100 / 274
多元迴歸的自變數比較與多元共線性之影響：效果量、優勢性與相對權數指標的估計與應用
100
Furthermore, in addition to the technical aspects, the mechanisms behind the predictors
and the model deserve more attention. For example, the complex relations among the
predictors may reflect not only the confounding effects of predictors, but also the
possibilities of causal impact or interaction existing among the variables. Researchers have
to be aware of the effects in terms of the mediation as well as the moderation and then
incorporate them into the empirical examinations (Baron and Kenny, 1986; Hayes, 2013).
Finally, the high correlation between the predictors may be due to the fact that the two
predictors are almost the same, a situation where there is lack of discriminant validity instead
of the confounding effect among different predictors (Hair et al., 2006; Nunnally and
Bernstein, 1994). In this case, removing one of the highly correlated predictors or combining
the similar variables into a single predictor may be a better solution.
5. Contributions
Rather than simply relying on hypothesis testing and interpretations of regression
coefficients, this paper presents a comprehensive review on several effect size index of
regression models. Two recently proposed strategies for evaluating the relative importance of
predictors, the RWA and the DA, are introduced along with a list of traditional statistics such
as the correlation coefficient, the beta coefficient, the structure coefficient, and the product
measures. The major contribution is to examine the impacts of multicollinearity, including
the enhancement, suppression, and redundancy effects, on the evaluation of the effect sizes
and several statistics of relative importance of predictors. The results from the simulation and
empirical study support that the statistics based on the RWA and the DA are recommended
for evaluating the relative importance of predictors.