Multicollinearity: detection and correction

Multicollinearity: detection and correction

Published September 28, 2026

Multicollinearity: detection and correction

On this page

Learning Outcome Statement

Explain multicollinearity and how it affects regression analysis.

On the exam, the vignette gives you the regression specification, sample frequency, coefficient and standard-error evidence, and one or more auxiliary-regression R-squared values. You may need to calculate the variance inflation factor, diagnose multicollinearity, interpret both statistical and economic conclusions, and justify a correction in an item-set or MCQ format.

Multicollinearity means that independent variables carry overlapping information. Imagine a regression explaining a fund's monthly excess return with several independent variables. Pick one and call it $X_i$ when you need to examine it. If the remaining independent variables move with $X_i$, the regression has trouble separating their individual contributions. The overall fit can still look good while one coefficient becomes unstable or statistically weak.

DiagnosticEvidenceSignal
Pairwise correlationA pair of independent variablesPairwise screen
Auxiliary regression$X_i$ on remaining variablesJoint overlap
Auxiliary $R_i^2$R-squared from that regressionExplained share
VIF for $X_i$Auxiliary $R_i^2$Variance inflation
Statistical significanceCoefficient, standard error, tailEvidence against null
Economic significanceCoefficient and exposurePractical importance

Pairwise correlation is a screening device, not the full check for joint overlap. VIF uses the auxiliary regression of one independent variable on the other independent variables, so it can catch overlap that pairwise checks miss. Statistical significance tells you how much evidence the sample gives against a null, while economic significance asks whether the effect matters to the investment decision. The evidence comes from monthly observations, with one period equal to one month, and the inputs are not annualized.

To turn the overlap into a diagnostic, use the variance inflation factor for $X_i$, written as $\mathrm{VIF}_i$. It is dimensionless. The auxiliary-regression R-squared is written as $R_i^2$, and the auxiliary regression includes an intercept. Use this calculation.

$$ \mathrm{VIF}_i = \frac{1}{1 - R_i^2} $$

As $R_i^2$ rises, the other variables explain more of $X_i$. The unexplained share, $1-R_i^2$, shrinks, so its reciprocal rises. A high VIF means that less independent variation remains for estimating the coefficient. Standard errors usually rise and $t$-statistics usually fall, but multicollinearity does not necessarily bias ordinary least squares coefficient estimates.

Picture $X_i$ sharing information with the other independent variables. The original regression uses all variables to explain the fund's excess return. The auxiliary regression asks how much of $X_i$ the other variables already explain.

When the case supplies the auxiliary regression, follow this sequence.

Before you calculate, keep the two regressions straight. The original model explains the fund's return, while the auxiliary model explains one independent variable using the others. Its R-squared measures how much of that variable is predictable from its peers. The unexplained share is the part that remains available to identify its separate contribution. Use that share in the denominator before interpreting the result.

Loading diagram…

Run the diagnostic before changing the model. The exam's reference level is a guide, not an automatic deletion rule. Remove a variable or combine variables only when they carry redundant economic information and your theory supports the choice. Deleting a meaningful variable can damage the specification, so a weak $t$-statistic alone is not enough. Read the sequence as a calculation aid: it identifies the selected variable, specifies the auxiliary regression, and connects the resulting VIF to a decision grounded in theory. A large diagnostic flags imprecision; it does not decide the model for you.

Worked example

You are handed the case values below for a factor regression of a fund's monthly excess return. One period is one month, and the inputs are not annualized. The case gives the sample size, number of independent variables, estimated size coefficient, standard error, and an auxiliary regression of the size variable on the other factors. The exam asks whether multicollinearity is present and whether the size coefficient is statistically significant.

