Written and maintained by CASRAI Editorial Board
Published · Last updated
On this page
Multicollinearity does not bias your regression coefficients, and it does not mean your model is “wrong.” What it does is inflate the standard errors of the coefficients for the correlated predictors, which widens their confidence intervals, weakens their individual t-tests, and can make coefficient signs flip or swing wildly between similar samples — even while the model’s overall fit and its predictions stay perfectly valid. That distinction is the one most searches for this topic are actually trying to resolve: a high variance inflation factor (VIF) on a predictor is a warning about how confidently you can interpret that predictor’s coefficient, not a verdict on the regression as a whole.
What Multicollinearity Is
Multicollinearity occurs when two or more predictor variables in a multiple regression model are highly linearly correlated with each other. It is a property of the predictor set, not of the outcome variable or the relationship being modeled. It exists on a spectrum — some degree of correlation among predictors is normal and harmless; the practical concern starts when that correlation is strong enough to make the model’s coefficient estimates unstable.
A useful distinction: perfect (or near-perfect) multicollinearity — one predictor is an exact or near-exact linear combination of others (e.g., including both “height in cm” and “height in inches,” or including a subtotal alongside the components that sum to it) — makes the regression mathematically unsolvable or numerically unstable, and statistical software will usually drop a variable or throw an error. Ordinary multicollinearity — strong but imperfect correlation, which is the far more common real-world case — is what the rest of this guide covers.
What It Does — and Doesn’t Do — to Your Model
It helps to be precise about which parts of a regression output multicollinearity actually touches:
- It does inflate standard errors for the correlated predictors, which widens their confidence intervals and lowers their t-statistics, making it harder to detect a real effect for those specific variables (a Type II error risk).
- It does make coefficients unstable across samples or when a variable is added or dropped — small changes in the data can produce large, sometimes sign-flipping, changes in the coefficient estimates for the correlated predictors.
- It does not bias the coefficient estimates. Under the standard OLS assumptions, the coefficients remain unbiased even in the presence of multicollinearity; the estimator is still, on average, centered on the true population value — it is just less precise.
- It does not reduce the model’s overall predictive accuracy or R². If your goal is prediction rather than interpreting individual predictors, multicollinearity among the predictors is largely irrelevant, because the combined effect of the correlated predictors on the outcome is still estimated well.
- It does not affect predictors that aren’t part of the correlated cluster. Multicollinearity is a pairwise-and-beyond property among specific predictors; a variable uncorrelated with the others in the model is unaffected.
Variance Inflation Factor (VIF): Definition and Formula
The variance inflation factor is the standard diagnostic for quantifying how much multicollinearity is inflating the variance (and therefore the standard error) of a given predictor’s coefficient, relative to what it would be if that predictor were uncorrelated with the others in the model.
For predictor Xj, VIF is calculated by regressing Xj on all the other predictors in the model, taking the R² from that auxiliary regression (call it Rj²), and computing:
VIFj = 1 / (1 − Rj²)
The logic: if Xj is well explained by the other predictors (high Rj²), the denominator shrinks toward zero and VIF climbs. A VIF of 1 means Rj² = 0 — that predictor is completely uncorrelated with the rest of the model and its standard error is not inflated at all. A VIF of 5 means Rj² = 0.80 for that predictor against the others; a VIF of 10 means Rj² = 0.90. The related statistic tolerance is simply 1/VIF and is reported by some software instead of, or alongside, VIF itself.
VIF is calculated per predictor, not once for the whole model — a regression with six predictors produces six VIF values, and it is entirely normal for some to be low while one or two are elevated.
