Model Validation
9.5. Model Validation
The dataset is imported into SmartPLS 264 and seven variables are computed with the research questions identified from SPSS. The estimation is based on specific performance measurements, such as R square, CA, Average Variance Extracted (AVE), and T- values. 265
R square should be optimized, so a high level of explanation of loyalty is present, which is essential for the validity and usability of the model. According to ECSI measurements, R square should be >0.65. 266 This is fulfilled by the variable satisfaction loyalty with a R square of 0.78, which is a high coefficient of determination and favorable for the validity of the model. However, the R square for perceived value has not quite meet the requirement, but is very close with a R square of 0.59, hence, it is seen as usable. Satisfaction loyalty is the aim of the model and it is therefore much more important that R square is above 0.65 for that result-variable, than it is for perceived value.
In SmartPLS CA is also estimated to show the variable’s capability of creating consistent results. In all variables the CA is above the recommended 0.6 267 and this is a very good indicator of the fact that each variable can create consistent results.
AVE should secure that the model explains more valid variance in measurement than error and should be >0.50. 268 Again, the variables perform well because they are all above 0.50. Hereby, the relationships between the latent variables and its indicators are high and the error is diminutive (see figure 14).
It is important that the path coefficients (impact scores) between the variables all are significant (p<0.05), which tell that there is causal relationships between the variables. All the outer weights are high meaning they are all above 0.724. An optimization should also have taken place, but since the values are so high and rather stable, there is no need to optimize further.
263 Appendix P (disc) Lecture notes from Markedsanalyse, Lars Grønholdt, lecture 23 2007, slide 12
264 www.smartpls.de
265 See the specific values of the WLN Model of Reader Loyalty in Appendix 10 - Output from SmartPLS (Generic Model)
266 Kristensen, Kai et al. (2000), p. 57 267 Malhotra, Naresh K. Birks, David F. (2003),p. 314 268 Fornell, Claes (1992), p. 15
It is now evident in SmartPLS that AVE and CA is still not high enough with the four questions remaining in the variable others, therefore the question with the lowest outer-
weight (Q 32 ) is removed. Q 32 concerns job advertisement and it is a very small area for
most WLNs and it can easily be defended that this is also removed.
The variable others is then left with three measurement variables (Q 34 ,Q 38 and Q 39 ) and
they all concern topics of generic character (Names, offers and layout). Others is now a part of the WLN Model of Reader Loyalty, represented by subjects that are generic for the industry, with a CA of 0.614. It should be kept in mind that if a specific WLN has emphasis on one or more of the removed subjects in others should anyhow include them in order to test their ability to contribute to loyalty.
Generic WLN Model of Reader Loyalty
Figure 14 - Source: Output from SmartPLS 269
The T-values should be >1.65 with p<0.05 in order to show significance. 270 This points to that there, with at least 95 certainty, is a significant relationship between two variables. The relationship between housing and perceived value is –0.087. A t-test is conducted and it is now known that this path is insignificant. Therefore this path is removed. The t- test was conducted again and it showed that all the path coefficients were well above
1.65 ranging from 1.657 to 6.679 as the highest.
269 Appendix 10 - Output from SmartPLS (Generic Model) 270 Sirdeshmukh, D. (2002), p. 26