InputValue
Sample120 monthly sample observations
Independent variables3 explanatory variables
Selected auxiliary $R_i^2$0.8 dimensionless R-squared
Size coefficient $\hat{\beta}_{\mathrm{size}}$0.12 percentage points of monthly excess return per one-unit factor exposure
Standard error0.1 percentage points of monthly excess return per one-unit factor exposure
Hypothesis$H_0$ has $\beta_{\mathrm{size}}=0$ versus $H_1$ has $\beta_{\mathrm{size}}\neq 0$
Tail and significanceTwo-tailed, $\alpha=0.05$
VIF reference5 heuristic reference

The line gives you a visual check on the denominator. Every plotted value is repeated in the comparison table below, so you can see how VIF responds as the selected variable becomes more predictable from the other variables. As auxiliary R-squared rises, the unexplained share falls and VIF rises. Track that inverse relationship before using the case's selected value: the curve is a quick check on direction, while the table makes each calculation easy to compare.

Loading diagram…

Read the curve from left to right: greater explanatory power in the auxiliary regression corresponds to a larger VIF. The value at 0.80 reaches the stated heuristic reference, and 0.90 produces still more inflation. The comparison below gives the exact pairs, so you can check the selected case value against nearby levels without treating the reference as a hard cutoff.

Auxiliary $R_i^2$VIF
0.201.25
0.502.00
0.805.00
0.9010.00
  1. Use the auxiliary regression's R-squared, because VIF measures overlap among independent variables. For the size variable, $R_i^2=0.8$.
  2. Subtract the auxiliary R-squared from $1$ to find the unexplained share. The calculation gives $1 - 0.8 = 0.2000$.
  3. Invert that share to obtain the variance inflation factor. The result is $\mathrm{VIF}_{\mathrm{size}}=\frac{1}{0.2000}=5.00$.
  4. Define $t$ as the estimated coefficient divided by its standard error. Here $t=\frac{0.12}{0.1}=1.20$, so the two-tailed test at $\alpha=0.05$ does not support rejecting $H_0$.

For the fund investor, the positive estimate is 0.12 percentage points of monthly excess return per one-unit factor exposure. It can still matter economically even though the coefficient is not statistically significant. Because the VIF reaches the heuristic reference level of $5$, treat the estimate as imprecise and remove the variable only if its economic content is redundant.

Changed condition

Now the same case reports auxiliary $R_i^2=0.2$ for the size variable, while the coefficient and standard error stay unchanged. Predict the direction before reading the table.

MeasureOriginalChanged
Auxiliary $R_i^2$0.80.2
Unexplained share0.20000.8000
VIF5.001.25

Lower auxiliary $R_i^2$ leaves more of the selected variable unexplained. The denominator grows, so VIF falls from 5.00 to 1.25 and the multicollinearity concern weakens. The coefficient and its $t$-statistic do not move because their inputs did not change. Before reading on, calculate the changed VIF yourself.

Common error

You may substitute the regression’s dependent-variable R-squared for the auxiliary-regression R-squared in VIF because both are called R-squared. That mistake feels right because the original regression’s fit seems relevant, and the VIF formula appears to need an R-squared. The wording that gives it away is the explicit regression of one independent variable on the remaining independent variables.

Check yourself

Practice

An analyst estimates a fund's monthly excess return with 3 independent variables using 120 monthly sample observations. For the size variable, the auxiliary regression on the other independent variables has an R-squared of 0.8. The estimated coefficient is 0.12 percentage points of monthly excess return per one-unit factor exposure, with standard error 0.1. The test is two-tailed at 0.05, and the VIF reference is 5.

Question 1

Calculate the VIF for the size variable.

Question 2

Which evidence is the direct VIF diagnostic?

Question 3

Which conclusion best describes the size coefficient?

Question 4

Which correction is most defensible when factors measure the same economic exposure?

The auxiliary R-squared changes from 0.8 to 0.2. The coefficient remains 0.12 and its standard error remains 0.1. Which outcome follows?

