A MPLIACIÓN DE LOS RESULTADOS DE SIMULACIÓN

Following modification to obtain the best fitting model, the next stage in CFA is to assess the reliability and construct validity of the model, to demonstrate the accuracy with which the latent variables are represented by their indicators (Diamantopoulos and Siguaw, 2000; Hair, Black, Babin, Anderson and Tatham, 2006).

162 5.4.7.1 Reliability

Reliability is defined as “the degree to which measures are free from error” (Peter, 1979, p. 16) and therefore measures the consistency of the results

(Diamantopoulos and Siguaw, 2000). In SEM, reliability is assessed by examining factor loadings of each indicator on to its latent variable. In particular the squared multiple correlations of the indicators on to latent variables show the proportion of variance in an indicator that is explained by its underlying latent variable after measurement error is accounted for (Diamantopoulos and Siguaw, 2000). A value in the squared multiple correlation of above 0.5 means that at least half of the variance in an indicator is explained by its latent variable.

Composite reliability (CR) (also known as construct reliability) can be calculated for each latent variable. It is a measure of the proportion of variance in the latent variable shared between all its indicators. Hence it is a measure of convergent validity or the degree to which the indicators converge on to the construct. CR is calculated from the squared sum of the (standardised) factor loadings divided by the sum of that value and the sum of the error variances for each factor loading, and should ideally exceed 0.7 for good reliability (Hair, Black, Babin, Anderson and Tatham, 2006).

5.4.7.2 Unidimensionality

A well-specified measurement model assumes that constructs are unidimensional, and that the measures in a scale associated with a latent construct measure that construct only (Gerbing and Anderson, 1984). Unidimensionality means that each indicator has only one underlying construct, and a unidimensional scale consists of only unidimensional indicators (Ping, 2004; Steenkamp and van Trijp, 1991).

Where a model contains more than one construct, unidimensionality exists when there are no cross loadings ( i.e. no indicators which load on more than one construct and cross-loadings are set to zero), and all covariances between error variances are also zero (Hair, Black, Babin, Anderson and Tatham, 2006).

Therefore in SEM the researcher specifies the measurement model using unidimensional measures and a good model fit suggests that unidimensionality

163 exists (Ping, 2004). A model which fulfils these conditions is known as a

congeneric model (Hair, Black, Babin, Anderson and Tatham, 2006).

5.4.7.3 Validity

Validity is “the degree to which instruments truly measure the constructs which they are intended to measure” (Peter, 1979, p. 16). Construct validity is assessed through face validity, nomological validity, convergent validity and discriminant validity. Face validity concerns whether the scale of items captures what it seeks to measure (Churchill, 1979), and how well the items match the definition of a construct as it was originally conceptualised (Ping, 2004). Nomological validity refers to whether the predicted relationships in the model are supported by the measures. The researcher identifies theoretical relationships between constructs from prior research and then assesses whether the scale similarly measures these relationships. If these relationships are confirmed by the measures, nomological validity can be said to exist (Hair, Black, Babin, Anderson and Tatham, 2006).

Convergent validity represents how well measures of the same construct are correlated (De Vellis, 2003). A measure of convergent validity is Average Variance Extracted (AVE). AVE measures the amount of variance captured by the intended construct relative to the variance relating erroneously to other constructs not hypothesised as related in the model (Diamantopoulos and Siguaw, 2000). The AVE of a construct is the sum of the squared (standardised) factor loadings divided by the number of indicators (Hair, Black, Babin, Anderson and Tatham, 2006; Ping, 2004). AVE values of more than 0.5 show that the scale of items will converge on to that construct more than on any other construct in the model (Fornell and Larcker, 1981; Ping, 2004).

Discriminant validity, on the contrary, tests that a construct does not correlate highly with any other construct to which it is not conceptually related (Churchill, 1979; Ping, 2004). Discriminant validity of a construct can be tested by assessing correlations with other constructs. Correlations between two constructs which are below an absolute value of 0.7 (or squared correlations below 0.5) show that these pairs of constructs are distinct and hence discriminant validity exists (Ping, 2004).

164 A more rigorous test of discriminant validity is to use AVE tests. As described by Ping (2004): “If the squared correlation between constructs (r²) is less than either of their individual AVEs, this suggests the constructs each have more error-free (extracted) variance than variance shared with other constructs (r²). In other

words, they are more internally correlated than they are with other constructs. This in turn suggests discriminant validity” (p. 132). This is the method used to assess discriminant validity in the data analysis reported in the following Chapter.

5.4.7.4 Common method variance

In addition to reliability and validity, another concern is the issue of common method variance (CMV). CMV is “variance that is attributable to the measurement method rather than to the constructs the measures represent” (Podsakoff,

MacKenzie, Lee and Podsakoff, 2003, p. 879). This is of particular concern when the survey method is cross-sectional, when self-reported perceptual data on both the independent and dependent variables is gathered concurrently from

respondents (Chang, van Witteloostuijn and Eden, 2010). CMV can create a false internal consistency whereby correlations are found between variables due to their common source, rather than the existence of a valid relationship (Chang, van Witteloostuijn and Eden, 2010).

Methods to avoid CMV have been discussed in the previous chapter (Sections 4.5.2 and 4.5.3), and include using different sources of information, creating temporal separation, randomly mixing the order of questions and varying the response formats (Podsakoff, MacKenzie, Lee and Podsakoff, 2003; Rindfleisch, Malter, Ganesan and Moorman, 2008). In the town centre image survey, it was not possible to implement the first two methods, for practical reasons, although

questions were randomised and response formats varied.

There are also several statistical remedies available to control for CMV. One of the most common is Harman‟s single-factor test. This method is based on the

assumption that a substantial amount of CMV is present if, using either exploratory or confirmatory factor analysis, all variables can be demonstrated to load on to one factor, or a single factor accounts for the majority of covariance among measures

165 (Chang, van Witteloostuijn and Eden, 2010; Podsakoff, MacKenzie, Lee and

Podsakoff, 2003). A more rigorous method is to use the marker variable technique (Lindell and Whitney, 2001). A marker variable is included in the analysis which is not theoretically related to the other variables in the study. Any relationships observed between the marker and the other variables not hypothesised to be related to it are assumed to be due to CMV (Podsakoff, MacKenzie, Lee and Podsakoff, 2003).