ACADEMIC WRITING GUIDE
How to Check Multicollinearity in Dissertation Regression (With Worked VIF Examples)
Regression can reveal how an outcome is associated with several predictors while accounting for their relationships with one another. But what if two of those predictors convey almost the same information? You might obtain a model that predicts reasonably well while its individual coefficients become uncertain or unstable.
That is the practical problem of multicollinearity. It deserves attention whenever your dissertation interprets the separate contributions of predictors in multiple regression or a related model. Fortunately, you can investigate it systematically rather than deciding that any correlation between variables is a problem.
This guide explains what multicollinearity means, how to calculate and interpret variance inflation factors (VIFs) and tolerance, what software diagnostics can and cannot tell you, and how to report an evidence-based response. The numerical values and study settings are hypothetical teaching examples, not findings collected from real students.
What is multicollinearity?
Multicollinearity occurs when one explanatory variable can be predicted closely from a combination of the other explanatory variables in the model. For example, a study might include both total weekly study hours and hours spent studying independently, alongside hours spent in scheduled study groups. Those measurements could overlap substantially.
At one extreme, perfect multicollinearity occurs when a predictor is an exact linear combination of other predictors. The model cannot uniquely estimate every separate coefficient in the usual way. More commonly, strong but imperfect multicollinearity inflates uncertainty around individual coefficient estimates and can make them sensitive to small changes in the model or sample.
Multicollinearity does not necessarily mean that a whole model is useless, that its predictions are automatically poor, or that a coefficient is biased solely because predictors overlap. The effect depends on the intended interpretation, estimator and research design. It particularly matters when your dissertation claims to identify a predictor's unique association with the outcome, holding other variables constant.
The issue belongs within your full analysis strategy; see Acadelyra's dissertation data analysis plan for guidance on matching models and checks to the research question.
Why does it matter for a dissertation?
Consider a hypothetical model relating students' examination scores to weekly study hours, attendance and a second measure of study engagement. If study hours and engagement largely reflect the same behaviour, the model has difficulty distinguishing their independent statistical associations. You may find that:
- coefficients change noticeably when a nearly redundant predictor is added or removed;
- standard errors and confidence intervals become wider;
- individual p-values become less informative about each predictor's separate association;
- coefficient signs or magnitudes fluctuate across plausible specifications; or
- interpreting an isolated coefficient becomes substantively awkward.
These symptoms are signals to investigate—not automatic proof that the model should be discarded. A non-significant coefficient, on its own, does not establish multicollinearity; limited precision can have many causes.
VIF and tolerance: what they measure
For an ordinary least squares regression containing a predictor named Xj, calculate an auxiliary regression in which Xj becomes the outcome and all the other original predictors are the explanatory variables. Obtain the auxiliary regression's R-squared, denoted Rj².
The standard VIF formula is:
VIFj = 1 / (1 − Rj²)
Its reciprocal is tolerance = 1 / VIFj = 1 − Rj².
VIF quantifies the degree to which the variance of a coefficient estimate is inflated by linear relationships among predictors, compared with an otherwise comparable situation in which that predictor is orthogonal to the others. Under the relevant classical regression assumptions, the associated standard-error inflation factor is the square root of VIF, not VIF itself.
A VIF of 1 indicates that the specified predictor has no linear association with the other predictors in the auxiliary regression. VIF values cannot be below 1 in the standard calculation using an intercept and ordinary R-squared. The larger the VIF, the more strongly the remaining predictors collectively explain that predictor.
The Pennsylvania State University statistics notes explain the derivation and use of VIF in regression diagnostics: Penn State STAT 462 — Detecting Multicollinearity Using Variance Inflation Factors.
Worked example 1: Calculate VIF by hand
Suppose a hypothetical dissertation models examination performance using study hours, class attendance and prior grades. To assess the study-hours predictor, the researcher regresses study hours on attendance and prior grades. This auxiliary regression yields an illustrative Rj² of 0.75.
The calculations are:
- Tolerance: 1 − 0.75 = 0.25.
- VIF: 1 / 0.25 = 4.00.
- Standard-error inflation factor: square root of 4.00 = 2.00.
| Auxiliary R-squared | Tolerance | VIF | Square root of VIF |
|---|---|---|---|
| 0.00 | 1.00 | 1.00 | 1.00 |
| 0.50 | 0.50 | 2.00 | 1.41 |
| 0.75 | 0.25 | 4.00 | 2.00 |
| 0.90 | 0.10 | 10.00 | 3.16 |
Table 1. Illustrative mathematical values; the rows are examples of the formula, not a dataset or software output.
