Multicollinearity is one of the most common reasons a regression model looks confusing. Your overall model is significant, yet none of the individual predictors are. Or a predictor you expected to matter suddenly has the “wrong” sign. Often, the predictors themselves are the problem — they overlap so much that SPSS cannot tell their effects apart. This tutorial shows you how to check for that overlap in SPSS using two simple numbers: tolerance and the variance inflation factor (VIF).
What Is Multicollinearity?
Multicollinearity happens when two or more predictors in a regression model are strongly correlated with each other. In plain terms, they carry much of the same information.
Imagine you predict exam performance from hours studied, number of revision sessions, and total study time. Those three predictors overlap heavily. If you already know hours studied, total study time adds little new information. The model then struggles to assign credit: is it the hours, the sessions, or the total time that matters? That uncertainty shows up as large standard errors, unstable coefficients, and p-values that change dramatically when you add or remove a single predictor.
Multicollinearity does not make your model useless. It makes the individual coefficients hard to trust. That is why checking it is a standard step before you interpret or report a regression.
Perfect multicollinearity — where one predictor is an exact combination of others — is rare and SPSS will usually drop a variable or warn you. The everyday problem is high but imperfect multicollinearity, which is exactly what tolerance and VIF are designed to detect.
Tolerance and VIF Explained
For each predictor, SPSS runs a small side calculation: it predicts that predictor from all the other predictors in the model and looks at how much variance is left unexplained.
Tolerance is the share of a predictor’s variance that is NOT explained by the other predictors. It ranges from 0 to 1. A tolerance close to 1 means the predictor is largely unique. A tolerance close to 0 means it is almost entirely predictable from the other predictors.
VIF is simply the reciprocal of tolerance:
VIF = 1 / Tolerance
So a tolerance of 0.20 gives a VIF of 5, and a tolerance of 0.10 gives a VIF of 10. They tell the same story in opposite directions: low tolerance means high VIF, and both point to overlap.
How should you judge the numbers? There is no single magic cut-off, and different textbooks draw the line in different places, so treat these as practical guides rather than hard rules:
- VIF below 5 (tolerance above 0.20): generally acceptable. Some overlap is normal, especially in social-science data.
- VIF between 5 and 10 (tolerance between 0.10 and 0.20): moderate concern. Look closely, especially if coefficients look unstable or contradict theory.
- VIF above 10 (tolerance below 0.10): serious concern. Action is usually needed before you interpret individual predictors.
Also remember that VIF values are about the predictors, not about the outcome. A high VIF does not mean your predictors are bad measures — it means they overlap. And multicollinearity is mainly a problem when you care about individual predictors. If you only care about overall prediction, it matters much less.
How to Check Multicollinearity in SPSS: Click-by-Click
You check multicollinearity inside the linear regression dialog, so start with your regression set up as usual. Your data should have one continuous outcome variable and two or more predictors, with no missing-value surprises you have not already handled.
- Click Analyze > Regression > Linear.
- Move your outcome (dependent) variable into the Dependent box.
- Move your predictors into the Independent(s) box.
- Click the Statistics button.
- Tick “Collinearity diagnostics” (under Regression Coefficients). While you are here, ticking “Estimates” and “Model fit” is good practice if they are not already selected.
- Click Continue, then click OK.
That is all you need to add. SPSS runs the same regression as before, plus a “Coefficients” table that now includes Tolerance and VIF columns, and a “Collinearity Diagnostics” table with eigenvalues and condition indices.
Paste-Ready SPSS Syntax
If you prefer syntax — and for a thesis, you should, because syntax makes your analysis reproducible — the command below runs a regression and requests collinearity diagnostics. Replace the variable names with your own:
REGRESSION
/MISSING LISTWISE
/STATISTICS COEFF OUTS R ANOVA COLLIN TOL
/CRITERIA=PIN(.05) POUT(.10)
/NOORIGIN
/DEPENDENT outcome_variable
/METHOD=ENTER predictor1 predictor2 predictor3.
The key part is COLLIN TOL in the /STATISTICS line: COLLIN requests the collinearity diagnostics table, and TOL adds the tolerance and VIF columns to the coefficients table. METHOD=ENTER puts all predictors in at once, which is the usual choice when you are checking multicollinearity rather than doing stepwise selection.
Save your syntax file alongside your data file. When a supervisor asks how you checked your assumptions, you can show exactly what you ran.
How to Interpret the Output
The Coefficients table: Tolerance and VIF
Find the Tolerance and VIF columns in the Coefficients table. Read them row by row, one predictor at a time.
A predictor with a VIF of 1.4, for example, is fine — its standard error is only slightly larger than it would be if it were unrelated to the other predictors. A predictor with a VIF of 8 is a warning sign: much of its information is shared with other predictors, so its coefficient and p-value should be treated cautiously. A predictor with a VIF above 10, with tolerance below 0.10, is the clearest signal that something needs to change.
Pay attention to pairs, too. If two predictors both have high VIFs, they are probably overlapping with each other. Check their bivariate correlation as well — a correlation above about 0.80 between two predictors is a strong hint, though VIF can also catch overlap spread across three or more variables that no single correlation reveals.
The Collinearity Diagnostics table: Condition indices
The second table gives eigenvalues, condition indices, and variance proportions. The condition index is a summary of how much overlap exists in each dimension of the predictors.
