This is very basic, but I have been stuck here for a while.

Consider an AR(1) model $Y_t = c+\phi Y_{t-1} +\epsilon_t$, where $c$ is a constant. If $\epsilon_t \sim i.i.d. N(0, \sigma^2),$ then $Y_1, \dots, Y_T$ are also Gaussian, where $Y_1$ is the first observation in the sample.

I don't quite understand how we have each single realization of $Y_t$ Gaussian. It seems that the conditional distribution $Y_t|Y_{t-1}$ is Gaussian, but why $Y_t$ is Gaussian unconditionally?

  • $\begingroup$ For last question: Because $Y_{t+1}$ and $\epsilon_t} are normal distributed. See here for proof. mathworld.wolfram.com/NormalSumDistribution.html $\endgroup$ – user158565 Oct 21 '18 at 20:14
  • $\begingroup$ @a_statistician Wait, why $Y_{t+1}$ is Gaussian? $\endgroup$ – Vivian Miller Oct 21 '18 at 20:38
  • $\begingroup$ typo. Should be $Y_{t-1}$ and $\epsilon_t$ are normal, so their linear combination (one of them is $Y_t$) is normal according to linked site. $\endgroup$ – user158565 Oct 21 '18 at 20:54
  • $\begingroup$ @a_statistician How do you know $Y_{t-1}$ is normal? $\endgroup$ – Vivian Miller Oct 21 '18 at 21:05
  • 1
    $\begingroup$ Think from beginning when $t=1$. Then $t=2$ ... $t=t-1$ $\endgroup$ – user158565 Oct 21 '18 at 21:12

You are right to be confused here. Strictly speaking, the asserted conclusion in the highlighted statement is a non sequiter. (Can you please add the source of the statement?) The element $Y_1$ is defined recursively in terms of $Y_0$ in the specified recursive equation. Since there is no specification of the distribution of $Y_0$, the distributions of the observable values is not determined.

What they should have specified is that $Y_0 | \boldsymbol{\epsilon} \sim \text{N}.$ Unfortunately, people are notoriously sloppy in setting up time-series models, and it is commonly the case for the model to not be properly specified. With a bit of practice you get used to "reading between the lines" to figure out what was intended.

  • 2
    $\begingroup$ Actually, specifying that $Y_0$ is Gaussian without also saying that $Y_0$ is independent of all the $\varepsilon_i$'s (or at least jointly Gaussian with all the $\varepsilon_i$'s) is also sloppy because without such a clause, $Y_1 = c + \phi Y_0 + \varepsilon_1$ is not necessarily Gaussian: the sum of two marginally Gaussian random variables is not Gaussian unless joint Gaussianity is also assumed. Perhaps it is simpler to set $Y_0=0$ and avoid the extra complications. $\endgroup$ – Dilip Sarwate Oct 21 '18 at 22:40
  • $\begingroup$ Good point - edited. $\endgroup$ – Ben Oct 21 '18 at 23:14

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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