Revision snapshot

$$ \mathrm{VIF}_i = \frac{1}{1-R_i^2} $$

VIF is $1$ divided by $1$ minus the R-squared from the auxiliary regression. Use the regression of one independent variable on the remaining independent variables. Pairwise correlation is a screening device. Do not substitute the dependent-variable R-squared. Multicollinearity usually increases coefficient standard errors and reduces $t$-statistics without necessarily creating bias in OLS coefficient estimates. Separate statistical significance from economic significance before deciding whether to remove a variable. On the exam, calculate VIF, diagnose the precision problem, and justify a correction with theory and economic reasoning.

Unlock the full lesson and continue reading

Signup to continue reading this lesson and unlimited access questions, flashcards, AI notes, and more

Scan to download app

Scan to download
UNLOCK FREE ACCESS

Practice Questions: Multicollinearity: detection and correction

  1. 1. Northbridge Investment Research evaluates an actively managed US equity fund using a monthly factor regression. The research team estimates fund excess return as a function of market excess return, a size-factor return, and a value-factor return. The 60 observations are consecutive monthly observations, and all return variables are stated in percentage points. The model is estimated using ordinary least squares with an intercept. The team has confirmed that the dependent variable and each factor-return series are stationary over the sample period. Exhibit 1 presents coefficient estimates, conventional ordinary least squares standard errors, p-values based on those conventional standard errors, and variance inflation factors. A senior analyst notes that the market coefficient should be interpreted conditionally, with the other included factor returns held constant. She also cautions that association in the regression does not, without an identification strategy, establish a causal effect. The research team uses a 5% significance level for all tests. Exhibit 2 reports model fit and diagnostic statistics. The overall F-test has the null hypothesis that all three slope coefficients equal zero. The Breusch–Pagan test has the null hypothesis that the regression error variance is constant. The Durbin–Watson statistic is also reported because consecutive monthly residuals may be serially correlated; a value materially below 2 is consistent with positive serial correlation. Two of the analysts disagree about the diagnostic implications. One argues that high variance inflation factors necessarily bias all coefficient estimates. The other responds that multicollinearity principally increases the uncertainty of the affected coefficient estimates while leaving ordinary least squares coefficients unbiased if the remaining assumptions, including exogeneity, hold. They also consider whether conventional standard errors remain suitable if the Breusch–Pagan test rejects its null hypothesis. No evidence indicates that the regression's functional form is nonlinear, and the team does not intend to remove an explanatory variable solely because its individual p-value exceeds 5%. The final report must distinguish model fit, individual coefficient significance, joint significance, and causal interpretation.

    • 2. A quantitative analyst at Sterling Capital estimates a multiple regression to explain the monthly excess returns of a small-cap equity portfolio. The explanatory variables are the domestic equity market's monthly excess return and a small-minus-big size factor return. Both variables are measured in percentage points, and the data comprise 60 consecutive months. The analyst intends to use the regression both to interpret historical factor exposures and to support risk forecasts. Before doing so, she evaluates coefficient significance, overall model fit, and the residual assumptions underlying ordinary least squares inference. Exhibit 1 presents the regression estimates and selected diagnostics. The null hypothesis for each coefficient t-test is that the population coefficient equals zero. The joint F-test has the null hypothesis that both slope coefficients equal zero. The analyst recognizes that statistical significance does not by itself establish economic importance. She also recognizes that R-squared measures the proportion of sample variation in the dependent variable explained by the fitted model; it does not demonstrate that changes in the regressors cause portfolio returns. Accordingly, Statement 1 in Exhibit 2 must be assessed separately from the numerical fit of the model. For residual diagnostics, the Breusch–Pagan test uses homoskedasticity as its null hypothesis; a high p-value fails to reject that null. The Durbin–Watson statistic is interpreted relative to 2: values near 2 indicate the absence of material first-order serial correlation, values materially below 2 are associated with positive serial correlation, and values above 2 are associated with negative serial correlation. Variance inflation factors are used to assess whether correlations among explanatory variables materially inflate estimated coefficient variances; the market excess return and size factor are known to be correlated in many samples, raising the possibility of multicollinearity. The analyst notes that multicollinearity can inflate coefficient standard errors and destabilize estimates even when the residual assumptions of homoskedasticity and no serial correlation are satisfied. She therefore plans to consider dropping or combining highly correlated regressors, or collecting additional data, before relying on the reported t-statistics for final investment decisions.

