For PLS-SEM, a VIF of 5 or above indicates critical collinearity, and values should ideally be close to 3 or lower (Hair et al., 2019). In multiple regression, the older and more lenient rule is VIF below 10, equivalent to tolerance above .10 (Hair et al., 2019b). For common method bias in PLS-SEM, the full collinearity test uses a stricter cut-off of 3.3 (Kock, 2015). Pick the threshold that matches your analysis and cite it; do not mix them.
What VIF measures
The variance inflation factor tells you how much the variance of a regression coefficient is inflated because the predictor is correlated with the other predictors. It is computed by regressing each predictor on all the others: VIF = 1 / (1 − R²), and tolerance = 1 − R² = 1 / VIF. A VIF of 1 means no overlap at all. High VIF values make coefficients unstable: standard errors grow, signs can flip, and a predictor that matters may look non-significant.
Where the cut-offs come from
| Cut-off | Context | Source |
|---|---|---|
| VIF < 10 (tolerance > .10) | Multiple regression, traditional rule | Hair, Black, Babin & Anderson (2019) |
| VIF ≥ 5 critical; ideally ≤ 3 | PLS-SEM structural model and formative indicators | Hair, Risher, Sarstedt & Ringle (2019) |
| VIF > 3.3 | Full collinearity test for common method bias in PLS-SEM | Kock (2015) |
O'Brien (2007) warned that VIF rules of thumb, especially 10, are arbitrary: whether a given VIF is a problem depends on sample size and on the variance of the predictor. That is a good reason to cite a source for whichever value you use and to read VIF together with the coefficients, rather than treating it as a pass or fail.
Worked example
A student regresses Purchase Intention on four predictors: Price, Quality, Brand Image and Brand Trust. To compute VIF for Brand Image, she regresses Brand Image on the other three predictors and gets R² = .75.
VIF = 1 / (1 − .75) = 1 / .25 = 4.0; tolerance = .25.
Under the regression rule (VIF < 10), 4.0 is fine. Under the PLS-SEM guideline it is between 3 and 5: not critical, but worth noting. If R² had been .80, VIF would be 5.0, the critical level for PLS-SEM; at R² = .90, VIF would be 10. In this case the correlation between Brand Image and Brand Trust is .82, which explains most of the overlap. A sensible response is to check whether the two constructs are discriminant (HTMT), and if theory treats them as distinct, keep both and report the VIF; if they measure the same idea, combine them.
What to do if VIF is high
- Find the pair. High VIF usually comes from two predictors that correlate strongly. Look at the correlation matrix.
- Check for construction artefacts. Including a variable and its square, or an interaction without centring, inflates VIF mechanically. Mean-centre before forming products.
- Revisit the model. Merge redundant constructs, or model them as dimensions of a higher-order construct if theory supports it (Hair, Risher, et al., 2019).
- Do not simply delete a predictor your hypotheses need. Dropping it changes the research question.
Common method bias with VIF
Kock (2015) proposed regressing a random variable on all latent variables in a PLS model. If any VIF exceeds 3.3, the model may be contaminated by common method bias. Many examiners accept this test alongside Harman's single-factor test. Report it as: "All full collinearity VIFs were below 3.3 (range 1.42–2.87), suggesting the model is free of common method bias (Kock, 2015)."
Reporting example
Multicollinearity was not a concern: VIF values ranged from 1.38 to 4.00, below the critical value of 5 (Hair, Risher, et al., 2019), with the highest value for Brand Image.
Run it free
The free survey data analysis tool reports VIF and tolerance for every regression predictor, and the PLS-SEM tool reports inner and outer VIF with the Hair, Risher, et al. (2019) cut-offs.
FAQ
Which VIF threshold should I use in a regression thesis?
Many supervisors accept 10 (Hair, Black, et al., 2019), but reporting that all values are below 5 is stronger. Name your source either way.
Is VIF the same as tolerance?
They carry the same information: tolerance is 1/VIF. A tolerance of .20 equals a VIF of 5.
Can VIF be too low?
No. A VIF near 1 simply means a predictor is unrelated to the others.
Why does my VIF in SmartPLS differ from SPSS?
SmartPLS computes it on latent variable scores, SPSS on whatever composite you built (for example, item means), so values differ slightly.
References
- Hair, J. F., Black, W. C., Babin, B. J., & Anderson, R. E. (2019). Multivariate data analysis (8th ed.). Cengage Learning.
- Hair, J. F., Risher, J. J., Sarstedt, M., & Ringle, C. M. (2019). When to use and how to report the results of PLS-SEM. European Business Review, 31(1), 2–24. https://doi.org/10.1108/EBR-11-2018-0203
- Kock, N. (2015). Common method bias in PLS-SEM: A full collinearity assessment approach. International Journal of e-Collaboration, 11(4), 1–10. https://doi.org/10.4018/ijec.2015100101
- O'Brien, R. M. (2007). A caution regarding rules of thumb for variance inflation factors. Quality & Quantity, 41(5), 673–690. https://doi.org/10.1007/s11135-006-9018-6