The question is pretty much contained in the title: in a linear model $Y=X\beta +\epsilon$, with $\epsilon\sim N_n(0,\sigma^2 I_n)$, does the Cook's distance, defined as $$ D_i:=\frac{\frac{1}{p}||X(\hat{\beta}_{(-i)}-\hat{\beta})||^2}{\tilde{\sigma}^2}, $$ follow an $F_{p,n-p}$-distribution?

Wikipedia seems to suggest it does, but the sentence in the Wikipedia article is cryptic and says "Since Cook's distance is in the metric of an F distribution...".

Can someone point me to a formal proof of the fact that $D_i\sim F_{p,n-p}$?

I am aware that $D_i=\frac{1}{p}\frac{p_i}{1-p_i} \hat{\eta}_i^2$, but what this seems to indicate is that $D_i\sim \frac{1}{p}\frac{p_i}{1-p_i} F_{1,n-p}$, not $F_{p,n-p}$.

  • $\begingroup$ I would think that this is a change in variables problem, but I don't have the time to attempt it right now. $\endgroup$
    – Galen
    Nov 23, 2023 at 20:29
  • $\begingroup$ Thank you for answering. Unfortunately I don't see how it could be, because it seems like, even supposing the leverage of $X_i$ is nice, say $p_i \approx 1/2$, we get $D_i \sim 1/p F_{1,n-p} $ which should be approximately $1/p F_{p,n-p}$ and not $F_{p,n-p}$ (using $\chi^2_p/p\approx \chi^2_1$). $\endgroup$
    – No-one
    Nov 23, 2023 at 20:35


Your Answer

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

Browse other questions tagged or ask your own question.