Multicollinearity occurs when two or more regression predictors are strongly correlated, making their separate effects difficult to estimate reliably. It can inflate explained variance because the model effectively counts overlapping information more than once. section 5.1 section 5.7.4
Check and respond to multicollinearity during regression development:
- Inspect correlations among continuous or numeric variables before fitting or refining the model. section 5.1
- Use car::vif() to calculate the variance inflation factor for each predictor. section 5.1 section 5.7.4
- When a predictor’s VIF exceeds 5, remove offending variables and fit a reduced model to improve validity and reliability. section 5.7.4 section 5.9
NBA hustle-statistics regression
A multiple linear regression tested hustle statistics as predictors of regular-season wins. VIF values ranged from 1.06 to 3.05, so none exceeded the threshold of 5 and no predictor required removal for multicollinearity. section 5.7.4 section 5.9
vif(fit1) ## screen_assists_pts deflections loose_balls ## 1.155995 1.062677 1.314189 ## contested_2pt contested_shots ## 3.052934 2.637588
Use the displayed VIF listing to compare predictors and identify values that may warrant a reduced model. section 5.7.4