VIF Threshold Table: Why 5 and 10 Are Conventions, Not Laws
There is no statistical test that defines a single correct cutoff for “too much” multicollinearity. The commonly cited thresholds are methodological conventions that have become standard practice through repetition in textbooks, not derived rules with a fixed significance level behind them:
| VIF | Corresponding Rj² | Conventional interpretation |
|---|---|---|
| 1 | 0 | No correlation with other predictors — no inflation. |
| 1 – 5 | 0 – 0.80 | Moderate correlation; generally considered acceptable in most applied work. |
| 5 – 10 | 0.80 – 0.90 | Elevated; worth investigating, and a cutoff of 5 is also commonly used as the point at which to start paying attention (Sheather, 2009, as cited in the Wikipedia VIF article). |
| > 10 | > 0.90 | The most commonly cited “high” threshold, attributed to Kutner, Nachtsheim & Neter (2004), Applied Linear Statistical Models, as cited in the Wikipedia VIF article. |
Two things matter more than memorizing a single cutoff. First, the cutoff in use varies: O’Brien (2007) records that a VIF of 10, or as low as 4 (tolerance 0.10 or 0.25), is used as a rule of thumb, so an analysis that needs precise coefficient interpretation may reasonably apply the stricter end. Second, methodologists including O’Brien (2007, Quality & Quantity) have argued explicitly against treating any fixed VIF cutoff as a rule, on the grounds that how much variance inflation is “too much” depends on the sample size, the size of the effect you are trying to detect, and how much you actually care about that individual coefficient versus the model’s overall predictive performance. Report the VIF values themselves alongside whatever cutoff convention you apply, rather than only reporting a pass/fail against a threshold.
Worked Example: Reading a Coefficient Table Before and After
The numbers below are an illustrative composite, constructed to show a realistic pattern — not a citation to a specific published dataset or study. Consider a regression predicting resting metabolic rate from three predictors: age, lean body mass, and total body weight. Lean body mass and total body weight are strongly correlated with each other (heavier individuals in the sample tend to also carry more lean mass).
Model 1 — all three predictors included:
| Predictor | Coefficient (B) | Std. Error | p-value | VIF |
|---|---|---|---|---|
| Age | −4.1 | 1.3 | 0.003 | 1.2 |
| Lean body mass | 18.9 | 9.7 | 0.058 | 8.6 |
| Total body weight | −2.3 | 7.1 | 0.749 | 8.9 |
Read literally, this table is confusing: total body weight carries a negative coefficient, which contradicts the well-established physiological relationship between body mass and metabolic rate, and neither weight nor lean mass reaches conventional significance despite the model’s overall F-test being highly significant. The VIFs for lean body mass and total body weight (8.6 and 8.9) explain why — the two predictors are competing for the same variance, so the model cannot cleanly apportion the effect between them, and the standard errors on both are inflated accordingly.
Model 2 — total body weight dropped:
| Predictor | Coefficient (B) | Std. Error | p-value | VIF |
|---|---|---|---|---|
| Age | −4.0 | 1.3 | 0.003 | 1.1 |
| Lean body mass | 16.2 | 2.4 | <0.001 | 1.1 |
With total body weight removed, lean body mass’s standard error drops from 9.7 to 2.4, its coefficient becomes statistically significant, and its VIF falls to 1.1. Age’s estimate is essentially unchanged in both models, consistent with it not being part of the correlated pair.
The sentence a researcher would write in a results section: “Total body weight and lean body mass showed high collinearity (VIF = 8.6 and 8.9, respectively); total body weight was removed from the final model. In the reduced model, lean body mass was a significant positive predictor of resting metabolic rate (B = 16.2, SE = 2.4, p < .001), consistent with established physiology.”
Reading VIF as a Standard-Error Multiplier
Because VIF is a ratio of variances, its square root is a ratio of standard errors: a predictor with VIF = 4 has a coefficient standard error twice as large as it would be if that predictor were uncorrelated with the others, and a VIF of 5.27 corresponds to a standard error about 2.3 times larger (the worked figure given in the Wikipedia VIF article). Confidence intervals widen by the same factor, so a VIF of 10 stretches them by roughly 3.2 times. The data table below this guide sets out VIF, the implied auxiliary R², tolerance and the standard-error multiplier side by side for the values that published sources use as reference points.