Here, VIF 4.00 indicates that the coefficient's variance is four times the corresponding orthogonal-predictor benchmark; its standard error is approximately twice the benchmark, holding the other relevant model quantities fixed. It does not mean that 75% of the original dissertation outcome is explained by study hours. The 0.75 belongs to the auxiliary regression predicting study hours from the other predictors.
If instead the auxiliary R-squared were 0.90, VIF would rise to 10.00. This makes it harder to distinguish the study-hours coefficient from overlapping explanatory variables. It does not prove that study hours have no substantive relationship with examination performance.
How high is too high? Avoid a magic threshold
Students often see rules of thumb such as VIF above 5 or above 10. These can be useful prompts for examination, but no universal cutoff establishes whether an analysis is valid. Consequences depend on study goals, sample size, model specification, desired precision and the underlying scientific relationships.
The UCLA regression diagnostics tutorial describes VIF values above 10 as a possible reason for closer investigation, while stressing their effects on coefficient stability and standard errors: UCLA OARC — Regression Diagnostics.
| Diagnostic observation | What it reasonably suggests | What it does not prove |
|---|---|---|
| VIF near 1 | Little linear overlap with the other predictors | Every regression assumption holds |
| Moderate VIF | Some coefficient variance inflation | The predictor must be removed |
| High VIF | Substantial predictor overlap worth investigating | The model is necessarily invalid |
| Very small tolerance | Remaining predictors explain most variation in this predictor | The outcome is caused by those predictors |
Table 2. Interpret diagnostics in context, not as universal pass/fail scores.
Report the actual values rather than claiming that all VIFs are “acceptable” merely because they fall under a threshold. A model can have VIF values near 1 and still have serious problems with functional form, independence, measurement or confounding.
Step-by-step: How to check multicollinearity
Step 1. Specify the model and why each predictor belongs
Start with the research question and the conceptual framework. List the outcome variable, predictors, transformations, interactions and planned adjustment variables. Record why each is needed. Avoid choosing predictors solely because their individual p-values are small.
If two questionnaire scores represent almost the same construct, the central question may be conceptual before it is numerical: do they measure distinct things worth estimating separately?
Step 2. Inspect definitions and coding
Check for duplicated variables, different units of the same measurement, category indicators entered alongside a complete set of categories plus an intercept, or derived totals entered with all of their constituent parts. These can cause exact or near-exact linear dependencies.
Interactions and polynomial terms can also create expected relationships among predictors. Centering a variable can sometimes reduce nonessential collinearity between main-effect and interaction terms, but it is not a universal cure and does not change the underlying information in the study. Retain the terms required to respect the scientific model.
Step 3. Review the predictor correlation matrix
A correlation matrix is a helpful first look at pairwise linear relationships. But low pairwise correlations do not rule out multicollinearity, because a predictor can be explained by a combination of several others even when no single pair is exceptionally correlated.
A high pairwise correlation is also not, by itself, an instruction to delete a variable. Consider conceptual meaning and the intended estimand before changing the model.
Step 4. Request VIF or tolerance diagnostics
Most common statistical applications can report VIF for an ordinary linear regression. In SPSS, for example, regression collinearity diagnostics can be requested as part of the linear regression statistics settings. UCLA provides a worked SPSS regression diagnostics guide.
Use the VIF values for your actual final candidate model, with its actual predictors and coding. Do not paste a VIF table from another study or interpret diagnostic values generated from a different specification.
Where categorical predictors have several degrees of freedom, software may use a generalized VIF or another diagnostic. Such numbers may not be directly comparable with the usual one-degree-of-freedom VIF. State what the software reported and do not apply conventional scalar cutoffs blindly.
Step 5. Inspect coefficient precision and stability
Check standard errors, confidence intervals, estimated signs and the substantive plausibility of predictor effects. Where justified, compare carefully selected, theoretically meaningful specifications. If dropping an overlapping variable changes a coefficient greatly, describe what the new specification estimates—it may be answering a different question.
Also assess model-specific issues, such as residual patterns, influential observations, missing data or category sparsity where relevant. Our quantitative data analysis guide explains why model assumptions should be assessed rather than merely named.
Step 6. Decide how to respond and document the reasoning
Options include retaining predictors with an explicit precision warning, removing a genuinely redundant measure for conceptual reasons, using a justified composite, collecting better information, or employing an appropriately validated regularized approach when the research goal is prediction. These options are not interchangeable.
