Refining and testing the self-construal scale

The aim of these analyses was to find a final set of items which would make up the optimal measurement of the seven self-construal factors and show the most cross- cultural comparability. In a similar way to Study 1 (see Chapter 5), I ran a series of single group and multigroup CFA in MPlus version 5. Once again I modelled

acquiescence as a separate uncorrelated factor which loaded onto every indicator at a fixed value of 1(Welkenhuysen-Gybels et al., 2003). Model fit was assessed using the Comparative Fit Index (CFI), Root Mean Square Error of Approximation (RMSEA), and the Standard Root Mean Squared Residual (SRMR). Values of CFI above .90, RMSEA up to .08, and SRMR up to .10 are seen as acceptable (Kline, 2005).

Achieving configural invariance. Firstly, I tested the seven-dimensional model

with all 58 items in the four samples separately. Any items that did not load

significantly on their respective factor or loaded opposite to the intended direction in any of the four national samples were removed from further analyses, which brought the number of items down to 56 (items 8 and 9 removed, see Appendix C). The fit of the model with the remaining items was still poor in all four samples and it was clear from

the modification indices that several items had a large number of cross-loadings. In order to identify the most problematic items, I inspected the modification indices closely and I ran partial correlations where each item was set to correlate with the means of the other dimensions, while controlling for the mean of its intended dimension. From these analyses I found 21 problematic items, which were removed from further analyses. I tested the resulting model in each national sample separately (see Model 1 – 4, Table 7.2) and as can be seen from the table, all three fit indices suggested good fit in Romania and the UK and RMSEA and SRMR indicated good fit in Thailand and Malaysia (Appendix C shows which items were included in the final model).

This difference between fit indices deserves some further discussion. CFI has been criticised for being based on an independence model, which assumes zero

covariances between variables. Such a baseline or null model is unrealistic in most SEM research (see Kline, 2005). Rigdon (1996) argued that compared to CFI, RMSEA is more useful in re-evaluating earlier research and confirming existing models because no such baseline model is involved in the calculation of RMSEA. SRMR is similar to RMSEA in that it is based on residuals rather than comparison with baseline model. Hu and Bentler (1998) recommend using SRMR in combination with either RMSEA or CFI. In line with the argument above, considering the aim of Study 6 was to confirm the seven-dimensional model identified in Study 4, it seems that SRMR and RMSEA are more useful for evaluating model fit in the present analyses.

Testing metric and scalar invariance. I then created a multigroup model, again

including a method factor, analyzing all samples simultaneously (see Table 7.2, Model 5), and I tested the impact on model fit of constraining first the factor loadings (for metric invariance: Model 6) and then the intercepts (for scalar variance: Model 7) to be

Table 7.2 Single-group and Multigroup Invariance Analysis Study 6

Model χ² df CFI RMSEA

RMSEA

90% CI SRMR LL UL

Model 1, Malaysia seven-

factor model 1527.51*** 538 .81 .06 .06 .07 .08

Model 2, Romania seven-

factor model 1045.21*** 538 .90 .05 .04 .05 .06

Model 3, Thailand seven-

factor model 1251.60*** 538 .86 .05 .05 .06 .07

Model 4, UK seven-factor

model 862.75*** 538 .90 .05 .05 .06 .08

Model 5, multigroup

configural invariance 4687.06*** 2152 .87 .06 .05 .06 .07 Model 6, factor loadings

constrained 5221.18*** 2257 .84 .06 .06 .06 .10

Model 7, factor loadings

and intercepts constrained 6110.40*** 2338 .80 .06 .06 .07 .10 Model 8, factor loadings

and 28 intercepts constrained

5580.22*** 2317 .83 .06 .06 .06 .10 Model 9, factor loadings

and 21 intercepts constrained

5380.65*** 2296 .84 .06 .06 .06 .10

Note. CFI = Comparative fit index; RMSEA = Root mean square error of

approximation. CI = Confidence intervals; LL = Lower level; UL = Upper level. Multilevel analyses did not provide confidence intervals for RMSEA.