      • 3. Meridian Securities analyst Elena Rossi investigates whether a company's monthly stock excess return is related to the market and size factors. She estimates an ordinary least squares multiple regression using 60 consecutive monthly observations. The dependent variable and both factor returns are expressed in percentage points. Her initial results appear in Exhibit 1. Rossi uses a two-sided 5% significance level for individual coefficient tests. She observes that the model has a relatively high R-squared and a statistically significant F-test. She cautions, however, that overall model significance does not establish that every classical linear regression assumption is satisfied or that the estimated relationships are causal. Rossi next performs residual and regressor diagnostics, summarized in Exhibit 2. For the Breusch–Pagan test, the null hypothesis is that the residual variance is constant. She also examines the Durbin–Watson statistic because the observations are ordered through time. Finally, she calculates the correlation between the two included factors and their variance inflation factors. She understands that high multicollinearity can make individual coefficient estimates imprecise even when the regression has substantial explanatory power. A colleague proposes dropping the size factor solely because its coefficient is negative. Rossi responds that coefficient sign alone is not a valid variable-selection criterion. The sign should instead be interpreted in light of the specified model, statistical evidence, economic rationale, and possible misspecification. Additional research identifies momentum as a potentially omitted factor. Prior evidence applicable to the sample indicates that momentum has a positive partial effect on the company's excess return. Momentum is also negatively correlated with the included size factor. Rossi assumes that momentum is not perfectly collinear with the included regressors and that adding it would otherwise leave the model suitable for ordinary least squares estimation. She evaluates the likely direction of omitted-variable bias before re-estimating the model.

        • 4. Northbridge Research develops a model to explain monthly changes in investment-grade corporate credit spreads. The dependent variable is measured in basis points. The explanatory variables are the contemporaneous monthly equity market return, measured in percentage points, and the monthly change in the central bank policy rate, also measured in percentage points. The analyst estimates the model using ordinary least squares and reports the results in Exhibit 1. The research director asks whether the equity market coefficient is statistically different from zero using a two-sided test at the 1% significance level. She emphasizes that statistical significance does not by itself demonstrate economic importance or causality. She also asks the analyst to interpret R-squared strictly as a sample measure of goodness of fit. As part of model validation, the analyst plots residuals against fitted spread changes. The residuals remain centered around zero, but their dispersion increases substantially as fitted values rise. The observations are displayed in chronological order only incidentally; the values along the horizontal axis represent fitted values, not time. The analyst has not yet applied heteroskedasticity-consistent standard errors. A credit strategist then identifies market liquidity as a potentially omitted variable. Deteriorating liquidity tends to widen credit spreads and is also associated with negative equity market returns. The team therefore concludes that liquidity both affects the dependent variable and is correlated with an included regressor. No interaction term or liquidity proxy is present in the reported model. The team is not claiming that the regression establishes a structural causal relationship. It is evaluating whether the reported inferential statistics are reliable and whether the coefficient estimates may be biased by model misspecification. The stated critical value applies directly to the requested test, and candidates need not calculate additional degrees of freedom.

