and security risk contributed significantly only to the formative index of
the complete dataset. Although not
all facets of risky behavior contributed significantly, VIF was less than 2
within all three datasets, confirming
indicator validity. Consequently, low
redundancy of indicators’ information was confirmed.
The results show performance
risk contributed significantly to over-
all risk perception in dataset B and
C, while security risk contributed to
overall risk perception in only the
complete dataset C. Privacy risk did
not significantly contribute to per-
ceived risk, regardless of dataset. And
performance expectancy, effort expec-
tionnaires, we received a total of 402
Clustering for espoused cultural
values. We conducted exploratory factor analysis, a statistical method used
to uncover the underlying structure
in a large set of variables, to test the
“unidimensionality” of the measurement items for IC and UA. It revealed
three items measuring IC and two
items measuring UA load with a high
coefficient on the factors they are intended to measure (loadings > 0.79).
Using the factor scores of these items,
we then conducted a K-means clustering. Cluster analysis revealed two
clusters (see Table 1) where the first
cluster (referred to as A) encompassed
respondents with high IC scores (
cluster center 0.13) and low UA scores
(cluster center −0.61), and the second
cluster (referred to as B) encompassed
respondents with rather low IC scores
(cluster center −0.20) and rather high
UA scores (cluster center 0.93).
We characterized the respondents
in group A as more individualistic
and less risk-averse than in group B.
Although both groups encompass
only millennials, they showed different characteristics when it comes to
the formation of behavioral intention.
One could argue A is the more forthcoming, self-centered, and aggressive,
while B represents the more group-oriented and considerate.
Measurement model assessment.
We tested our model with partial
least squares using SmartPLS 3.0 with
1,000 samples bootstrapping, assessing the measurement model with the
complete dataset C, as well as with
clusters A and B.
We measured behavioral intention
reflectively (loadings of the indicators were above 0.95 and significant at
the .001 level) and confirmed internal
consistency by assessing Cronbach’s
Alpha (CA) and Composite Reliability (CR) measures. Both exceeded the
threshold of 0.90 for all datasets:
A CA=0.90, CR=0.95; B CA=0.92,
CR=0.96; C CA=0.91; and CR=0.96.
The average variance extracted was
greater than 0.50 (A 0.91; B 0.93; and
C 0.92), demonstrating sufficient
convergent validity. Finally, we used
cross-loading analysis to confirm that
all constructs load highest with their
The formative measures of “
perceived benefits” were significant, at
least at the .05 level, and path coefficients were greater than . 1, suggesting the chosen characteristics of each
category were relevant for the formation of the construct (see Table 2).
Moreover, the variance inflation factor
(VIF) was less than 2, supporting our
assumption for indicator validity.
The formative measures of “
perceived risks” revealed mixed results
regarding the risk facets’ contribution to the formation of the formative
index. We found privacy risk was not
relevant regardless of dataset used;
performance risk was significant only
in subset B and the complete dataset;
Figure 2. Structural model assessment.
*p < 0.05
**p < 0.01
***p < 0.001
Table 2. Formative constructs measurements.
Construct Facet Cluster A Cluster B Complete Set C
Performance 0.740 0.631* 0.713**
Privacy -0.315 0.133 -0.124
Security 0.668 0.530 0.613*
0.522*** 0.430** 0.491***
EffortExpectancy 0.325* 0.230* 0.292***
Compatibility 0.349** 0.538*** 0.420***
*** p < 0.001; p < 0.01; p < 0.05