Understand Variance Inflation Factor (VIF) in Regression Analysis

Search Fintrev

Search our guides, reviews and glossary.

The variance inflation factor (VIF) measures how much the variance of an estimated regression coefficient is increased because that independent variable is correlated with the other independent variables in the model. It is the standard test for multicollinearity in multiple regression. A VIF of 1 means the variable shares no linear overlap with the other predictors. A VIF of 5 means the variance of its coefficient estimate is five times what it would be with no overlap.

What VIF measures: the multicollinearity problem

Multiple regression estimates the effect of several inputs (independent variables) on one outcome (the dependent variable). Each coefficient is meant to show the effect of one input while the others are held constant. That interpretation works cleanly only when the inputs carry separate information.

Multicollinearity exists when one or more independent variables are linearly related to the others. When two predictors move closely together, the model has little independent variation to tell their effects apart.

The consequences show up in the coefficients, not in the overall fit:

• Larger standard errors. Coefficient estimates become less precise, so variables that matter can appear statistically insignificant.
• Unstable estimates. Small changes in the sample or the model specification can produce large, erratic swings in coefficients, sometimes flipping their signs.
• Ambiguous attribution. The model may explain the outcome well as a whole while being unable to say which input is doing the work.

This matters in econometrics and finance research, where many models exist precisely to test the relationship between a specific input and an outcome. Multicollinearity undermines that test while leaving the headline R² looking healthy.

The VIF formula

For the ith independent variable:

VIFᵢ = 1 / (1 − Rᵢ²)

where Rᵢ² is the unadjusted coefficient of determination from an auxiliary regression of variable i on all the remaining independent variables. The dependent variable plays no part in the calculation. VIF describes relationships among the predictors only.

The denominator, 1 − Rᵢ², is called tolerance. Tolerance is the share of a variable’s variation that the other predictors cannot explain. VIF is its reciprocal, so a tolerance of 0.2 corresponds to a VIF of 5, and a tolerance of 0.1 to a VIF of 10.

The name comes from what the number does to the coefficient’s sampling variance. The variance of the estimated coefficient equals the variance it would have with uncorrelated predictors, multiplied by VIF. Because standard error is the square root of variance, the standard error is inflated by the square root of VIF. A VIF of 4 doubles the standard error. A VIF of 9 triples it.

How to calculate VIF step by step

1. Pick one independent variable and treat it as the outcome.
2. Regress it on all the other independent variables (with an intercept).
3. Record the R² from that auxiliary regression.
4. Compute 1 / (1 − R²).
5. Repeat for every independent variable in the model.

How to interpret VIF values

Because R² ranges from 0 to 1, VIF ranges from 1 to infinity.

• VIF = 1: Rᵢ² is 0. The variable is uncorrelated with the other predictors, and no multicollinearity affects its coefficient.
• VIF from 1 to 5: moderate correlation. Usually acceptable, depending on the purpose of the model.
• VIF above 5: high correlation. The coefficient deserves scrutiny.
• VIF above 10: commonly treated as significant multicollinearity that needs correcting. At this level, more than 90% of the variable’s variation is explained by the other inputs.

These cutoffs are conventions, not statistical tests, and a stricter guideline holds that a VIF of 3 or below is not a cause for concern. Whether a given value is a problem depends on whether you need to interpret the coefficient on that variable.

Worked example (hypothetical)

An economist wants to test whether the unemployment rate has a statistically significant relationship with the inflation rate. To control for labor-market conditions, the model also includes new initial jobless claims. Both series track the same underlying labor-market weakness, so they tend to rise and fall together.

Suppose regressing the unemployment rate on jobless claims and the other predictors produces an R² of 0.90.

VIF = 1 / (1 − 0.90) = 1 / 0.10 = 10

The variance of the unemployment coefficient is ten times what it would be with uncorrelated inputs, and its standard error is about 3.16 times larger (the square root of 10). The model could show strong explanatory power overall yet fail to show whether the effect on inflation comes mainly from unemployment or from jobless claims.

Compare a variable with an auxiliary R² of 0.40: VIF = 1 / 0.60 ≈ 1.67, and the standard error is inflated by only about 29%. That variable’s coefficient can be read with reasonable confidence.

The same pattern appears in stock research. A model explaining returns with both the P/E ratio and earnings yield (its inverse) would show high VIFs for both, because the two measure the same valuation information in different forms.

What to do about a high VIF

The right fix depends on the hypothesis being tested.

• Drop a variable. If two predictors carry redundant information, removing one often resolves the problem. In the example, the economist might keep the unemployment rate and remove jobless claims.
• Combine variables. Averaging, summing or building an index from closely related inputs captures their joint effect in one term.
• Use principal components analysis (PCA). PCA transforms correlated predictors into a smaller set of uncorrelated components.
• Use partial least squares (PLS) regression. PLS builds new uncorrelated variables that are also chosen for their relationship with the outcome, instead of using ordinary least squares on the raw inputs.
• Apply regularization. Ridge regression shrinks coefficients and stabilizes estimates when predictors overlap. This is common in machine learning, where the goal is prediction rather than interpreting individual coefficients.
• Center variables before building interaction or squared terms. A term like X × Z, or X², is mechanically correlated with X. Subtracting the mean before multiplying usually lowers these VIFs without changing the substance of the model.

Sometimes the right response is no response. If the high-VIF variables are controls whose coefficients you will not interpret, and the variable of interest has a low VIF, the multicollinearity does not affect your conclusion.

Limitations of VIF

• Thresholds are arbitrary. A VIF of 6 in a large sample may matter less than a VIF of 4 in a small one, since sample size also drives standard errors.
• It does not name the partner variable. VIF shows that a variable overlaps with the others, not which ones. Inspecting the auxiliary regression or a correlation matrix fills that gap.
• It captures linear relationships only. Strong nonlinear dependence between predictors can go undetected.
• Categorical variables need care. A set of dummy variables for one category can produce high VIFs among themselves. A generalized VIF, which treats the set as one unit, handles this better.
• It says nothing about predictive accuracy. A model with high VIFs can still forecast well if the correlation structure among inputs stays the same in new data.

VIF vs. related measures

VIF vs. pairwise correlation. A correlation matrix shows how pairs of variables move together. VIF captures how one variable relates to all the others at once. Three variables can each have modest pairwise correlations while one is almost a linear combination of the other two. VIF catches that case. The correlation matrix does not.

VIF vs. R² of the main model. The model’s R² measures how well all the predictors explain the dependent variable. The Rᵢ² inside the VIF formula measures how well the other predictors explain one predictor. A high model R² is usually welcome. A high auxiliary R² is the warning sign.

FAQs

The Bottom Line

VIF matters most to analysts who need to interpret individual coefficients, such as economists and finance researchers testing a specific input against an outcome. A high VIF calls for scrutiny or a fix like dropping a redundant variable, though it can be left alone when it affects only control variables. Because VIF misses nonlinear links and does not name the overlapping partner, pairing it with a correlation matrix or auxiliary regression gives a fuller picture.