CHAPTER 3 METHODOLOGY
3.6 STAGE 5: SECOND DATA COLLECTION AND ITEM ANALYSIS
Although an EFA was performed to examine the extent to which the scale items assess the content domains, a significant limitation of the analysis is its incapability to statistically quantify the model fit of the factor structure (Long, 1983). Unlike CFA that specifies the structure of the scale and the relationships among the variables of interest in advance, EFA only provides a post hoc interpretation of the outcomes (Hinkin et al., 1997). Consequently, items that load fittingly in an EFA sometimes show a lack of fit in a model with a different sample due to lack of external consistency (Gerbing & Anderson,
1988). It is recommended that a scale developer must perform a CFA regardless of whether an exploratory analysis have been conducted or not (Hinkin et al., 1997). 3.6.1 PARTICIPANTS AND SAMPLE SIZE
The same sampling framework was applied to recruit qualified participants: adult sports fans who had a recent history of attending live sporting events. In the previous data collection, eligible participants were those who attended any live sporting event within the last three month, which was due to potential time constraint of recruitment. However, it took much less time for collecting the satisfactory number of complete surveys in the previous stage. Thus, in this stage more rigorous restriction as “within the last month” was applied to the screening process.
When it comes to sample size, the same rules can be applied to decide appropriate number of observations in both CFA and SEM (Kline, 2013) since CFA is a type of SEM that is designed to assess the goodness-of-fit of models. No consensus exists in the
literature about the satisfactory number of observations, but scholars at least present some evidence that a simple CFA or SEM model could be rigorously tested even if the sample size is relatively small (Hoyle, 1999; Hoyle & Kenny, 1999; Marsh & Hau, 1999). Quite often, 100 to 150 is considered the minimum sample size (Tabachnick & Fidell, 2001; Ding, Velicer, & Harlow, 1995). Other researchers recommend a larger sample size of at least 200 (Boomsma & Hoogland, 2001; Kline, 2013). In the situation of no missing data, an acceptable sample size for a CFA model can be about 150 (Muthén & Muthén, 2002). The sample size is often also considered in relation to the number of observed variables. Bentler and Chou (1987) propose that a ratio of 5 cases per parameter to estimate would be sufficient when latent variables have multiple indicators with normally distributed data.
Another widely accepted direction is ten observations per an item as adequate sample size (Nunnally, 1967). The researcher took a conservative approach collecting 510 observations, which meets all of the guidelines mentioned above.
Fifty-six percent of the respondents were between 30 and 49 years of age and other age groups were 33% of 18 to 29, 10% of 50 to 69, and less than 1% of 70 or older. Sixty-six percent of the sample was male, and the racial breakdown of the participants was 73% Caucasian, 12% African American, 6% Hispanic, 8% Asian, and 1% self- identified Others. Demographics of the sample were consistent with the population of sports fans in the United States (“Demographics of Sports Fans,” 2017).
3.6.2 INSTRUMENT AND DATA COLLECTION PROCEDURE
A new questionnaire was distributed through MTurk (see Appendix E for the questionnaire). The same two screening questions used in the first data collection were also presented to check whether respondents were old enough and recently attended sporting events. Total 510 participants completed the survey receiving $1 each as
compensation. The questionnaire for the second data analysis included four measures: the SIF, event experience satisfaction, revisit intention, and post familiarity with the local culture. Demographic information was also gathered. The SIF scale was measured with the 26 items evaluated in the previous EFA stage. Participants rated each item on a 7- point Likert scale ranging from (strongly disagree) to 7 (strongly agree). The other three measures were included in the survey for the criterion validity check afterward. Event experience satisfaction was measured with the three items from Bitner and Hubbert (1994), with a 7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree). Revisit intention was measured with a single item from Lee et al. (2012) using a
7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree). Post familiarity was measured with a revised four-item measure derived from Toyama and Yamada (2012) using a 7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree). Those respondents who successfully completed the survey received a confirmation code at the end, and the research exported the usable data to RStudio for subsequent analysis.
3.6.3 DATA ANALYSIS
Once a factor structure had been identified, a CFA was conducted to assess the quality of the factor structure by statistically testing the significance of the overall model, as well as the relationships among items and the scale. Both global and local fit indices were evaluated to examine model fit. It is recommended to investigate many fit indices as they focus on different parts of the model (Fan, Thompson, & Wang, 1999, Hu & Bentler, 1999, Schumacker & Lomax, 2012). Global fit focuses on indices that investigate the “overall” model fit. The indices evaluated in the study include 1) Chi-Square test, 2) Comparative Fit Index (CFI), 3) Tucker-Lewis Index (TLI), 4) Root Mean Square Error of Approximation (RMSEA), and 5) Standardized Root Mean Square Residual (SRMR). Chi-square test should be rejected (p > 0.05) for exact fit (Barrett, 2007). The
Comparative fit index family (CFI and TLI) compares the fit of the model to its baseline model. These should be higher than .95 or at least .90 to show acceptable fit (Hair et al., 2010). MacCallum, Browne, and Sugawara (1996) suggest RMSEA should be under 0.05 and 0.08 for good and mediocre fit respectively. SRMR that shows the amount of average residual across the model should be under .08 for an acceptable model or .05 for a good fitting model (Kline, 2013). Table 3.1 presents the fit indices used for the assessment.
Table 3.1. Fit indices used in the CFA
Measure Description Cut off for
acceptable fit Chi-
Square
Access model fit by comparing the discrepancy between the sample covariance and the fitted covariance matrices. Null hypothesis: the model fits perfectly
P-value > .05 TLI An TLI of .90 implies the model improves the fit by 90%
compared to the null model. TLI³ .90
CFI Revised TLI that is not very sensitive to sample size. CFI ³ .90 RMSEA The average of residuals between the sample and the
fitted model matrices.
RMSEA< .08 SRMR The difference between the square-rooted residuals of the
sample covariance and the fitted model. SRMR < .08
Regarding local fit criteria, three indices were examined: 1) parameter estimates to make sure the values are significant and in the correct direction, 2) standard errors of parameter estimate to make sure the values are similar across the parameter estimates, and 3) R-squared values showing the amount of variance explained in the indicator by the factors.