Two properties help in interpretation. A VIF of 1 leaves the slope variance uninflated, and VIF does not change when a predictor is rescaled by a constant, so changing units (metres to centimetres) cannot create or remove a VIF problem. The same source states the underlying reason multicollinearity matters: when predictors are not orthogonal, each one’s slope estimate, standard error and p-value depend on which others are in the model (James et al., 2021, p. 116, as cited there).
Rules of Thumb and Who Cites Them
The most-cited paper on this question is O’Brien (2007), “A Caution Regarding Rules of Thumb for Variance Inflation Factors,” Quality & Quantity 41(5):673–690 (doi 10.1007/s11135-006-9018-6). Its abstract notes that a VIF of 10, or as low as 4, is used as a rule of thumb, and argues that such thresholds must be judged together with the other factors that affect the variance of a regression coefficient. It also cautions that common remedies, including ridge regression and combining variables into an index, can create problems of their own. In practice the VIF is one multiplicative term in the coefficient variance; the residual variance and the spread of the predictor (which grows with sample size) are the others, so a given VIF is less costly in a large sample than in a small one.
The practical reading is: cite the convention you applied (4, 5 or 10), name the source, and report the VIF values themselves so a reader can apply a different cutoff.
Generalized VIF (GVIF) for Categorical and Polynomial Terms
An ordinary VIF assumes each term has a single coefficient. A multi-level factor entered as several dummy variables, or a polynomial entered as X and X², has more than one coefficient, and its dummy columns are correlated with each other by construction. For these terms the R package car computes generalized variance inflation factors (GVIF; Fox & Monette, 1992, Journal of the American Statistical Association 87:178–183), which measure the inflation in the size of the joint confidence ellipse (or ellipsoid) for the term’s coefficients relative to what orthogonal data would give. Because GVIF grows with the number of coefficients in the term, car::vif also prints GVIF^(1/(2×df)), a value intended to be comparable across terms with different degrees of freedom. Per its documentation, the function works for linear, generalized linear, weighted and mixed models, so the same diagnostic applies to a logistic regression.
Detecting Multicollinearity Beyond VIF
VIF is the most commonly reported diagnostic, but it is not the only one, and checking more than one is good practice when the stakes of a wrong coefficient interpretation are high:
- Bivariate correlation matrix. A simple correlation coefficient matrix among predictors will surface obvious pairwise problems, though it misses multicollinearity that only shows up across three or more variables jointly.
- Tolerance. 1/VIF, reported by default in some software (notably SPSS); the same information as VIF on an inverse scale.
- Condition number warnings in software. The statsmodels OLS summary prints a note when the condition number is large; in its Longley example the warning reads that the condition number is large, 4.86e+09, which “might indicate that there are strong multicollinearity or other numerical problems.” The same example notes that values over 20 are worrisome (citing Greene, section 4.9), so treat a large condition number as a prompt to compute VIFs rather than as a verdict.
- Condition index / eigenvalues of the predictor matrix. A condition index above roughly 30, combined with two or more predictors loading heavily on the same low-eigenvalue dimension, is the classic Belsley, Kuh & Welsch (1980) diagnostic and can catch multivariate collinearity that pairwise correlations and even VIF can miss.
What to Do About It
There is no single correct fix — the right response depends on why the correlated predictors are both in the model and what the analysis is for:
- Remove one of the correlated predictors if it is conceptually redundant with another (as in the worked example above). This is usually the simplest and most defensible option when both variables are proxies for a similar underlying construct.
- Combine the correlated predictors into a single composite or index, or extract a principal component from them, if both carry information you don’t want to fully discard.
- Center predictors before creating interaction or polynomial terms. A large share of the multicollinearity found in practice is “structural” — it’s created by including X and X², or X1 and X1×X2, in the same model. Mean-centering the component variables before forming the product or power term removes most of this induced correlation without changing the substantive model.
