Multicollinearity and the variance inflation factor

Multicollinearity is when two or more predictor variables in a regression carry overlapping information, so the model cannot separate their individual effects. It does not bias the predictions, but it inflates standard errors, which makes coefficients unstable, widens confidence intervals, and can leave a predictor non-significant even though the variable genuinely matters. The standard diagnostic is the variance inflation factor, and a VIF above 5 is the usual warning line.

Why it damages a dissertation model specifically

The harm is precise, and knowing exactly what it does and does not touch saves a great deal of unnecessary worry.

What survives multicollinearity: the model's overall fit, the R-squared, the F test, and the predicted values. A model riddled with correlated predictors can still forecast well. What breaks: every claim about an individual predictor. Standard errors inflate, so t-statistics shrink and p-values rise; coefficients swing wildly, sometimes flipping sign, when you add or remove a variable or change the sample slightly.

That distinction decides whether you need to act. If your research questions are about prediction, high correlation among predictors may be tolerable. If your hypotheses name individual predictors, which is the usual case in doctoral work, then unstable coefficients undermine the exact claims your thesis rests on. A hypothesis stating that organisational support predicts retention over and above job satisfaction cannot be tested when the two are measured almost interchangeably.

Where it comes from

• Redundant measurement. Two scales that tap the same construct, or a total score entered alongside its own subscales.
• Structural multicollinearity. Created by the model rather than the data: an interaction term or a squared term is correlated with the variables it was built from. Centring the predictors before forming the term usually removes it.
• Variables that are definitionally linked. Age, years of experience, and years since qualification move together because they measure overlapping things.
• Too many predictors for the sample. With a small sample and many variables, correlations arise by chance alone.
• The dummy variable trap. Entering every category of a nominal variable instead of leaving one out as the reference produces perfect collinearity, and most software will drop a variable on its own.

How to detect it, in the order worth using

1. The predictor correlation matrix

Run it before the model. Correlations above about .80 between two predictors are an early warning, and this step often identifies a redundant pair before you have fitted anything. It is not sufficient on its own, because multicollinearity can involve three or more variables with no single pairwise correlation looking alarming.

2. The variance inflation factor

The variance inflation factor is the proper diagnostic because it accounts for a predictor's relationship with all the others at once, not one at a time. Each predictor is regressed on the remaining predictors, and the resulting R-squared becomes the VIF:

VIF = 1 / (1 - R²)

A VIF of 1 means a predictor is uncorrelated with the others. A VIF of 5 means its R-squared against them is .80, so the variance of its coefficient is five times what it would be if the predictors were independent.

Tolerance is simply 1 divided by the VIF, so the two carry identical information and reporting either is enough. Be aware that the thresholds are conventions, not tests. Some fields apply a stricter cut of 2.5, and a VIF of 8 on a control variable you never intend to interpret matters far less than a VIF of 6 on your key predictor.

3. Condition index and eigenvalues

For a stubborn case, the collinearity diagnostics table adds a condition index. Values above 30 combined with two or more variance proportions above .50 on the same row identify which specific variables are entangled, which VIF alone cannot tell you.

Checking multicollinearity in SPSS

SPSS does not show these diagnostics by default; they sit behind a checkbox that is easy to miss.

1. Open Analyze, then Regression, then Linear.
2. Put your outcome in Dependent and your predictors in Independent(s).
3. Click Statistics and tick Collinearity diagnostics. Tick Estimates and Model fit as well if they are not already selected.
4. Click Continue, then OK. Read the Tolerance and VIF columns on the far right of the Coefficients table, and the Collinearity Diagnostics table beneath it for the condition index.

The same check in R and Stata

VIF in R

In R the diagnostic takes one line once the model exists. Fit the regression with lm(), then call vif() from the car package:

model <- lm(retention ~ support + satisfaction + tenure, data = df)
library(car)
vif(model)

The output lists one value per predictor, read against exactly the same thresholds as the SPSS table above. For a categorical predictor with more than two levels, use vif(model, type = "predictor") so you read a generalised VIF for the variable as a whole rather than one value per dummy. The wider procedure is in building a regression with several predictors and fitting the same model in R.

VIF in Stata

In Stata, run the regression first and ask for the diagnostic immediately afterwards:

regress retention support satisfaction tenure
estat vif

Stata prints a VIF for each predictor plus the mean VIF, and the 1/VIF column is the tolerance. Judge each predictor individually against the thresholds above; the mean VIF is a summary, not a decision criterion, and a clean mean can hide one badly inflated predictor.

What to do when the VIF is too high

Ranked by how often they are the right answer in dissertation work.

• Do nothing, and justify it. Legitimate more often than students expect. If the inflated VIF belongs to a control variable, or the model exists to predict rather than to explain, report the diagnostic and explain why it does not threaten your conclusions.
• Drop one of a redundant pair. Keep the one with stronger theoretical justification, not the one with the better p-value. Say which you kept and why.
• Combine them. If two predictors measure one underlying construct, average or sum them into a single index, after confirming they belong together through a check on scale reliability.
• Centre before forming interactions. Where the problem is structural, subtracting the mean from each predictor before multiplying them removes most of the inflation and changes nothing about the model's fit.
• Collect more data. A larger sample shrinks standard errors and can make an otherwise unstable coefficient usable. Rarely available mid-dissertation, but worth stating as a limitation.

One tempting route to avoid: dropping variables one at a time until the VIFs look acceptable. Letting a diagnostic select your model turns a confirmatory analysis into an exploratory one, and an examiner will ask what theory justified the final specification.

Reporting it

Multicollinearity belongs in the assumptions paragraph of your results chapter, alongside the other checks a regression model has to pass. State that you examined it, give the diagnostic and the range of values, and say what you concluded.

A clean sentence reads: multicollinearity was assessed using variance inflation factors, all of which fell below 2.4 (tolerance above .42), indicating that the predictors were sufficiently independent for the individual coefficients to be interpreted. If a value did exceed the threshold, report it, name the variables involved, and state the action taken and the reasoning behind it. Reviewers accept a documented, justified decision far more readily than a diagnostic that goes unmentioned.