Meaning
Statistical modeling requires that explanatory variables in a regression analysis be independent to ensure the reliability of estimated coefficients. The phenomenon of variance inflation occurs when high correlation between predictor variables artificially expands the standard errors of these estimates. This effect is measured on a cardinal scale known as the variance inflation factor, where values above a certain threshold signal that multicollinearity is undermining the model’s reliability.
The metric is not calculated when analyzing single-variable models or when predictors are orthogonal by design.
Estimate Distortion
When independent variables share too much information, the regression model struggles to isolate the individual effect of each predictor. This confusion leads to unstable coefficients that fluctuate widely with minor changes in the data. Under variance inflation, true effects may be masked by the large standard errors, causing researchers to draw incorrect conclusions.
The model loses its predictive and explanatory power.
Diagnostic Threshold
Identifying this problem involves calculating the diagnostic metric for each independent variable. Most practitioners use a baseline threshold of five or ten to flag problematic variables. When a variable exceeds this limit, action must be taken to resolve the overlap.
This procedure is standard in predictive modeling.
Remedial Action
Resolving multicollinearity often requires removing one of the highly correlated variables from the analysis. Alternatively, researchers can combine the overlapping variables into a single composite index or use regularization techniques like ridge regression. These methods reduce the variance of the estimates at the cost of some bias.
The choice depends on the goals of the study.
