I came across the term "Multicollinearity" in statistics, particularly statistics. However, I never really understand mathematically why highly correlated (almost linearly dependent) columns in the design matrix $X$ lead to higher variance of the regression coefficient, given that the $$Var(\hat{\beta}) = \sigma^2 (X^TX)^{-1}$$ ?

Can someone explain to me mathematically the idea behind please ?

  • $\begingroup$ That formula doesn't even make sense if the columns of $X$ are linearly dependent, because then $X^TX$ is not invertible. Multicollinearity in statistics usually refers not to the case where the columns of $X$ are exactly dependent but to the case where the explanatory variables are highly correlated. $\endgroup$ – symplectomorphic Mar 6 '15 at 16:42
  • $\begingroup$ I meant to say high correlated columns, not perfectly correlated columns. $\endgroup$ – mynameisJEFF Mar 6 '15 at 16:47

Lets start with your formula:


First, note that is the columns of X are linearly dependent, then $X'X$ will not be invertible and have a determinant of 0. And, I won't prove this next part, but if $X'X$ is "close" to being linearly dependent, then the determinant will be "close" to 0.

Now, the inverse of a matrix A is related to the inverse of its determinant:

$A^{-1} = \frac{1}{det(A)}adj(A)$

where $adj(A)$ is the adjugte of the matrix A.

Therefore, as $X'X$ becomes more linearly dependent, its determinant becomes closer to 0, which means the elements of the inverse get larger.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.