1
$\begingroup$

I'm a maths student, this is my first statics class and I'm studying Confidence Intervals. When it's time to estimate the CI for the mean of $\{X_i\}_{i=1..N}$ - gaussian random variables i.i.d. with unknown standard deviation - we use the estimator $\hat{\sigma}$ to compute the quantity

$$T:=\frac{\overline{X}_N - \mu}{\frac{\hat{\sigma}}{\sqrt{N}}}$$

where $\overline{X}_N$ is the sample mean, $\mu$ the mean and

$$\hat{\sigma}:=\sqrt{\frac{1}{N-1}\sum^N (X_i-\overline{X}_N)^2}$$

I would prove that $T \sim t(N-1)$, so:

$$T=\frac{\sigma}{\sigma} \frac{\overline{X}_N - \mu}{\frac{\hat{\sigma}}{\sqrt{N}}}=\frac{\frac{(\overline{X}_N - \mu)\sqrt{N}}{\sigma}}{\frac{\hat{\sigma}}{\sigma}}$$

Now the numerator is a standard gaussian and I'll call it $\tilde{Z}$. Hence,

$$T=\frac{\tilde{Z}\sqrt{N-1}}{\frac{\hat{\sigma}}{\sigma}\sqrt{N-1}}$$

In order to have a Student t-distribution I have to prove $(\frac{\hat{\sigma}}{\sigma}\sqrt{N-1})^2\sim\chi^2(N-1)$.

Solution:

There's a clear proof of that here.

$\endgroup$
8
  • 5
    $\begingroup$ You can't. It isn't. $\endgroup$
    – Glen_b
    Jul 13, 2015 at 12:39
  • $\begingroup$ I think you'll find a lengthy explanation/proof here $\endgroup$
    – user82170
    Jul 13, 2015 at 12:39
  • $\begingroup$ You might want to read the question more closely and consider whether the "lengthy explanation/proof" that you link to really does answer the question asked. Moderator Glen_b's comment on the question is right on target $\endgroup$ Jul 13, 2015 at 13:06
  • $\begingroup$ Still can't, still isn't. Under certain conditions $(n-1)s^2/\sigma^2$ is $\chi^2$ -- and this is covered by a number of answers on site. $\endgroup$
    – Glen_b
    Jul 13, 2015 at 14:10
  • $\begingroup$ I think I have some issues in my handouts. Can you suggest me some good readings about this? Thanks $\endgroup$
    – Phugo
    Jul 13, 2015 at 14:16

1 Answer 1

0
$\begingroup$

Let $V=\sum_{i=1}^n(\frac{X_i-\mu}{\sigma})^2$, you know for each $\frac{X_i-\mu}{\sigma} $it has a $N(0,1)$ distribution because the $X_i$ are $N(\mu, \sigma)$, and you should know that square of a normal distribution has a $\chi^2(1)$ distribution.

Therefore, $V=\chi^2(1)+...+\chi^2(1)=\chi^2(n)$

$V=\sum_{i=1}^n(\frac{X_i-\mu}{\sigma})^2=\sum_{i=1}^n(\frac{(X_i-\bar{X})+(\bar{X}-\mu)}{\sigma})^2=\sum_{i=1}^{n}(\frac{X_i-\bar{X}}{\sigma})^2+2*\sum_{i=1}^n(X_i-\bar{X})(\bar{X}-\mu)/\sigma+(\frac{n(\bar{X}-\mu)}{\sigma})^2$

Let $W:= \sum_{i=1}^{n}(\frac{X_i-\bar{X}}{\sigma})^2$, which is our quantity of interest.

The term $2*\sum_{i=1}^n(X_i-\bar{X})(\bar{X}-\mu)/\sigma$ is equal to 0.

Let $U:= (\frac{n(\bar{X}-\mu)}{\sigma})^2$. $U$ is is another $\chi^2(1)$, using that the square of a $N(0,1)$ is $\chi^2(1)$ and that $\frac{n(\bar{X}-\mu)}{\sigma}$ has a $N(0,1)$ distribution because $X_i$ are $N(\mu, \sigma)$.

Now you have $V=W+U$, with $V$ following a $\chi^2(n)$ distribution and $U$ following a $\chi^2(1)$ distribution.

Therefore, $\sum_{i=1}^{n}(\frac{X_i-\bar{X}}{\sigma})^2$ has a $\chi^2(n-1)$ distribution.

So $W = \frac{\sum_{i=1}^{n}(X_i-\bar{X})^2}{\sigma^2}$ has a $\chi^2(n-1)$

Usually people define sample variance as $s^2=\frac{\sum_{i=1}^{n}(X_i-\bar{X})^2}{n-1}$ which is an unbiased estimator of $\sigma^2$

Therefore, it should be $\frac{(n-1)s^2}{\sigma^2}$ has a $\chi^2(n-1)$ distribution.

Therefore your question has a problem there, you should be very careful.

$\endgroup$
5
  • $\begingroup$ Yeah. I just edited the question $\endgroup$
    – Phugo
    Jul 13, 2015 at 13:57
  • $\begingroup$ Your question still has problems. You may post the original question, i think. $\endgroup$
    – Deep North
    Jul 13, 2015 at 14:00
  • $\begingroup$ With the help of Zhanxiong I understood what was the right question and this is the answer for that. Thank you @Deep North! I would add a reference for future readers: onlinecourses.science.psu.edu/stat414/node/174 This is a more rigorous proof following the scheme up here. $\endgroup$
    – Phugo
    Jul 14, 2015 at 22:38
  • $\begingroup$ The proof writes 𝑉=π‘Š+π‘ˆ, with 𝑉 following a πœ’2(𝑛) distribution and π‘ˆ following a πœ’2(1) distribution. How does one conclude that π‘Š is indeed πœ’2(𝑛-1)? Proof is incomplete. Once can write π‘Š=𝑉-π‘ˆ, but that looks like a πœ’2(𝑛+1), except for dependence between 𝑉 , π‘ˆ. $\endgroup$ Aug 25, 2022 at 0:53
  • $\begingroup$ see stats.stackexchange.com/questions/586626/… which provides proof. $\endgroup$ Aug 25, 2022 at 18:35

Your Answer

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

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