          • 5. Meridian Securities analyst Priya Shah is developing a model to explain a company's monthly stock return using the broad-market return and the monthly change in the term spread. All returns and changes in the term spread are entered as decimals; therefore, 1.00% is entered as 0.0100. A decrease of 0.25 percentage points in the term spread is entered as −0.0025. Exhibit 1 reports the ordinary least squares estimates from 60 monthly observations. Shah initially evaluates the fitted model by examining the coefficient estimates, their standard errors, and the overall model statistics. She tests each slope coefficient using a two-sided null hypothesis that the population coefficient equals zero, and she also constructs confidence intervals around the coefficient estimates. All tests use a 5% significance level. She also evaluates whether the model is statistically significant as a whole, but she recognizes that the overall F-test does not replace tests of the individual coefficients. For a scenario analysis, Shah assumes a monthly market return of 1.00% and a 0.25-percentage-point decrease in the term spread. She intends to calculate the fitted stock return directly from the reported intercept and slope coefficients. She does not add a separate residual forecast because the conditional expected value of the error term is assumed to be zero. Shah next examines the residual diagnostics in Exhibit 2. The Breusch–Pagan test has a null hypothesis of homoskedastic residual variance. The Durbin–Watson statistic is used to assess first-order serial correlation, with a value near 2 indicating no serial correlation. Variance inflation factors are used to identify whether linear relationships among explanatory variables materially inflate coefficient variances. The research director states that the regressors may be treated as exogenous, but asks Shah to identify any condition that could make conventional ordinary least squares standard errors unreliable. If such a condition is present, she should select a correction that preserves the estimated coefficients while improving inference. The director also cautions her not to infer causality solely from R² or statistical significance and not to diagnose nonstationarity using a test designed for residual variance.

            • 6. An analyst at Atlas Advisory evaluates the monthly excess returns of the Atlas Tactical Fund. She estimates a multiple regression in which the dependent variable is the fund's monthly return minus the one-month risk-free rate. The explanatory variables are the global equity market's monthly excess return and the monthly change in the 10-year minus 2-year government bond yield spread. Both the dependent and explanatory return series are expressed in percentage points, while the term-spread change is also measured in percentage points. The ordinary least squares estimates appear in Exhibit 1. The analyst states that each slope coefficient should be interpreted while holding the other included explanatory variable constant. She uses a two-tailed 5% significance level for individual coefficient tests. With 60 observations and three estimated parameters, including the intercept, the applicable degrees of freedom are 57. The stated critical value is supplied so that no distribution table is required. Before using the model for inference, the analyst performs the diagnostics in Exhibit 2. The Breusch–Pagan test has a null hypothesis of homoskedastic errors. The Durbin–Watson statistic is close to 2, and the variance inflation factors are well below commonly used levels of concern. She intends to retain the economic specification unless a diagnostic result identifies a problem that requires a different estimation or inference procedure. A colleague then suggests that the regression may omit a traded liquidity factor. Additional research indicates that this factor has a positive partial effect on the fund's excess return. It is also positively correlated with the global equity market excess return. The analyst assumes that the liquidity factor is otherwise measured without error and focuses specifically on the likely direction of omitted-variable bias in the estimated market coefficient. She does not interpret adjusted R-squared as proof of causality or as a test of any individual coefficient.

              • 7. Meridian Asset Consulting is evaluating whether the Meridian Tactical Fund's monthly excess returns are related to the equity market and changes in the term spread. Analyst Priya Shah estimates an ordinary least squares regression using 60 consecutive monthly observations. The dependent variable is the fund return minus the one-month risk-free rate. The explanatory variables are the market excess return and the monthly change in the ten-year government yield minus the three-month government yield. Exhibit 1 presents the estimated coefficients and model statistics. Shah tests each slope coefficient using a two-sided test at the 5% significance level. The null hypothesis for each test is that the relevant population slope coefficient equals zero. The critical value shown in Exhibit 1 reflects the applicable residual degrees of freedom and must be used instead of a normal-distribution critical value. Shah also considers the overall F-test, for which the null hypothesis is that both slope coefficients jointly equal zero. A colleague makes three observations. First, he argues that the market coefficient cannot be statistically significant because its estimate exceeds 1.0. Second, he concludes that the term-spread coefficient is significant because its sign is negative and the overall regression has a low F-test p-value. Third, he states that an R² of 0.380 proves the explanatory variables cause 38% of the fund's monthly excess return. Shah must evaluate these statements using standard regression interpretation. Before using the model for manager assessment, Shah examines the residuals. Exhibit 2 reports a Breusch–Pagan test, the Durbin–Watson statistic, and variance inflation factors. At the 5% level, a Breusch–Pagan p-value below 0.05 is evidence that the error variance is not constant. For this monthly regression, a Durbin–Watson statistic materially below 2.0 is consistent with positive residual serial correlation. VIF values below 5.0 are not considered evidence of serious multicollinearity under the firm's diagnostic policy. Shah does not intend to alter the estimated coefficients solely because of the diagnostic results. Her immediate objective is to obtain reliable hypothesis tests and confidence intervals. Any proposed treatment must therefore address the effect of the detected error structure on estimated standard errors without deleting economically justified regressors.