Do not remove a scientifically important confounder merely to make VIF smaller. And do not report a model selected after looking at outcomes as though it were preregistered.
Worked example 2: Read a hypothetical dissertation VIF table
A hypothetical study of undergraduate academic engagement specifies a multiple linear regression with predictors: perceived feedback usefulness, lecturer communication quality, access to course resources, and weekly study hours. The following auxiliary R-squared values and derived diagnostics are intentionally illustrative; they are not claimed to come from a fitted participant dataset.
| Predictor | Auxiliary R-squared | Tolerance | VIF |
|---|---|---|---|
| Feedback usefulness | 0.80 | 0.20 | 5.00 |
| Lecturer communication | 0.75 | 0.25 | 4.00 |
| Resource access | 0.20 | 0.80 | 1.25 |
| Study hours | 0.50 | 0.50 | 2.00 |
Table 3. Hypothetical diagnostic illustration; each VIF is calculated from the auxiliary R-squared shown. The rows are not asserted to form a jointly fitted covariance matrix.
Feedback usefulness and lecturer communication warrant closer attention because other predictors explain much of their variation. However, no single VIF determines whether either measure should remain. The researcher should revisit their definitions, inspect coefficient precision, and decide whether the dissertation seeks a distinct association for each construct.
A defensible draft results paragraph could read:
“Collinearity diagnostics for the proposed predictor set produced illustrative VIF values ranging from 1.25 to 5.00. The highest values occurred for perceived feedback usefulness (VIF = 5.00) and lecturer communication quality (VIF = 4.00), suggesting appreciable overlap in their linear relationships with the other covariates. These diagnostics alone do not establish that either variable is redundant. Final interpretation would require examination of measurement definitions, coefficient uncertainty and the study's conceptual model.”
In a real dissertation, replace all illustrative figures with calculated output from your own analysis, and write in the appropriate past tense only when the study is completed. For broader result-writing standards, see our statistical results reporting guide.
Worked example 3: What to do when predictors overlap
Imagine a study using total study time and independent study time as separate predictors, alongside participation in scheduled study groups. The researcher notices that the total-hours measure includes independent study hours by definition. This is not merely a technical VIF issue: the variables' meanings overlap.
| Investigation | Question | Defensible response |
|---|---|---|
| Review measurement definitions | Does total study time include independent study time? | Document the overlap and revise ambiguous definitions |
| Check diagnostics | Do auxiliary regressions show strong linear overlap? | Report VIF/tolerance and coefficient precision |
| Revisit the research aim | Is the interest in total workload or study-mode differences? | Align model choice with the research question |
| Consider revised specification | Would distinct time categories better represent the constructs? | Specify categories clearly and avoid an exact linear dependency |
| Compare interpretation | What does each coefficient mean after a change? | Explain why the estimand changes; do not compare p-values alone |
Table 4. Conceptual and analytic decisions; no fictitious before-and-after coefficients are presented.
For example, the researcher may decide that total study time alone matches a question about overall workload. Alternatively, a question comparing independent and collaborative study may justify including non-overlapping components, with suitable coding and an interpretation that respects their relationship. The right choice comes from the research question and data-generating process, not a mechanical instruction to keep whichever option yields the smallest VIF.
When should you keep correlated predictors?
Sometimes overlapping predictors are scientifically necessary. Age and experience may be related but represent distinct theoretical ideas. Intervention indicators and their interaction terms can also be essential to interpreting heterogeneous effects. Including relevant controls can be crucial even if their inclusion increases uncertainty for particular coefficients.
If the dissertation is primarily explanatory, be particularly careful about presenting unstable coefficients as independent effects. If the goal is prediction, assess predictive performance on appropriately separate data rather than assuming that VIF alone indicates forecast quality. High predictive accuracy still does not make a causal explanation valid.
UCLA's logistic regression diagnostics guide also discusses multicollinearity in models with binary outcomes. The same predictor-overlap concern extends beyond ordinary linear regression, although diagnostics and their interpretation may depend on the model.
How to report multicollinearity in your methodology and results
A strong dissertation distinguishes the planned diagnostic procedure from the observed diagnostic findings.
Proposed methodology example:
“Before interpreting individual regression coefficients, the study will examine predictor definitions, pairwise correlations and variance inflation factors. The assessment will focus on whether substantial overlap undermines the precision and interpretability of particular coefficient estimates. Any modifications to the model will be justified against the research questions and transparently recorded.”
