Does VIF four mean standard error is four times larger?

No; the factor is two under the stated comparison.

AI Glossary / Statistics and probability / Intermediate

Variance inflation factor ( VIF )

Definition

The variance inflation factor, or VIF, measures how linear relationships among predictors inflate the variance of an estimated ordinary least squares regression coefficient under the standard model. For predictor j, VIF is 1/(1 minus Rj squared), where Rj squared comes from regressing that predictor on the other predictors. It multiplies coefficient variance, not standard error; the corresponding standard-error factor is the square root of VIF. This distinction is essential when interpreting the number.

What it means

When predictors carry overlapping information, separating their individual associations with the outcome becomes harder. A coefficient may therefore be estimated imprecisely even when the model predicts reasonably well overall. VIF quantifies one part of that precision problem under the ordinary linear-model assumptions. It does not measure predictive accuracy, prove a coefficient is biased or identify which variable should automatically be removed. The meaning of a predictor and the purpose of the model remain relevant.

How it works

For each nonconstant predictor, fit an auxiliary regression using the remaining predictors and an intercept, then compute its R squared. Insert that value into VIF = 1/(1 minus R squared). Under a homoscedastic full-rank linear model, the coefficient variance can be written as residual variance divided by the predictor's centered sum of squares, multiplied by VIF. The comparison holds residual variance and predictor spread fixed. Perfect linear dependence gives R squared equal to one and prevents separate ordinary least squares identification. Generalized VIFs and models without an intercept need their own definitions.

Three practical examples

Example 1: A regression includes two measurements of nearly the same usage behavior, making their separate coefficients unstable. Example 2: An AI demand model contains several overlapping activity indicators that jointly predict another predictor well. Example 3: A polynomial regression includes a variable and its square, producing coding-dependent collinearity that may be reduced by centering. These situations do not all call for the same remedy. Removing an essential adjustment variable can damage interpretation even if VIF falls.

Think of it this way

Think of trying to determine which of several nearly synchronized instruments contributed to a combined signal. If they move almost together, many allocations of influence fit similarly. VIF describes how that overlap enlarges uncertainty in separating one contribution under a linear model. It does not say the combined signal is necessarily uninformative.

Worked example

Suppose regressing predictor X1 on the other predictors gives R1 squared = 0.75. Then VIF1 = 1/(1 minus 0.75) = 4. Under the standard coefficient-variance formula, suppose the corresponding no-collinearity variance baseline, holding residual variance and X1 spread fixed, is 0.09. The coefficient variance is then 0.36, and the standard error is 0.60 rather than 0.30. Thus variance is inflated fourfold but standard error only twofold. The auxiliary regression uses X1 as its response, not the original outcome Y. A large overall outcome-model R squared is therefore not the same quantity as the R1 squared used in this calculation.

Illustrative case study

In this fictional case study, Sycamore Forecasting includes visits, sessions and active minutes in a regression. Their overlap produces high VIF values and unstable individual coefficients. Analysts review whether the goal is prediction or explanation, inspect the feature definitions and compare suitable alternatives, including combining redundant measures or regularizing. They do not discard variables solely to pass a numerical threshold. The report explains that coefficient uncertainty and predictive performance answer different questions.

Common misconceptions and mistakes

Common mistakes include saying VIF multiplies standard error directly, using the original model's R squared instead of the auxiliary regression's value, and assuming pairwise correlations reveal every collinearity problem. A predictor can be well explained by a combination of several others without any single pairwise correlation being extreme. Another mistake is treating a threshold such as five or ten as an automatic decision rule. VIF also does not correct heteroscedasticity, omitted variables, dependence or a poorly defined estimand.

Business use

Businesses should use VIF as one diagnostic when interpreting linear-model coefficients. Review uncertainty intervals, predictor definitions, sample size and the model's purpose. If stable prediction is the objective, evaluate out-of-sample performance and consider regularization rather than focusing only on coefficient separation. If explanation or adjustment is central, preserve substantively necessary variables and discuss the resulting uncertainty. Changing coding can change some VIF values without changing the fitted model's substantive information, so diagnostics require context.

Check your understanding

What VIF corresponds to auxiliary R squared 0.75?

Answer: Four.

What happens to a baseline variance of 0.09?

Answer: It becomes 0.36.

What is the corresponding standard error?

Answer: 0.60.

Which response is used in the auxiliary regression?

Answer: The predictor whose VIF is being calculated, not the original outcome.

Related terms

Tolerance
Related terms include multicollinearity
auxiliary regression
coefficient variance
standard error and regularization
so it is the reciprocal of VIF
while near-collinearity produces unstable or imprecise coefficient estimates

Learn next

Learn next how the inverse design matrix determines ordinary least squares coefficient uncertainty. Compare pairwise correlation checks with auxiliary regressions, and examine how centering affects polynomial terms. Then study regularization and substantive variable selection as different responses to unstable coefficients, each with its own interpretation.

Ordinary Least Squares

Sources

Sources: Penn State, Detecting Multicollinearity Using Variance Inflation Factors, https://online.stat.psu.edu/stat462/node/180/; University of Vermont, Statistical Methods for Psychology glossary, https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html. The coefficient-variance formula clarifies that VIF inflates variance, while its square root inflates standard error, correcting the source glossary's loose standard-error wording. The numerical example is original and independently checked; the forecasting case is fictional. The statsmodels primary implementation independently confirms the auxiliary-regression calculation: https://www.statsmodels.org/stable/_modules/statsmodels/stats/outliers_influence.html. Penn State was not retrievable during this review.

Previous: Variance estimation Next: Variance Sum Law