As a rough guide, a condition index above 30, combined with two or more predictors loading heavily (variance proportions above 0.50) on that same dimension, points to a collinearity problem involving those predictors. Used together with VIF, this table helps you pinpoint which variables are tangled together, rather than just telling you that a problem exists somewhere.
Do not over-interpret small condition indices, and do not panic at a single high index if the VIFs are all low. The two tables should tell a consistent story before you act.
What high VIF does — and does not — mean
High VIF inflates standard errors, which makes confidence intervals wider and p-values larger. That is why a predictor can look non-significant even when it matters: the model cannot separate its effect from its overlapping partner. Coefficients can also flip sign or swing sharply when a correlated predictor is added or removed.
High VIF does not mean your data are wrong, your measures are invalid, or your overall model is worthless. The model’s overall fit (R-squared) is not distorted by multicollinearity in the same way. The damage is specifically to the interpretation of individual predictors — which, for most theses and journal articles, is exactly what you need to report.
What to Do If VIF Is Too High
There is no single fix, because the right fix depends on why the predictors overlap. Work through these options in order:
- Check for coding errors first. A variable entered twice, a total score entered alongside its components, or a dummy variable for every category (the “dummy variable trap”) can create artificial overlap. Fix the coding before anything else.
- Remove or combine overlapping predictors. If two variables measure almost the same thing, keep the one with the stronger theoretical justification or the cleaner measure, and drop the other. Alternatively, combine them into a single index or composite score if that makes conceptual sense.
- Centre your predictors when the overlap comes from interaction or polynomial terms. If you created an interaction term (for example, X times Z) or a squared term, centring X and Z first (subtracting their means) often reduces the collinearity between the main effects and the product term. Centring does not fix overlap between naturally correlated predictors, but it helps with overlap you created yourself.
- Collect more data, if that is a realistic option. Larger samples shrink standard errors and can make moderate multicollinearity less damaging, though they do not remove the overlap itself.
- Consider advanced methods only with justification. Techniques such as ridge regression handle collinearity by shrinking coefficients, but they change what the coefficients mean and are harder to report in a standard thesis. Do not reach for them simply to make a high VIF disappear — a supervisor or reviewer will ask why, and “to fix VIF” is not a sufficient answer on its own.
Whatever you do, report what you did. Dropping a predictor silently is worse than reporting a high VIF honestly with a reasoned decision.
Reporting in APA Style (7th ed.)
APA style does not prescribe a single table format for collinearity diagnostics, but reviewers expect you to state that you checked and what you found. A short paragraph in your results or assumptions section is usually enough:
“Prior to interpreting the regression model, multicollinearity was assessed using tolerance and variance inflation factor (VIF) values. Tolerance values ranged from .42 to .87, and VIF values ranged from 1.15 to 2.38, indicating no problematic multicollinearity.”
Adapt the numbers to your own output, and if you took action, say so: “One predictor showed a VIF of 12.40 and was removed due to substantial overlap with [the other predictor], after which all VIF values were below 3.”
If your field or supervisor expects a table, add tolerance and VIF as columns in your regression coefficients table. Keep decimal places consistent — two decimals are standard — and note the thresholds you used so the reader can judge your decision.
Key Takeaways
- Multicollinearity means your predictors overlap, making individual coefficients unstable and hard to interpret.
- Tolerance is the unique share of a predictor’s variance; VIF is 1 divided by tolerance.
- In SPSS: Analyze > Regression > Linear > Statistics > tick “Collinearity diagnostics”.
- VIF below 5 is generally fine; 5 to 10 warrants a closer look; above 10 usually needs action.
- Check coding errors first, then remove or combine overlapping predictors — and always report what you checked and what you did.
- High VIF harms interpretation of individual predictors, not the overall fit of the model.
FAQs
Do I need to check multicollinearity for every regression?
Yes, whenever your model has two or more predictors and you plan to interpret individual coefficients. It takes one extra tick box in SPSS, and it protects you from reporting a coefficient that is really an artefact of overlap.
What is a good VIF value in SPSS?
There is no universal cut-off. VIF values below 5 are generally considered acceptable, values between 5 and 10 suggest moderate collinearity worth investigating, and values above 10 indicate a serious problem. Report the values you found and the threshold you applied, rather than just saying multicollinearity was “fine.”
Can multicollinearity exist with only two predictors?
Yes. With two predictors, VIF is driven by their correlation with each other. If they correlate very highly, both will show the same elevated VIF. With more predictors, overlap can also be spread across several variables, which is why VIF is more informative than a single correlation.
Does centring variables fix multicollinearity?
Centring helps when the overlap is created by interaction or polynomial terms, because it reduces the correlation between a variable and its own product or square. It does not fix overlap between two naturally correlated predictors, such as income and education — for that, you need to remove, combine, or justify keeping one of them.
Should I delete a variable just because its VIF is high?
Not automatically. First check for coding mistakes, then consider theory: is the variable essential to your research question? If two variables are near-duplicates, dropping one is reasonable. If the variable is theoretically central, discuss the limitation honestly, consider combining measures, or seek advice from your supervisor before dropping it.
Need this done on your data? Get a free quote → /services/