VIF and the Multicollinearity Threshold: When to Worry About Your Regression

The reviewer’s comment is three lines long: “The VIF for BMI is 8.4. Please address multicollinearity.” Your regression has age, sex, BMI, and waist circumference predicting fasting glucose, and BMI is not even the variable you care about. The question is whether a VIF of 8.4 crosses a threshold that should worry you, or whether the honest answer is that this multicollinearity does not affect the coefficient your paper is about.

Both answers are defensible depending on which coefficient carries your argument. Working out which one applies takes three checks, and none of them is a lookup against a cutoff.

What VIF Measures

The variance inflation factor for predictor j is computed by regressing that predictor on all the others and taking

where \(R^2_j\) is the proportion of variance in predictor j explained by the remaining predictors. So VIF is not a separate diagnostic quantity — it is a restatement of how predictable one predictor is from the rest.

The number practitioners actually need is the square root. \(\sqrt{VIF_j}\) is the factor by which the standard error of that coefficient is inflated relative to a hypothetical version of the model where the predictor was uncorrelated with the others.

Every point on that curve is \(\sqrt{VIF}\) from the table above, so you can check any of them by hand. Reading the reviewer’s number off it: VIF 8.4 means the BMI coefficient’s standard error is about 2.9 times larger than it would be if BMI were uncorrelated with age, sex, and waist circumference. Its confidence interval is roughly three times wider. Nothing else about the model has changed — the coefficient is still unbiased, the fitted values are still correct, and \(R^2\) for the model as a whole is unaffected.

Which VIF Threshold Should Worry You in a Regression?

The commonly quoted cutoffs are VIF > 5 (moderate concern) and VIF > 10 (serious). They are rules of thumb, not tests. There is no null distribution being evaluated and no α being controlled, which is why you can find defensible published advice at 2, at 2.5, at 5, and at 10 without any of it being wrong.

Paul Allison, who has written the standard applied texts on regression, sets his own bar lower than most: he becomes concerned above VIF 2.50, corresponding to an \(R^2_j\) of 0.60. That is stricter than the usual advice, and it is consistent with the table — at VIF 2.5 the standard error is already inflated 58%.

The reason a single cutoff cannot do the job: whether inflated standard errors matter depends entirely on whether the affected coefficient is one you interpret, and whether the inflation is structural or substantive. Those are the next two checks.

Check 1: Which Coefficients Do You Actually Interpret?

Multicollinearity inflates the standard errors of the collinear predictors. It does not inflate the standard errors of predictors that are not part of the collinear set, and it does not affect the model’s predictions.

Allison’s first situation where high VIF can be ignored is exactly the case in the scenario above: collinear variables that appear only as controls. If your argument is about a treatment indicator with VIF 1.2, and BMI and waist circumference are adjustment variables you never interpret, their mutual collinearity costs you nothing you were going to use. Dropping one of them to make a diagnostic table look tidier would change the adjustment and therefore change the estimate you do care about — a worse trade.

Write down, before touching the model, which coefficients your Results section reports and interprets. If none of them has a high VIF, you are done, and the response to the reviewer is an explanation rather than a re-analysis.

Check 2: Is the Collinearity Structural?

Two common model features generate high VIFs mechanically, and in both cases the inflation is an artefact of how the terms are constructed rather than a problem with your data.

Polynomial and interaction terms. If the model contains \(x\) and \(x^2\), or \(x\), \(z\), and \(x \times z\), those terms are correlated by construction. The p-value for the interaction or quadratic term is not affected by this, and the collinearity carries no inferential penalty.

Dummy variables from a categorical predictor with a small reference group. Indicator variables from the same categorical variable are necessarily correlated, and the correlation is larger when the reference category is small. Allison’s worked case: with 45% never married, 45% married, and 10% formerly married, the VIFs for the married and never-married indicators will be at least 3.0 by construction. The joint test that all indicators equal zero is unaffected.

Check 3: Choosing a Remedy

If a coefficient you interpret has a high VIF and the collinearity is substantive rather than structural, then you are choosing between genuine options, all of which have costs.

• Combine the correlated predictors. BMI and waist circumference both measure adiposity. A single index, or one of the two chosen on theoretical grounds, is often the honest model. Say which you chose and why.
• Drop one, on subject-matter grounds. Defensible when the two variables are near-substitutes for the same construct. Not defensible when they are conceptually distinct and both belong in the causal structure — dropping a confounder to reduce a VIF trades a variance problem for a bias problem.
• Keep both and report the inflation. Legitimate, and often the most honest choice. Report the VIFs, state that the estimates are unbiased with wider intervals, and let the intervals speak.
• Collect more data. Standard errors shrink with n even at fixed VIF. Rarely available mid-review, but it is the only option that improves precision without changing the model.

One case that gets misfiled as collinearity: two predictors correlate because one of them lies on the causal path between the other and the outcome. Adjusting for a mediator does not remove confounding — it removes part of the effect you are trying to estimate, and the elevated VIF is a symptom rather than the problem. If that is the structure you have, the analysis you want is a mediation analysis with bootstrapped confidence intervals, not a collinearity remedy.

The University of Virginia’s StatLab write-up on addressing multicollinearity works through the diagnostics and the trade-offs with R code if you want to see the remedies applied to a worked dataset.

Three Fixes That Do Not Work

Centering does not remove multicollinearity. Centering \(x\) and \(z\) before forming \(x \times z\) does reduce the correlation between a predictor and its own product term, which is why it helps the structural case. It does not reduce the correlation between \(x\) and \(z\) themselves, so it does nothing at all for the substantive case. Centering is an interpretability aid — it makes main effects readable as effects at the mean of the other variable — and describing it as a collinearity fix is a claim it cannot support.

Stepwise selection does not resolve it either. Letting an algorithm drop whichever collinear predictor happens to lose on this sample produces a model whose specification is a function of sampling noise, and the reported p-values no longer mean what they claim.

Inspecting the pairwise correlation matrix is not sufficient. A predictor can have modest pairwise correlations with every other predictor and still be almost perfectly predicted by them in combination. That is precisely what VIF catches and a correlation matrix does not.

Writing the Response, and the Sample Size Behind It

A response that satisfies reviewers has the same three parts as the checks: state the VIFs, state which coefficients you interpret and whether they are affected, and state what you did and why. “VIFs for BMI and waist circumference are 8.4 and 7.9; both are adjustment variables and neither is interpreted. The exposure coefficient has VIF 1.2 and is unaffected. We report both adiposity measures because omitting either changes the adjustment set” is a complete answer, and it is shorter than a re-analysis. Our guide to reporting statistics in a methods section covers the surrounding conventions.

Underneath all of this sits a sample-size question that is easier to fix before data collection than after. The conventional planning rules are 10 to 15 observations per predictor for linear regression, and about 10 events per predictor for logistic regression — rules of thumb rather than derived limits, but useful because a model near those limits has unstable coefficients whether or not any VIF is elevated. Our free sample size and power calculator includes a multiple-regression mode that takes the number of predictors and your target power directly, and it can also run in reverse to report the minimum effect you could detect with the n you already have.

The reviewer’s comment is worth treating as a question about which coefficient your paper rests on. Answer that first, and the VIF number usually stops being the decision.