Given regression equation Y=Xβ + ε which satisfies all classical assumptions including those of homoscadasticity

Estimated regression equation using OLS is the following -

Y=Xb + e

Where, b is the OLS estimator for β and e is the estimated residual term for equation where E(e|x) =0

Note: b=(X'X)$^{-1}$X'Y

We can find, Cov(b,e|x) = E(b.e|x) - E(b).E(e|x)

=E(b.e|x) = E((X'X)$^{-1}$X'Y.e|x)=E((X'X)$^{-1}$X'(Xβ + ε).e|x)

=E((β.e + (X'X)$^{-1}$X'ε.e)|x)


Solving for E(ε.e) gives σ²(n-k) [I wrote e= Y-Xb, substituting for b and some manipulation gave me e= ε + X(X'X)$^{-1}$X'ε]

This gives Cov(b,e|x)=(X'X)$^{-1}$X'σ²(n-k)

I think in this expression - (X'X)$^{-1}$X' - I shouldn't have X' in the right (as the formula of var(b) suggests). What am I missing?


  • 1
    $\begingroup$ In your last equality E(β.e|x + (X'X)−1X'ε.e|x) is kind of vaudou for most humans. $\endgroup$
    – Bertrand
    Sep 10, 2019 at 13:48
  • $\begingroup$ @Bertrand Please have a look at the edited question. Does this make sense now? $\endgroup$ Sep 10, 2019 at 14:24
  • $\begingroup$ i) You should be careful with the dimensions of your vectors: $b$ is $K \times 1$ and $e$ is $N \times 1$, so the covariance is $K \times N$. It is better to transpose $e$ in $E[be^T]$. ii) After taking carefully into account the dimensions, it turns out that $E[\varepsilon e^T]$ is a $N \times N$ matrix, and is different from what you mention. $\endgroup$
    – Bertrand
    Sep 10, 2019 at 19:22
  • $\begingroup$ @Bertrand Thank you! Doing so, I am getting 0 covariance. Is it right? I mean why does it be so? $\endgroup$ Sep 10, 2019 at 23:48
  • $\begingroup$ The 0 answer seems to be right. You may also want to prove it using projections matrices $P_X$ and $M_X$. $\endgroup$
    – Bertrand
    Sep 11, 2019 at 6:41


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Browse other questions tagged or ask your own question.