                • 8. An equity analyst at Northbridge Securities evaluates the monthly excess returns of Delta Industries. She estimates a multiple regression using 60 consecutive monthly observations. The dependent variable is Delta's return minus the one-month risk-free rate, expressed in percentage points. The explanatory variables are the domestic market excess return, also in percentage points, and the monthly change in an investment-grade credit spread, measured in percentage points. Exhibit 1 presents ordinary least squares estimates using conventional standard errors. For the credit-spread coefficient, the analyst tests the null hypothesis that the coefficient equals zero against the two-sided alternative that it does not equal zero. At the 5% significance level, the applicable critical t-value is 2.00 in absolute value. She separately uses the reported F-test to assess the joint null hypothesis that both slope coefficients equal zero. The analyst then reviews the residual evidence in Exhibit 2. The Breusch–Pagan test is conducted at a 5% significance level. A residual plot shows that dispersion increases with the magnitude of the market excess return. The Durbin–Watson statistic is materially below 2, creating an additional concern about positive serial correlation. The analyst can obtain heteroskedasticity-consistent standard errors without changing the estimated coefficients. She understands that this adjustment addresses inference under heteroskedasticity but does not itself correct serial correlation or omitted-variable bias. To quantify the serial correlation and decide on a joint correction, she also approximates the first-order autocorrelation coefficient implied by the Durbin–Watson statistic and considers whether an autocorrelation-and-heteroskedasticity-consistent standard error, such as Newey-West, is required instead of a heteroskedasticity-only correction. A colleague identifies an omitted monthly investor-sentiment measure. Economic research indicates that higher sentiment increases Delta's return after controlling for the included regressors. In this sample, sentiment is positively correlated with the market excess return. The analyst assumes this correlation is not caused by measurement error and asks how omitting sentiment affects the estimated market coefficient. Finally, the analyst cautions that neither statistical significance nor an R-squared of 0.48 establishes causality. She will use the regression for conditional return analysis only after correcting inference and assessing whether the model remains stable in a later out-of-sample period.

                  Flashcards: Multicollinearity: detection and correction

                  1. 1. What is incorrect variable form mispecification?

                    Show answer

                    the relationship between dependent & independent is non linerar

                  2. 2. What is Breush-Pagan test used for?

                    Show answer

                    conditional heteroskedasticity (autocorrelation)\ to see if the size of the independent variables influence the size of the residuals

                  3. 3. Four types of model specification errors:

                    Show answer

                    1\. ommision of an important independent variabel\ 2\. innappropriate variable form\ 3\. innappropriate variable scaling\ 4\. data improperly pooled

                  4. 4. What is the BP test statistic and df?

                    Show answer

                    n\R2\ df = k (parameters)

                  5. 5. What model qualities lead to Type I errors?

                    Show answer

                    - conditional heteroskedasticity - positive serial correlation

                  6. 6. What is look-ahead bias?

                    Show answer

                    using info that was unavailable at the time of the investment decision

                  7. 7. What may data improperly pooled cause?

                    Show answer

                    serial correlation

                  8. 8. What may ommision of an important independent variable cause?

                    Show answer

                    serial correlation\ model coefficients will be biased/inconsistent

                  Everything you need for CFA prep

                  Get full Oncourse access with lessons, practice questions, flashcards and AI study tools.

                  GET STARTED FOR FREE
                  75% OFFLimited time offer
                  GET 75% OFF