Mathwords logoMathwords

Variance Inflation Factor — Definition, Formula & Example

The variance inflation factor (VIF) is a number that tells you how strongly one predictor in a multiple regression overlaps with the other predictors. A VIF of 1 means no linear overlap. Larger values mean that predictor's slope estimate is less stable.

For predictor xjx_j in a multiple linear regression, VIFj=1/(1−Rj2)\text{VIF}_j = 1/(1 - R_j^2), where Rj2R_j^2 is the coefficient of determination from regressing xjx_j on all the other predictors. VIFj\text{VIF}_j is the factor by which the sampling variance of the estimated slope bjb_j exceeds what it would be if xjx_j were uncorrelated with the other predictors.

Key Formula

VIFj=11−Rj2\text{VIF}_j = \frac{1}{1 - R_j^2}
Where:
  • VIFj\text{VIF}_j = Variance inflation factor for predictor j. It is always at least 1, and undefined when the other predictors explain predictor j perfectly.
  • Rj2R_j^2 = R-squared from regressing predictor j on the other predictors. The response variable y is not used.

Worked Example

Problem: Software fits a model that predicts home price from three predictors. The output lists these VIFs: square footage 4, number of bedrooms 12.5, age of home in years 1.25. For each predictor, find how much of its variation the other predictors explain, and decide which one is a multicollinearity concern.
Step 1: Solve the VIF formula for the R-squared. This undoes what the software did.
Rj2=1−1VIFjR_j^2 = 1 - \frac{1}{\text{VIF}_j}
Step 2: Substitute each reported VIF.
sq ft: 1−14=0.75,bedrooms: 1−112.5=0.92,age: 1−11.25=0.20\text{sq ft: } 1 - \tfrac{1}{4} = 0.75, \quad \text{bedrooms: } 1 - \tfrac{1}{12.5} = 0.92, \quad \text{age: } 1 - \tfrac{1}{1.25} = 0.20
Step 3: Compare with the common rules of thumb. A VIF above 5 deserves attention, and a VIF above 10 usually signals serious multicollinearity. Bedrooms is above 10. Square footage is moderate, and age is fine.
Answer: The other predictors explain 75% of the variation in square footage, 92% in bedrooms, and 20% in age. Bedrooms (VIF 12.5) is the serious concern. Its slope estimate is imprecise, with a standard error about 3.5 times larger than with no overlap, because bedrooms is nearly predictable from the other variables.

Why It Matters

High VIFs make slope estimates vary much more from sample to sample. The standard error of bjb_j grows by a factor of VIFj\sqrt{\text{VIF}_j}, so confidence intervals for slopes widen and t-tests may miss real effects. Analysts in economics, public health, and real-estate pricing check VIFs before interpreting any individual coefficient.

Common Mistakes

Mistake: Thinking a high VIF means the predictor is strongly related to the response y.
Correction: VIF never uses y. It only measures how well the other predictors explain xjx_j. A predictor can have a huge VIF and still matter, or a VIF of 1 and be useless for predicting y.

Related Terms