Confirmatory factor analysis (CFA)

First, through EFA the items needed to represent the key constructs were determined (Hair et al., 2006). CFA was considered essential to validate the measurement scale. All CFA’s were analysed through AMOS (Analysis of Moment Structure) graphics 22.0 using the data stored in SPSS 22.0. CFA was carried out in two stages, namely, the measurement model and the structural model. The measurement models were used to determine how well the indicator variables determined through CFA are related to one another. The measurement model of CRM, the outcomes of CRM (repeat visitation and word-of-mouth) and variety-seeking behaviour were measured separately. The scales were validated through this process.

The model that was deemed fit for instrument validation and hypothesis testing was determined by standardised regression weights (>.5) (Hair et al, 2006; Holms-Smith 2010), squared multiple correlations (>.5), standardised residual covariance (between 2.5 to 4) and standardised effects, and the critical ratio (>1.96). The models that do not contain these criteria were improved based on the modification indices (Holmes-Smith, 2012). The model improvement due to specification was calculated with the chi square difference (∆ x2

). In addition to statistical criteria, the model specification was guided by the theoretical rationale (Hair et al, 2006; Kline, 2011). For example the items that were essential in representing the constructs were not removed despite their statistical cut off points.

Secondly, the structural models were used to determine the causal relationships necessary for hypotheses testing. The structural CFA models for hypothesis testing involved multi- group CFA, moderation and moderated mediation models. The groups for multi-group analysis were identified through the data collected by a categorical variable (please refer to Q.10). According to Hair et al. (2006) non-metric variables are often hypothesised as moderators. For example, the respondents belonging to each group were identified through a number. The survey respondents were categorised into four groups. Those who:

1. visited the same location and same hotel each time (112, 28%) 2. visited the same location and different hotels (158, 39.5%)

3. mostly visited different locations and different hotels (110, 27.5%)

4. mostly visited different locations and the same brand of hotel/chain (20, 5%)

Among them group four was not used for multi-group analysis due to an insufficient number of respondents compared to the other groups.

The mediating effect of VSB on the relationship between repeat visitation and word-of- mouth was evaluated. As the mediating effect varies across the three groups, the method adopted is referred to as moderated mediation (Baron & Kenny, 1986). The most popular approaches to determine indirect effects are causal steps strategy, distribution of the product strategies, resampling or bootstrapping or various products of coefficient strategies (Preacher, Rucker, & Hayes, 2007). Considering its popularity, this study used bootstrapping to assess the indirect effects, and resampling strategy for hypothesis testing (as cited in Kenny, 2013from Bollen & Stine (1990) and Shrout & Bolger (2002)). In AMOS, 2000 bootstrap samples and 95% biased correlated confidence intervals were set up. Both the direct (without the mediating effect) and indirect (with the mediating effect) effects were determined. This enabled the researcher to determine the existence of a mediating effect and whether mediation is stronger for variety seekers than for familiarity seekers.

3.8.5.1 Model goodness-of-fit criteria

Fit indexes were selected based on the criteria stipulated by Hair et al. (2006). Considering the sample size, model complexity and degree of error in model specification (Hair et al., 2006), the fit statistics used to determine the model fit are shown in Figure 3.4.

Table 3.3: Goodness-of-fit indices

Goodness-of-fit indices Ideal cut-off value

Sources

Chi-square (X2) P > 0.05 Holmes-Smith, 2012

X2/df (Normed Chi-square) >1 to <2 Holmes-Smith, 2012

GFI (Goodness-of-Fit) >.95 Holmes-Smith, 2012

CFI (Comparative Fit Index) >.95 Holmes-Smith, 2012 RMSEA (Root Mean-Square

Error of Approximation)

<.05 to <.8 Byrne, 2013

Even though the cut-off values given above were considered ideal for the measurement models since the sample size used for CFA was ˂250 (Hair et al. (2006), the cut off points: X2/df (Normed Chi-square) <5 and CFI˃.80 were also considered permissible (Hu

& Bentler, 1999).

3.8.5.2 The item reliability for CFA

Reliability is an assessment of the degree of consistency between multiple measures of a variable (Hair et al, 2006, p. 137). The item reliability of each latent variable was observed through the squared multiple correlations (Blunch, 2013, Hair et al, 2006, Holmes-Smith, 2012). According to Holmes-Smith (2012), item reliability between .3 and .5 was considered adequate. The rationale for this cut off has been explained by the level of variance exhibited by the items. Based on the variance exhibited by the items accounting for 50% of the variance are considered more reliable. Therefore >.5 is usually acknowledged as the cut-off point to determine validity of an item (Blunch 2013; Hair et al, 2006). However, items with squared multiple correlations <.5 were retained based on the model fit and depending on whether the item was a new one and developed through the qualitative data, since such items could be reformulated during future research.

3.8.5.3 The item validity for CFA

Validity is the extent to which a scale or set of measures accurately represents the concept of interest (Hair et al, 2006, p. 137). According to Hair et al. validity is categorised as convergent validity, discriminant validity and nomological validity. Broadly, the different categories of validities can be classified under the concept of construct validity (Kline, 2011). Convergent validity assesses the degree to which two measures of the same concept are correlated (Hair et al., 2006, p. 137). Discriminant validity is the degree to which two conceptually similar concepts are distinct (Hair et al., 2006, p. 137) and evaluates whether the two constructs are separate (Holmes-Smith, 2012). Nomological validity refers to the degree to which the summated scale makes accurate predictions of other concepts in a theoretically based model (Hair et al., 2006, p. 138).

Convergent validity is measured by standardised factor loadings or by average squared factor loadings (Hair et al., 2006). Both methods use the cut-off point >.5 or higher to

determine convergent validity. The convergent validity of the constructs involved in this study was determined through the standardised factor loadings of each item.

This study consisted of multiple factors that represent the same construct. In fact, CRM was measured both by purchase stage CRM and pre/post purchase stage CRM. The outcomes of CRM were measured through repeat visitation and word-of-mouth, whereas variety- seeking behaviour was measured through intrinsic factors and extrinsic factors. Therefore, determining the discriminant validity of the factors of each construct was important since the factors measuring the same construct were interrelated. Discriminant validity can be calculated through the average variance extracted method suggested by Fornell and Larcker, in 1981 (Hair et al., 2006; Holmes-Smith, 2012). To determine discriminant validity the variance extracted for the pairs of constructs was first calculated through the following formula:

Thereafter, the variance extracted by the two constructs was compared with the square of the correlation between the constructs. If the average variance extracted between the two constructs were greater than the square of the correlation between the constructs, the discriminant validity was assured (Hair et al., 2006; Holmes-Smith, 2012). This way the researcher ensured the distinctiveness between the constructs and also the unidimensionality of the individual items under a latent construct (Hair et al., 2006).