Intro
Econometrics students across California—including UCLA, USC, UC Berkeley, UC Irvine, UC Davis, UC Santa Cruz, UC Riverside, and the CSU system—encounter multicollinearity around mid‑semester. This is where regression intuition breaks: coefficients become unstable, standard errors explode, and t‑tests lose power. For additional support, visit our Econometrics Tutoring or explore related topics in the Statistics Post Hub.
Answer First
Multicollinearity inflates standard errors because highly correlated predictors make it difficult for OLS to isolate each variable’s unique contribution. When predictors move together, the design matrix becomes nearly singular, causing the variance of the estimated coefficients to increase dramatically.
Problem Setup
Consider the multiple regression model:
\[ y = X\beta + u \]
The OLS estimator is:
\[ \hat{\beta} = (X’X)^{-1}X’y \]
The variance of \(\hat{\beta}\) is:
\[ Var(\hat{\beta}) = \sigma^2 (X’X)^{-1} \]
Multicollinearity affects the matrix \(X’X\), making it nearly singular.
Step-by-Step Solution
1. Understand what multicollinearity means
Two or more predictors are highly correlated. This makes it difficult for OLS to determine which variable explains variation in the dependent variable.
2. Examine the role of \(X’X\)
The matrix \(X’X\) summarizes how predictors relate to each other. When predictors are highly correlated, \(X’X\) becomes close to singular.
3. Connect singularity to variance inflation
When \(X’X\) is nearly singular, its inverse contains very large values. Because the variance formula includes \((X’X)^{-1}\), standard errors become large.
4. Interpret the consequences
- Coefficient estimates become unstable.
- Small changes in data produce large changes in coefficients.
- t‑tests lose power because standard errors are inflated.
- Variables appear “insignificant” even when they matter.
5. Understand what multicollinearity does NOT do
- It does not bias coefficients.
- It does not affect predictions.
- It only affects inference.
Intuition
OLS tries to separate the effect of each predictor. When two predictors move together, OLS cannot tell which one is responsible for changes in the dependent variable. This uncertainty shows up as large standard errors.
Common Exam Mistakes
- Thinking multicollinearity biases coefficients (it does not).
- Confusing multicollinearity with heteroskedasticity.
- Dropping variables without considering theory.
- Assuming VIF thresholds are universal (they are not).
Why This Matters
Multicollinearity is common in real datasets—income and education, advertising channels, macroeconomic indicators, and demographic variables often move together. Understanding variance inflation is essential for interpreting regression output correctly.
Final Summary
Multicollinearity inflates standard errors because correlated predictors make the design matrix nearly singular. This causes the inverse matrix to contain large values, which increases the variance of coefficient estimates and weakens t‑tests. Coefficients remain unbiased, but inference becomes unreliable.