Completed-study template (replace brackets with real results):
“Collinearity was assessed for the predictors in the final [model name] using [method/software]. VIFs ranged from [minimum] to [maximum], with the largest value for [predictor]. Examination of [relevant coefficient precision and specification evidence] suggested [specific conclusion]. We [retained/revised] the specification because [research-based justification], and interpret [affected coefficient] with [specific caution].”
Do not write “multicollinearity was absent” as an absolute statement when what you mean is that the observed VIFs did not raise material concern under your chosen interpretation. Also avoid claiming that low VIFs establish validity, representativeness, causal identification or proper residual assumptions.
Common mistakes to avoid
Treating VIF = 5 or 10 as a universal verdict. Thresholds are screening aids, not laws. Explain implications for coefficient uncertainty and your research question.
Confusing the auxiliary R-squared with the outcome model's R-squared. VIF uses the R-squared from predicting one predictor with the other predictors.
Confusing variance inflation and standard-error inflation. Variance inflation equals VIF; under the standard formula, the standard-error factor is the square root of VIF.
Removing controls just to lower a VIF. This can change the estimand and introduce confounding. Prioritise scientific justification.
Assuming small pairwise correlations eliminate the issue. Multicollinearity involves relationships among combinations of predictors, not only pairs.
Reporting diagnostics from a different model. VIF depends on which predictors are included and how they are coded.
Claiming a low VIF proves every assumption. Regression requires separate consideration of functional form, error structure, influential observations and relevant design limitations.
Fabricating software output. Worked examples are fine when clearly labelled; findings and methodological claims in a real dissertation must be traceable to its actual analysis.
Frequently asked questions
Is multicollinearity the same as correlation?
No. Correlation usually describes an association between two variables. Multicollinearity concerns whether a predictor can be explained by a linear combination of the other predictors in a model. High pairwise correlation can contribute, but it is not the whole story.
Is a VIF of 4 always acceptable?
No single cutoff answers this. A VIF of 4 indicates fourfold variance inflation relative to an orthogonal benchmark, but whether that creates an important precision problem depends on the study and coefficient being interpreted.
Does multicollinearity make regression coefficients biased?
Not automatically. Under the usual correctly specified classical linear-model assumptions, correlation among predictors alone does not introduce bias into ordinary least squares estimates. It can make separate estimates imprecise or unstable; other model problems may still cause bias.
Can I use VIF for logistic regression?
Collinearity among predictors remains relevant in logistic regression, but software-specific diagnostics and conventions require care. Check the chosen diagnostic against the actual model and predictor coding rather than applying an OLS checklist without explanation.
Should I drop the predictor with the highest VIF?
Not automatically. First examine what the predictor measures, why it belongs, whether it duplicates another measure, and what changes in the scientific meaning of the model if it is removed.
Should multicollinearity appear in the methodology or results chapter?
The methodology can describe the planned diagnostics and decision principles. The results should report what the diagnostics actually showed and any justified specification decisions. Follow your institution's preferred chapter structure.
Final checklist
- Is each predictor defined and scientifically justified?
- Have overlapping totals, components, dummy variables and transformations been checked?
- Were VIFs computed for the appropriate final model and its actual sample?
- Are tolerance and VIF interpreted correctly, without magical thresholds?
- Have coefficient precision, confidence intervals and sensitivity to specification been considered?
- Are any model changes justified conceptually rather than solely by p-values?
- Is planned analysis distinguished from completed findings?
- Are the results and limitations reported without causal overclaiming?
Final thoughts
Multicollinearity is not a label that automatically invalidates your dissertation's regression. It is a way of recognising that related predictors may make individual coefficient estimates difficult to distinguish. Compute the appropriate diagnostics, examine the effect on precision, return to the conceptual framework and report exactly what the evidence supports.
If you need guidance reviewing how your regression decisions are explained and presented, Acadelyra's academic support services can help you improve the clarity, organisation and methodological transparency of your writing while keeping your research and academic responsibilities your own.
Further reading
- Penn State, STAT 462: Detecting Multicollinearity Using Variance Inflation Factors.
- UCLA OARC: Regression with Stata — Regression Diagnostics.
- UCLA OARC: Regression with SPSS — Regression Diagnostics.
Editorial integrity note: All worked values in this guide are hypothetical illustrations. The reader must analyse and accurately report their own data, approved methods and institutional requirements.