Multiple regression is one of the most widely used techniques in quantitative research. It allows you to examine how several independent variables together predict an outcome. However, when two or more predictors are strongly related to each other, regression results can become unstable and difficult to interpret. This problem is called multicollinearity, and checking for it is a standard step before you report your regression findings.
This post explains what multicollinearity is, why it matters, how to detect it using common diagnostics, how to check it in SPSS, and what to do if you find it.
What is multicollinearity?
Multicollinearity occurs when independent variables in a regression model are highly correlated with one another. In extreme cases, one predictor can be almost perfectly predicted from the others. This is called perfect or near-perfect multicollinearity.
For example, if you include both "years of work experience" and "age" as predictors of salary, these two variables may be strongly correlated. The model then has difficulty separating the unique effect of each one.
Why does multicollinearity matter?
Multicollinearity does not necessarily reduce the overall predictive power of the model. The R-squared may still be high. The problem lies in the individual coefficients:
- Inflated standard errors: coefficients become less precise, so predictors that are actually important may appear non-significant.
- Unstable coefficients: small changes in the data or model can cause large changes in coefficient size or even sign.
- Confusing results: a predictor may have a strong correlation with the outcome but a weak or opposite-signed coefficient in the regression.
- Difficult interpretation: it becomes hard to say which predictor is responsible for the effect.
If your goal is only prediction, moderate multicollinearity may not be a serious concern. If your goal is to interpret the effect of individual predictors, as in most theses, it matters a great deal.
How to detect multicollinearity
1. Correlation matrix
Start by examining correlations among your independent variables. Very high correlations, often described in textbooks as above about 0.8 or 0.9, are a warning sign. However, a correlation matrix only shows relationships between pairs of variables. Multicollinearity can also arise from combinations of three or more predictors, so you need further diagnostics.
2. Tolerance
Tolerance shows how much of the variance in a predictor is not explained by the other predictors. It is calculated as 1 minus the R-squared from regressing that predictor on all the other predictors. Low tolerance values indicate a problem. A tolerance below 0.1 is commonly treated as a serious concern, and some authors suggest caution below 0.2.
3. Variance Inflation Factor (VIF)
VIF is the reciprocal of tolerance (VIF = 1 / tolerance). It shows how much the variance of a coefficient is inflated because of correlation with other predictors. Commonly cited rules of thumb are:
- VIF of 1: no correlation with other predictors.
- VIF above 5: often treated as a sign of potentially problematic multicollinearity.
- VIF above 10: widely treated as a sign of serious multicollinearity.
These thresholds are conventions rather than strict rules. Different textbooks recommend different cut-offs, so cite the source you follow and interpret VIF in the context of your sample size and research purpose.
4. Condition index and variance proportions
Collinearity diagnostics in SPSS also report eigenvalues, condition indices and variance proportions. A high condition index, often above 30, combined with two or more predictors having high variance proportions (for example above 0.5) on the same dimension, suggests that those predictors are collinear. This method is useful for identifying which specific variables are involved.
How to check multicollinearity in SPSS
- Go to Analyze > Regression > Linear.
- Move your dependent variable into the Dependent box and your predictors into the Independent(s) box.
- Click Statistics and tick Collinearity diagnostics.
- Click Continue and then OK.
In the output, look at the Coefficients table, which now includes Tolerance and VIF columns for each predictor. A separate Collinearity Diagnostics table shows eigenvalues, condition indices and variance proportions.
How to report multicollinearity checks
In your thesis or article, briefly state that you checked for multicollinearity, which diagnostics you used and what you found. For example:
"Multicollinearity was assessed using tolerance and variance inflation factor values. All tolerance values were above 0.40 and all VIF values were below 2.5, indicating that multicollinearity was not a concern."
Replace the numbers with your actual results, and cite the threshold you used.
What to do if you find multicollinearity
There is no single fix. Choose a solution that fits your research question:
- Check for errors: make sure you have not included the same variable twice, a total score together with its components, or a full set of dummy variables without a reference category.
- Remove a redundant variable: if two predictors measure very similar concepts, consider keeping the one that is more theoretically important.
- Combine variables: if several predictors measure the same underlying construct, consider combining them into a composite score, supported by factor analysis and reliability testing.
- Centre variables in interaction models: when you include interaction or polynomial terms, mean-centring the variables can reduce the correlation between the main effects and the product term, and makes coefficients easier to interpret.
- Collect more data: larger samples can reduce the standard errors, although this does not remove the correlation itself.
- Use alternative techniques: methods such as ridge regression or partial least squares are designed to handle correlated predictors, but they change how results are interpreted.
- Accept and explain: if the collinear variables are only controls and your main predictors are unaffected, you may decide to keep them and explain your reasoning.
Common mistakes to avoid
- Relying only on the correlation matrix.
- Applying VIF cut-offs mechanically without justification.
- Dropping theoretically important variables just to reduce VIF.
- Forgetting to check multicollinearity after adding interaction terms.
- Not reporting the multicollinearity check at all.
Final thoughts
Multicollinearity makes it difficult to separate the effects of correlated predictors and can lead to unstable, misleading coefficients. Check for it routinely using correlations, tolerance, VIF and collinearity diagnostics, report what you found, and respond with solutions grounded in theory rather than simply deleting variables. A short, clear multicollinearity check strengthens the credibility of your regression analysis.