- Increase the sample size where feasible. Multicollinearity inflates variance, and a larger sample shrinks variance from the other direction; it does not remove the underlying correlation, but it can bring the standard errors back down to a usable range.
- Use a regularized estimator such as ridge regression, which is specifically designed to stabilize coefficient estimates in the presence of correlated predictors, at the cost of introducing a small amount of bias.
- Do nothing, and say so explicitly if the model’s purpose is prediction rather than interpreting individual coefficients. Since multicollinearity does not harm overall predictive accuracy, forcing a fix in a pure-prediction context can do more harm than the “problem” it’s solving.
When Multicollinearity Doesn’t Matter
Multicollinearity is only a problem relative to what you are asking of the model. If the analysis goal is to predict the outcome as accurately as possible and no individual coefficient needs to be interpreted or reported on its own, elevated VIFs among the predictors can typically be left alone — the combined predictive contribution of the correlated block is still estimated well, even if it can’t be cleanly split between its members. The concern is specific to situations where a reader or reviewer needs to trust the size, sign, and significance of one particular predictor’s coefficient, which is the normal case in explanatory research (e.g., “does controlling for confounders, does X independently predict Y”) but not in a purely predictive model.
Assumptions and Scope
- VIF and the diagnostics above apply to multiple regression models — linear, logistic, and their extensions — wherever more than one predictor is entered simultaneously. A simple bivariate regression with a single predictor cannot have multicollinearity by definition.
- Multicollinearity is a property of the predictors, not the outcome, and is unrelated to whether the outcome variable is normally distributed or how it was measured.
- It is distinct from confounding: a confounder biases the estimated relationship between a predictor and the outcome by being related to both; multicollinearity affects the precision of coefficient estimates without biasing them, and can occur even when none of the correlated predictors confounds the outcome relationship.
- It is also distinct from skewed predictor distributions, heteroscedasticity, and autocorrelation — each is a separate regression diagnostic with its own test and its own fix, and a model can have any combination of them independently.
How to Calculate VIF in R, Python and SPSS
VIF for predictor j is 1 / (1 − R²j), where R²j comes from regressing that predictor on all the other predictors (the auxiliary regression). Every major package computes it for you, so you rarely need to run the auxiliary regressions by hand.
| Software | How to get VIF | What to know |
|---|---|---|
| R | car::vif(model) |
For terms with more than one degree of freedom (a multi-level factor), it returns a generalized VIF (GVIF) with the adjusted value GVIF^(1/(2×df)), which is intended to be comparable across terms of different dimension (car documentation). |
| Python | variance_inflation_factor(exog, exog_idx) from statsmodels.stats.outliers_influence |
Returns the VIF for one column at a time, so loop over the columns of the design matrix. The function takes the design matrix containing all explanatory variables plus the column index; standardize defaults to True, and it does not keep the auxiliary regression (statsmodels documentation). To fit the model itself, see statsmodels OLS. |
| SPSS | Linear Regression > Statistics > Collinearity diagnostics | Adds Tolerance and VIF columns to the Coefficients table; see multiple regression in SPSS. |
Read each VIF against the thresholds above and in light of what you need the model for. A high VIF on a control variable you do not need to interpret is usually harmless, while a high VIF on your main predictor of interest is worth acting on.
If you need to keep correlated predictors, two common remedies are penalized estimators such as ridge or LASSO regression and principal component regression. Because multicollinearity works through the coefficient variance, it also helps to understand the standard error of the regression it inflates.
Frequently Asked Questions
What is a good VIF value?
A VIF of 1 indicates no multicollinearity. Values up to about 5 are generally treated as acceptable in most applied work; values between 5 and 10 warrant a closer look; values above 10 are the most widely cited threshold for a genuine problem. These are conventions, not statistical cutoffs with a fixed p-value behind them, and some fields apply stricter thresholds — report the actual VIF, not just whether it cleared a cutoff.
