Journal of Electronic Commerce Research, VOL 8, NO.1, 2007
units. The model parameters are estimated by weighted least squares algorithm [Jöreskog and Sörbom 1996a]. Model structures were learned and the parameter estimated in an explorative and iteratively way. The model is then tested on sample 2 in order to guarantee unbiased measures of goodness of fit.
Firstly, we present a factor analysis, secondly evidences derived from the model, and finally we close with remarks on privacy, trust and familiarity of respondents. 5.1 Factor Analysis
The factor analysis included all items from Table I of the appendix. The "Eigenvalue > 1" criterion lead to an initial five-factor model. However, a strong evident decline in the scree-plot after the sixth factor demanded a rotation with six factors. The extraction with principal component analysis (PCA), and oblimin rotation (delta = 0°) resulted in 74 % explained variance. The first factor has a relatively high fraction of the overall variance (33,9 %). After rotation, all factors had Eigenvalues above 2.
Four factors displayed medium intercorrelations (see Table 3), which underlines the necessity of an oblimin rotation. The pattern matrix of the rotated solution can be found in Table II of the appendix.
The factor loadings can be found in Table I of the appendix. All factors include three items each with high factor loading above .6, except for the last factor (trust), cf. -.52, -.58 and -.76. Each factor contains three items that comprise one scale of the questionnaire (compare Table I in the appendix). Thus, the confirmatory factor analysis supports the hypothesized scales from the questionnaire and allows testing of models of causal influence between factors. Accordingly, the reliability (cronbach’s alpha) of the scales is satisfactory (.72 for willingness to buy, .73 for perceived size, .86 for perceived reputation, .81 for trust, .95 for perceived privacy and .72 for risk perception). The medium factor correlations between the factors perceived reputation and perceived privacy, perceived reputation and perceived size, perceived reputation and risk perception, and perceived reputation and trust already indicate that influences between factors exist.
Table 3: Factor inter-correlation matrix
5.2 Linear Structural Models
To test our main five hypotheses, the six factors identified above are inserted into a linear structural model according to Figure 1.
The models were developed with the SIMPLIS command language [Jöreskog and Sörbom 1996a] and LISREL 8.54 [Jöreskog and Sörbom 2003]. Due to the fact that ordinal questionnaire data was used, the weighted least squares algorithm for polychoric correlations was employed, including the asymptotic covariance matrices [Jöreskog and Sörbom 1996a].
However, satisfactory model parameters for the original hypothetical model (compare Figure 1) could not be reached. In the non-fitting model, the path between risk perception and willingness to buy (Hypothesis 5) showed a non-significant t-value (r = .10, t = 1.77), so the factor willingness to buy was removed from the model. In the resulting second model, the path between Trust und Risk perception did show a highly significant coefficient of .76, but the model did not display satisfactory fit indices (RMSEA = .061). Thus, the factor Risk had to be removed as well, and hypotheses four and five do not receive empirical support in our study. In order to test for hypothesis one through three, a third model was set up. The model is the same as in Figure 1 but without the factors risk perception and willingness to buy. The underlying assumption of this model is that the factors perceived size, perceived reputation, and perceived privacy determine trust. This model produced stable parameter estimates and reached acceptable fit indices. The completed model for the first sample with all standard errors, factor loadings, and path coefficients is depicted in Figure 2.