Before testing the moderating effects, the two predictor variables (social support and family function) were standardized to reduce problems associated
with multicollinearity between the interaction term and the main effects (Frazier et al., 2004).
We did not include age as a covariate due to the high correlation between age and generation (r =.83) and the possible problems
with multicollinearity.
In spite of significant intercorrelations among some of the predictor variables, none of these correlations approached the level -LRB-.70 or higher) suggestive of significant problems
with multicollinearity (Tabachnick & Fidell, 1996, p. 86).
Not exact matches
To deal
with the issue of
multicollinearity in the data, all variables were examined using the Variance Inflation Factor (VIF) and Tolerance in SPSS.
Results from these tests showed that the assumption of
multicollinearity was not violated
with VIF values less than 1.25 and Tolerance values above 0.93.
We were able to estimate a stable model, though, by constraining ORs for particular types / numbers of disorders to be the same in predicting both planned and unplanned attempts unless the interaction of the predictor
with plans in the pooled model was significant at the α level of.05 and had an estimated variance inflation factor (a diagnostic test suggesting that a regression coefficient might be affected by
multicollinearity) of less than 10.0.
However, the
multicollinearity due to high comorbidity among respondents
with suicide plans made it impossible to estimate a model
with all 17 type / number coefficients separately for ideators
with and without a plan.
However, analysis of regression structure coefficients (child report of adherence rs =.67, parent report of adherence rs =.59), which are not suppressed or inflated by collinearity, demonstrates that beta weights for adherence are low because of
multicollinearity between predictors, not poor relations
with the outcome variable.