of 
the structural model is evaluated, and modification indices are again examined. In the 
structural model, modification indices indicate how much the model could be improved by 
adding in significant but unspecified paths (Schumacker and Lomax 2004). Nevertheless, it is 
vital to note that modifications of the structural model, based on the inclusion of paths shown 
to be significant but not theoretically predicted, is considered to be problematic because it 
increases the probability of Type I errors (Kline 2011). In addressing these concerns in this 
CR = (
∑
standardized loading)
2
/ [(
∑
standardized loading)
2
+ (
∑
indicator measurement error)] 
AVE = (
∑
standardized loading)
2
/ [(
∑
standardized loading
2
) + (
∑
indicator measurement error)] 

---

204
study, non-specified paths (between latent variables) were not added to the modified 
structural model.