response rates, as the purpose was not to analyze the aggregate results, but rather to study the
data on a stratum basis (Groves et al., 1992).
The data from the questionnaire were first factor analyzed in order to produce the
variables that were subsequently related with usage behavior through ordered logistic
regressions. Factor analysis was used to reduce the items of the questionnaire to a smaller set
representing the cognitive and affective variables. Ordered logistic regression was considered
to be the appropriate instrument to analyze the influence of affective and cognitive
determinants of usage, due to the ordered and discrete nature of the outcome variable. The
objective was to have the most parsimonious model that best describes the data (Hosmer and
Lemeshow, 1989). When the model contained the variables that were significant in the
correct functional form, the model’s goodness-of-fit was assessed with measures such as the
likelihood ratio test, the Wald test, the deviance, the Akaike Information Criterion, and the
Bayesian Information Criterion.
Usage frequency was operationalized as customers’ frequency of use (ranging from
“Never”, “Less than once a month”, “One/two times a month”, “One/two times a week” and
“Most days”) of the different delivery channels (branch network, debit card network,
telephone-based and Internet-based access). The definition of this dimension of usage follows
Ram and Jung (1991) and Zaichkowsky (1985) who defined usage frequency as how often
the product was used or the different applications for which the product was used.
DATA ANALYSIS
Factor Analysis
A factor analysis with orthogonal rotation was performed, as the objective in this exploratory
research was to produce uncorrelated constructs. This is considered adequate when the
objective is to reduce the number of original variables to a smaller set of uncorrelated
variables for subsequent use in prediction techniques, which was the case. Among the
orthogonal rotation methods, Varimax was used. The factor analysis conducted with the four
strata of respondents produced some differences across strata, as can be observed in Table 1:
14