Validating the measurement model

In the reflective mode, the inner measurement model was tested by examining a) internal consistency b) convergent validity and c) discriminant validity. From the theoretical framework we know that both of the latent constructs of EXPKNOW and CUSTMKNOW consist of latent variables of alliance learning (ALLLEA-CAP), alliance management (ALLMGMT(ALLLEA-CAP), marketing planning and

imple-mentation (MKTCAP) and new product development capabilities (NPDCAP).

The internal consistency of the latent variables was first of all assessed by the loadings values between the indicator and its latent variables. Two of the nega-tively worded items TARSAVOL and RPLACMNT were reversed before com-puting the individual items loadings. However, these were the only two items indicating < 0.50 loadings on their respective latent constructs. All the rest of the indicators in the reflective mode of the model loaded with ≥ 0.50 values on their respective latent variables indicating a high degree of individual item reliability.

The loadings for each of the individual indicators with their relevant latent vari-ables are given in Table 1, Appendix 2.

To further validate the internal consistency of the measures, the construct reliabil-ity for each of the main latent constructs (CUSTMKNO and EXPKNOW) by means of composite reliability was carried out. As can be seen from Table 1 in Appendix 2, the composite reliability values for the latent variables of all four kinds of capabilities are ≥0.80, thus showing a high degree of internal consis-tency of latent constructs.

To assess the convergent validity of the reflective block of the model, the average variance extracted (AVE) with a value higher than 0.5 has also been suggested in literature (Fornell and Larcker 1981). As can be seen from Table 1 Appendix 2, three latent variables in the reflective mode of the model demonstrated a ≥ 0.50 AVE. Values higher than and equal to 0.5 indicated that the measurement errors account for relatively less variance in the indicators than the latent variables.

Thus, the latent variables measured the real phenomenon with less degree of measurement errors. For that reason, all the latent constructs were found to be sound and satisfactorily valid.

Fornell and Larcker (1981) suggested that the AVE can also be used to assess the discriminant validity of the study. However, in such a case the AVEs of the latent variables should be greater than the square of the correlations with any other la-tent variables. This could indicate more variance shared between the lala-tent vari-able components and its block of indicators than with components of another la-tent variable representing a different block of indicators (Chin 1998; Fornell and Larcker 1981). From Table 2 Appendix 2, it can be seen that the square of the AVEs of the latent constructs is higher than the correlation among any other la-tent variables. Thus, indicators for each lala-tent variable shared more variance be-tween their respective latent variable components. As mentioned in Chapter 5.2, clear guidelines as to how much greater the squared AVE should be than the cor-relation among any other latent variables are not given in literature.

Analyzing the cross loadings of the latent variables have been suggested as an-other test for assessing the discriminant validity of the reflective mode of the model. Cross loadings were obtained by calculating the correlations between la-tent variable scores and other indicators beside the lala-tent variables’ own block (Chin 1998:321). It can be seen from Table 1 Appendix 2, that all the indicators loaded higher on their own latent variables and not a single indicator shared higher scores with a latent variable from another block. Thus, all the items re-flected their respective latent variables and measured what they were supposed to measure. As can be seen from Table 19, all the individual items loadings are ≥ 0.5 thresholds.

Table 19. Loadings, cross loadings, composite reliability and the AVE of the modified measurement model loadings on their latent constructs as less than 0.50 and were supposed to be re-moved from the measurement model. Thus, after removing these two items, the reliability and validity of the modified measurement model was assessed again by

following the same assessment criteria and procedure as mentioned for the origi-nal measurement model. Further, by removing the two indicators from the ALLMGMTCAP construct, its composite reliability value improved from 0.89 to 0.92. However, there was a minor decrease in the composite reliability for NPDCAP from 0.80 to 0.76. Overall, all the measures and latent variables scored a high degree of internal consistency for individual items reliability and construct reliability for the modified measurement model. Next, to assess the discriminant validity of the modified measurement model, the AVE for the latent variables was assessed. Besides ALLEARCAP, all the latent variables showed greater than .50 AVE. This is shown in Table 19.

To test the discriminant validity of the latent variables, the Fornell and Larcker criterion of the squared AVE was computed for the latent variables. As Table 20 indicates, all the squared AVEs were fairly larger than the correlations among the latent variables. Thus, the latent variables in the modified measurement model showed a high degree of discriminant validity at the construct level.

Table 20. Inter-construct correlations and the AVE along the diagonal Latent variables ^{ALLEARCAP ALLMGMT}

CAP

MKTCAP NPDCAP

ALLEARCAP 0.64

ALLMGMTCAP 0.385 0.84

MKTCAP 0.167 -0.047 0.81

NPDCAP 0.277 0.374 0.148 0.72

Next, to further assess the discriminant validity of the modified measurement model, the cross loadings of the latent variable and other indicators besides the latent variables’ own block were computed. As shown in Table 19, none of the indicators loaded higher with another latent variable than the one it was intended to measure. Thus, all the indicators reflected a true score for their own respective latent variable.

In sum, the modified measurement model demonstrated the required degree of individual item reliability, construct reliability and discriminant and convergent validity. In the next section, the quality of the inner structural model is evaluated.