The Impact of Cognitive versus Affective Aspects on Consumer Usage of Financial Service Delivery Channels

concluded that all variables were acceptable for face, content and discriminant validity
criteria.

Ordered Logistic Regressions

Model fitting efforts consisted of multivariate analyses of the data in order to explore and test
the hypotheses. By stratum, correlation and univariate analyses were also performed to best
understand the data, and to suggest a number of candidate variables to enter the multivariate
model.

For the logistic regressions, each variable’s estimated odds ratio, standard error, beta
coefficient, odds ratio confidence interval, and corresponding p-value are described, together
with the measures used to evaluate the model’s fit. In logistic regression, the variables’ beta
coefficients represent the impact of a marginal change of the independent variable over the
logit (i.e. the log odds) of the dependent variable. As sometimes, the effect over the logit is
not very useful or clear, it is preferable to interpret the impact of the independent variables in
terms of the odds. As such, both the odds ratio and the beta coefficient are presented. The
estimated odds ratio is the exponentiated estimated beta coefficient and represents the effect
of a marginal change of the independent variable over the odds of the dependent variable. For
example, an estimated beta coefficient of ‘desire for control’ of .29 or the estimated odds
ratio of 1.33 means that a marginal increase in desire for control will increase the likelihood
of more frequent usage of the channel vs. less frequent usage (or that the odds of a more
frequent usage of the channel are 33% higher for those who feel higher desire for control).

Care was also taken in relation to different types of data (numerical) problems that
interfere with model fitting by biasing the estimates and, consequently, leading to erroneous
conclusions. Namely, issues such as multi-collinearity, outliers and influential observations,
and complete discrimination were evaluated. When such problems were detected, the number
of observations was corrected, as suggested by Hosmer and Lemeshow (1989) and Hair,
Anderson, Tatham, and Black (1998). Additionally, for each model fitting, the ‘Proportional
Odds’ assumption was tested (the chi-square values for this test are presented by channel).
This is a critical assumption of the ordered logit model and states that the beta coefficients do
not vary according to the outcome category being considered (Borooah, 2001). Violation of
this assumption leads not to one model describing the outcome variable, but to k-1 models (k
being the number of categories of the outcome variable), each one representing a contrast
between one category of the outcome variable and the reference category. In such case, the

17

<<   15   16   17   18   19   20   21   >>

More intriguing information