Charles Blonstein, Alumnus, Fielding's School of Psychology (2004)
My dissertation study (Blonstein, 2004) supported original hypotheses that the Social Stress-Depression Relationship is moderated by Social Anxiety and jointly mediated by Sense of Belonging, as well as Social Self-Efficacy. In a high Social Anxiety subgroup (n=60), Belonging relative to Efficacy transmitted nearly eight times as much of the variance between predictors and criterion (i.e., ratio of the variables' squared Beta weights). An integrative reformulation of statistical suppression processes (see Friedman and Wall, 2005) holds implications for explaining this enigmatic differential.
This poster describes a statistical transformation of the Efficacy variable using my original dissertation data. Efficacy scale items were assigned to either of two subscales, based on their partial correlations with the criterion variable. By varying cut points anywhere along the continuum of partial values (r= -1.0 to +1.0), suppressor-enhancer dyads were constructed. Ultimately, I dichotomized the 24 full-scale items into 14- and 10-item, mutually exclusive subsets. Respectively, this combination reflected a ratio of 14 more negatively valenced partial values vis-a-vis 10 more positively valenced counterparts. This 14-10 variant was designed to optimize predictor-criterion mediation effects via reciprocal: (1) suppression (or statistical control) of variance that is irrelevant to Depression, and conversely (2) enhancement of variance that is relevant to this criterion. Hypotheses evaluated were that the above procedures would: (1) identify variance previously obfuscated by collinearity; (2) explain the magnitude differential between Efficacy versus Belonging effects; as well as (3) generate improvements in model fit indices.
In sum, numerous path analyses and other multivariate procedures yielded convergent findings. Relative to the original Efficacy variable, its transformed counterpart yielded stronger support for prior mediation hypotheses. Improved outcomes were evidenced by indices including: (1) R-Square for the criterion variable; (2) percent mediation between predictors and criterion; as well as (3) model fit criteria. Consistent with recent methodological trends (e.g., Kraha et al.; 2012), I discuss the interpretive value of identifying the multiple sources of variance in a data set.