*** p < .001.

equal across samples. The impact of constraining factor loadings and intercepts was small according to RMSEA and SRMR, which suggests that invariance is tenable. However, CFI dropped substantially, especially when the intercepts were constrained. I therefore investigated whether CFI could be improved by testing for ‘partial intercept invariance’ (Byrne et al., 1989) in the same way as in Chapter 5 (Section 5.2.2.1).

By inspecting modification indices, I identified seven problematic intercepts (one per factor). Freeing these intercepts improved CFI to some extent (see Model 8). Freeing another seven intercepts (again one per factor) provided an additional small

improvement (see Model 9) and reached the same level of invariance as when only factor loadings were constrained. Hence, it did not seem worthwhile freeing more intercepts.

Overall, these results provide a somewhat mixed picture. Configural, metric, and scalar invariance of the seven factor could be considered tenable when assessing model fit according to RMSEA and SRMR. CFI on the other hand suggested a poor fit for the constrained models. As noted above, there are arguments in favour of using RMSEA rather than CFI when confirming models from previous research and hence the present results provide some confidence in the cross-cultural validity of the 35 item self- construal scale.

Latent mean differences. Considering these indications of invariance, I then

moved on to look at how the four samples differed in their levels of endorsement of the self-construal dimensions. When investigating latent means, groups are compared against one reference group which has its mean set to zero. In line with common practice in the literature where other cultures are compared against the West (e.g. Oyserman et al., 2002), I chose the UK sample as the reference group. A significant result should therefore be interpreted as a significant difference from the UK sample. Table 7.3 shows that the Thai and Malaysian samples scored lower than UK participants on uniqueness whereas the Romania sample scored higher. The opposite pattern was observed for harmony. Considering Study 5 found self-differentiation (high uniqueness and low harmony) to be related to individualism, this pattern fits in with the common view in the literature which describes East Asia as less individualistic (e.g. Triandis, 1995). Interestingly, these results suggest that Romania may be more individualistic than the UK, which was also the case in Study 5. Malaysian and Thai participants scored lower on self-direction whereas the Romania sample did not differ from the UK

sample. British participants scored the lowest on inclusion and the Malay and Thai samples also scored higher on commitment, while Romanian participants scored lower than the UK sample. These results differ from those in Study 5 where the UK scored high on inclusion and commitment—a discrepancy which could be the result of the present student samples and the large proportion of rural UK participants in Study 5. In terms of self-reliance, the UK and Thai samples did not differ, whereas the Malaysian and Romanian samples scored higher. Finally, the Malaysian, Thai and UK samples did not differ on consistency but the Romanian sample scored higher. Mirroring results from Study 5, the Romanian sample was again found to score relatively high on self- reliance and consistency, whereas unlike the previous study where the Thai sample score higher on these dimensions, the present results suggested comparable levels between the Thai and UK sample.

Table 7.3 also shows reliabilities based on ipsatized scores for each of the seven dimensions, which were good or acceptable in all four samples.

Table 7.3 Latent Means and Reliabilities Study 6

Malaysia Romania Thailand UK

Mean α Mean α Mean α Mean α

Uniqueness - 0.24* .73 0.33** .78 - 0.50*** .80 0.00 .84 Harmony 0.98*** .62 - 0.23* .73 0.85*** .68 0.00 .80 Self- direction - 0.87*** .63 - 0.05 .76 - 0.43*** .72 0.00 .80 Inclusion 0.87*** .80 0.57*** .71 0.70*** .73 0.00 .77 Commitment 0.25* .65 - 0.50*** .70 0.37** .65 0.00 .77 Self-reliance 0.17** .73 0.76*** .84 0.00 .85 0.00 .91 Consistency 0.33 .65 0.75*** .86 0.18 .77 0.00 .92

Note. Table is showing standardized estimates.