4
$\begingroup$

I came across a paper that assumes the following has a t-distribution:

Let $W = \frac{\mathbf{a}'\mathbf{X}}{\sqrt{\mathbf{X}'\mathbf{X}}}$ and $\mathbf{a}' \in \mathbb{R}^n$ with $\mathbf{a}'\mathbf{a} = 1$. If $\mathbf{X} \sim N_n(\mathbf{0}, \mathbf{I}_n)$ then

$$\frac{\sqrt{(n-1)}W}{\sqrt{\left(1 - W^2\right)}},$$

follows a t-distribution with $n-1$ degrees of freedom.

I don't quite see how this is true though. I'm assuming that to start you'd need to show that $\mathbf{a}'\mathbf{X}$ and $\sqrt{\mathbf{X}'\mathbf{X}}$ are independent, but I'm not quite sure how to proceed in the multivariate case.

$\endgroup$
  • 1
    $\begingroup$ Hi. Welcome to CV. Showing that $\mathbf{a}'\mathbf{X}$ and $\sqrt{\mathbf{X}'\mathbf{X}}$ are independent would prove what $W$ follows t-distribution. And if so were the case, the statistic that you have shown would not follow t-distribution. So probably you need to first write your statistic in terms of $\mathbf{X}$. $\endgroup$ – Dayne Oct 14 at 9:59
  • 1
    $\begingroup$ $a'X$ and $\sqrt{X'X}$ are not independent. $\hat \sigma^2$ is independent on $\sqrt{X'X}$, but total sum of squares isn't. $\endgroup$ – carlo Oct 14 at 11:09
  • $\begingroup$ @carlo You're right. On second thought, I think the approach is to rather condition the statistic on $\mathbf{a}'$, i.e. the conditional distribution of the statistic given $\mathbf{a}$ is the $t_{n-1}$ distribution. You can do this since $\mathbf{a}'$ and $\mathbf{X}$ are independent. $\endgroup$ – user334916 Oct 14 at 11:31
4
$\begingroup$

Continuing from my comment above:

Let $$T=\frac{(\sqrt{n-1})W}{\sqrt{1 - W^2}}$$

Now, $$1-W^2 = 1-\frac{X'aa'X}{X'X}$$

$$\implies (1-W^2)X'X=X'X-X'AX=X'(I-A)X\,,\qquad A\equiv aa'$$

Therefore,

\begin{align} T&=\frac{(\sqrt{n-1})a'X}{\sqrt{X'X}}\frac{\sqrt{X'X}}{\sqrt{X'(I-A)X}} \\&=\frac{(\sqrt{n-1})a'X}{\sqrt{X'(I-A)X}} \end{align}

See that $A$ is a symmetric idempotent matrix with $\operatorname{tr}(A)=\operatorname{rank}(A)=1$. Similarly, $(I-A)$ is also symmetric idempotent so $\operatorname{rank}(I-A)=n-1$.

Now let $$Z=a'X \quad\text{ and }\quad V=X'(I-A)X$$

We have, $$X'X=X'AX+X'(I-A)X = Z^2+V$$

Since $\operatorname{rank}(A)+\operatorname{rank}(I-A)=n$, from Cochran's theorem:

$X'AX$ and $X'(I-A)X$ are independent.

Therefore $Z$ and $V$ are independent and $Z^2\sim \chi^2_{(1)}$ and $V\sim \chi^2_{(n-1)}$.

Finally, $$T=\frac{Z}{\sqrt{V/(n-1)}}\,,$$ which by definition now follows t-distribution with $(n-1)$ degrees of freedom.

$\endgroup$

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.