Does multicollinearity affect R²?
No. Multicollinearity affects the standard errors and stability of individual coefficient estimates, not the model’s overall R² or its predictive accuracy.
Can you have multicollinearity with only two predictors?
Yes. Two highly correlated predictors in an otherwise simple model can produce high VIFs for both; multicollinearity is not exclusive to models with many predictors, though the risk of it going undetected does increase as more predictors are added.
Is a VIF of 10 always disqualifying?
No. VIF > 10 is a widely used caution flag, not a rule that invalidates a model. Whether it matters depends on whether you need to interpret that specific predictor’s coefficient, your sample size, and the size of the effect you’re trying to detect. A high VIF in a model built purely for prediction is often not worth acting on at all.
What’s the difference between VIF and tolerance?
Tolerance is simply 1/VIF. A tolerance of 0.10 is equivalent to a VIF of 10. Some statistical software (SPSS in particular) reports tolerance by default alongside or instead of VIF; they carry identical information on an inverted scale.
Is a VIF of 5 acceptable?
It depends on the convention you adopt. A cutoff of 5 is commonly used (Sheather, 2009), and O’Brien (2007) records rules of thumb of 4 and 10; a VIF of 5 corresponds to an auxiliary R² of 0.80 and a standard error about 2.2 times larger than with no correlation. Whether that is acceptable depends on the coefficient’s importance and the sample size.
What is the square root of VIF?
It is the factor by which a coefficient’s standard error is larger than it would be if the predictor were uncorrelated with the others. A VIF of 4 means the standard error doubles.
How do you compute VIF for categorical variables?
Use generalized VIF. In R, car::vif returns GVIF and GVIF^(1/(2×df)) for terms with more than one degree of freedom, and the adjusted value is the one to compare across terms.
Does VIF work for logistic regression?
Yes. car::vif accepts generalized linear models, so it can be applied to a fitted logistic model.
What condition number indicates multicollinearity?
There is no single cutoff. The statsmodels OLS example warns on a very large condition number (4.86e+09 in the Longley data) and quotes values over 20 as worrisome (Greene); the Belsley, Kuh & Welsch convention above uses a condition index above roughly 30.
Sources
Data: VIF values, implied R-squared, tolerance and standard-error inflation, with the cutoffs published sources attach to them
Compiled by applying the VIF = 1/(1 - R^2) definition to the VIF values that O'Brien (2007), Kutner et al. (2004) and Sheather (2009) are cited as using as rules of thumb; R^2, tolerance and sqrt(VIF) are computed from that definition, the final column records which source the cutoff is attributed to.
Period covered: 2004/2009.
| Row | VIF | Auxiliary R-squared of predictor on others | Tolerance (1/VIF) | Standard error multiplier (sqrt VIF) | Source attribution |
|---|---|---|---|---|---|
| VIF 1 | 1 | 0 | 1 | 1 | No inflation (Wikipedia VIF article) |
| VIF 4 | 4 | 0.75 | 0.25 | 2 | Rule of thumb for serious multicollinearity cited by O'Brien (2007) |
| VIF 5 | 5 | 0.8 | 0.2 | 2.24 | Cutoff 'commonly used' (Sheather 2009, via Wikipedia) |
| VIF 5.27 | 5.27 | 0.81 | 0.19 | 2.3 | Worked example in Wikipedia article: SE about 2.3 times larger |
| VIF 10 | 10 | 0.9 | 0.1 | 3.16 | High multicollinearity (Kutner et al. 2004, via Wikipedia); also O'Brien (2007) rule of 10 |
Sources
- Variance inflation factor (Wikipedia) (accessed 2026-11-17)
- O'Brien (2007) A Caution Regarding Rules of Thumb for Variance Inflation Factors, Quality & Quantity 41(5) (accessed 2026-11-17)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







