2. METODOLOGÍA
2.2. Diseño de elementos y partes
2.2.2. Transmisión de potencia
Based on the research model in Figure 3.1, Measurement Model 5 illustrated in Figure 4.6 was designed to examine the relationships between the five higher-order constructs (service quality, customer satisfaction, perceived value, restaurant image and behavioural intentions) and the measured items. The preliminary measurement model is shown in Figure 5.15.
The preliminary model represented 20 items. The measured variances and covariances of the 20 measured items were v = 210 pieces of information
parameters (15 regression weights, 10 covariances, and 25 variances); the model was over- identified with 160df (210 pieces of information - 50 parameters). The preliminary measurement model of the causal path was statistically significant at the .001% level. Except GFI (see Table 5A.10, Appendix 4), the model-fit-indices for all indicators were more sufficient than the relevant recommended thresholds presented in Table 4.5
2160 469.754;2 df 2.936; NFI = 0.927; CFI = 0.950; RMR = .036; and RMSEA = .087). The GFI value (0.850) was slightly lower than the 0.90 threshold suggesting that some modifications were needed to improve it.Figure 5.15: Measurement Model 5 – Preliminary Model for Causal Path
In reviewing the correlation estimates on all pairs, six pairs PVCS (0.853), PVRI (0.850), SQRI (0.765), SQBI (.839), BICS (0.852) and PVBI (0.813) were less than the recommended value
r0.85
of Kline (2005) indicating the existence of discriminant validity. However, four pairs were higher than the recommended value: SQCS (0.888), BIRI (0.884), SQRI (0.862) and RICS (0.864), therefore some modifications were needed in order to improve the discriminant validity of the model. In addition to the information provided by model-fit-indices and discriminant validity of this model, it may be concluded that the preliminary measurement model required somemodification. In reviewing the MI, four miss-specified items were identified (see Table 5.22). Other suggestions in the MI were ignored in this study because they could not be supported by strong, substantive and empirical rationale (see: Chinna, 2009; Hair et al., 2010). Based on the MI information, the preliminary measurement model for causal path was re-specified four times. Through the model modification process, three items remained for all constructs except behavioural intentions with four items.
Table 5.22: Suggestions for Improving the Model-Fit-Indices from the MI
Items Cross loadings with Modifications Index Expected Par Change
D15 CS 6.077 -.032 RI 5.831 -.031 BI 11.482 .050 SQ 5.748 .038 D4 SQ 11.159 -.059 D16 RI 10.363 .042 BI 7.900 -.040 D10 CS 4.464 .030 RI 6.943 .038 PV 13.003 -.053
The modified measurement model represented sixteen items. The number of observed variances and covariances was v = 136 pieces of information
16
161
2136
and the number of estimated parameters was p = 42 parameters (11 regression weights, 10 covariances, and 21 variances); the model was over-identified with 94df (136 pieces of information - 42 parameters). The modified measurement model is shown in Figure 5.16.The modified measurement model was statistically significant at the .001% level and model-fit-indices sufficiently satisfied the relevant recommended thresholds as presented in Table 4.5 (
294 249.157; 2 2.651;df
NFI = 0.948; CFI = 0.967; RMR = .034; and RMSEA = .081), indicating that the model fitted the data. The GFI value (0.893) was little improved from the preliminary measurement model and was marginally adequate (Hair et al., 2006; Kim, 2003; Kline, 2005) (see Subsection 4.3.3.2.1). The modified measurement model was considered as the final measurement model for causal path, reflecting the best model fitting based on the aforementioned information of model fit and MI.
The improvement in model fit was also examined by subtracting the overall
2
preliminary model
2160 469.754
with the modified model ( 294 249.157
yielded adifference in 2 value of 220.597
266 220.597
. Since
266 220.597
852.965,.05, the modified model was statistically significant and indicated an improvement in the model-fit-indices.Figure 5.16: Measurement Model 5 – Modified Model for Causal Path
To verify construct validity, convergent and discriminant validities were measured (Janssens et al., 2008; Kline, 2005; Schumacker & Lomax, 2004). Unidimensionality was assessed. The standardized factor loadings ranged from 0.838 to 0.950, all well above the acceptable value
0.50
suggested by Bagozzi and Yi (1988) and statistically significant at the .001% level. This evidence supported the unidimensionality of each scale, which indicated that convergent validity was obtained. With unidimensionality satisfied, the reliability of the constructs must always be verified (Chinna, 2009; Hair et al., 2010; Janssens et al., 2008; Nokelainen, 2009). Reliability was verified with Cronbach alpha, composite reliability and AVE. Cronbach alpha for all constructs in the modified measurement model ranged from 0.907 to 0.952, which exceeded the threshold level of 0.60, indicating high internal consistency of the measurement scales (Churchill, 1979; Hair et al., 2006; Janssens et al., 2008). The results from the modified measurement model showed that composite reliability calculated using Equation 4.1 for all constructs ranged from 0.904 to 0.953 which exceeded the threshold level of 0.70. Using Equation 4.5, the calculated AVE values ranged from 0.758 to 0.868 which exceeded the minimum criterion
AVE0.50
. These five constructs indicated that the variance due to measurement error was less than the variance captured by the construct (Fornell & Larcker, 1981). This evidence supported the reliability of each scale, indicating both that convergent validity had been obtained and also therefore the high reliability and validity of the measurement items (Bagozzi & Yi, 1988; Chen, 2008; Fornell & Larcker, 1981). In reviewing the correlation estimates of all pairs, 10 pairs were less than the recommended value
r0.85
of Kline (2005) indicating the existence of discriminant validity.In summary, the modified measurement model generally exceeded the criteria established in Table 4.5 for model-fit-indices (see Table 5A.10 in Appendix 4). In addition, all conditions required for examining the convergent and discriminant validities recommended by several researchers (Anderson & Gerbing, 1988a; Byrne, 2009; Chinna, 2009; Kline, 2005; Nokelainen, 2009; Schumacker & Lomax, 2004) were satisfactorily met. The modified measurement model represented the best model fit for measuring the service quality structure with the present data; therefore, the model was used for the structural model.