38 / 414
服務主導邏輯之共同生產：前置因素與結果因素
38
for ANOVA analysis. By contrast, Palmatier, Jarvis, Bechkoff, and Kardes (2009) used a
single item (Percentage Figure) to measure share of wallet and then pooled the percentage
figure and Likert-type scale of other constructs for the subsequent structural equation
modeling (SEM) analysis. In the current study, the method of Palmatier et al. (2009) was
adopted, and hence, a five-point Likert-type scale and percentage figure were used for data
analysis. Most importantly, the use of different scale types and formats for the independent
and dependent variables can diminish common method variance (CMV) caused by
commonalities in anchoring effects. This method eliminates the effects of consistency
(Podsakoff, MacKenzie, Lee, and Podsakoff, 2003). In this study, the three managers from
the investment services industry were recruited to review specific items and the definitions
of all the constructs that were included in the questionnaires. To ensure the internal validity
of the measurement, the three managers were asked if the selected items were able to
measure the underlying constructs. Several items in the questionnaire were modified based
on their suggestions to suit the investment services context.
3.3 Validation of Measures
Consistent with the two-step approach advocated by Anderson and Gerbing (1988), this
study first developed the measurement model by conducting confirmatory factor analysis
(CFA). The structural equation modeling was then estimated for hypotheses testing. The fit
of the model was acceptable (chi-square (203) = 827.04,
p
= 0.00; GFI = 0.90; CFI = 0.96;
PNFI = 0.84; NNFI = 0.93; RMSEA = 0.07; RMR = 0.05). The Cronbach’s alphas of all
constructs were all greater than 0.80, supporting the reliability of the measurement. In
addition, all composite reliabilities were greater than 0.70 and all average variance extracted
(AVE) estimates were greater than 0.50 (Fornell and Larcker, 1981). As evidence of
convergent validity, all of the items had significant loadings on their respective constructs
(Anderson and Gerbing, 1988). Evidence of discriminant validity was supported by the fact
that none of the confidence intervals of the phi estimates among the pair of constructs
included one in this study (Anderson and Gerbing, 1988). Discriminant validity was also
tested among all constructs according to Fornell and Larckerʼs (1981) recommendations and
confirmed for all pairs of constructs. Specifically, AVE estimate for each construct was
greater than the squared correlation of all construct pairs.
Given that the data were self-reported, CMV was expected. One ex ante means of
avoiding or minimizing any potential CMV is that respondents are assured of the anonymity
and confidentiality of the study. Another approach is to apply ex post statistical approaches.