1
$\begingroup$

We have $X_i \sim ^{iid} N(\mu,\sigma^2)$with known mean $\mu$, and unknown $\sigma^2$. Let's suppose we're given $s^2$ (sample variance) and $n$ (sample size).

We know $\frac{\bar X-\mu}{S/\sqrt{n-1}} \sim t(n-1)$.

  • Is there a way to compute $P(\bar X>c)$?

Now, I've seen several times the following reasoning:

$$P(\bar X>c)=P\left(\frac{\bar X-\mu}{S/\sqrt{n-1}}>\frac{c-\mu}{s/\sqrt{n-1}}\right)$$

However, I find this odd, instead I think we should have

$$P(\bar X>c)=P\left(\frac{\bar X-\mu}{S/\sqrt{n-1}}>\frac{c-\mu}{S/\sqrt{n-1}}\right)$$

because we're supposed to divide by the same expression on both sides of the inequality.

(Note: By $s$ I mean the r.v $S$ evaluated at a given sample)

The motivation for this question is the following: Imagine I wanted to test the null $\mu \leq \mu_0$. If we knew the population variance, then we would expect to reject the null if $\bar X > c'>\mu_0$. So,

$$P(\bar X>c')=P\left(\frac{\bar X-\mu_0}{\sigma/\sqrt{n-1}}>\frac{c'-\mu_0}{\sigma/\sqrt{n-1}}\right)$$

and we would chose $\frac{c'-\mu_0}{\sigma/\sqrt{n-1}}:= q_{1-\alpha}$, the $1-\alpha$ quantile for $N(0,1)$.

However, this type of reasoning cannot be done for the case at the begining of this question. We can prove that a similar rejection interval does indeed come out from using the Likelihood Ratio Test(LRT). I was wondering if there was a way to reach the same rejection interval without having to use the LRT...

Any help would be appreciated.

$\endgroup$
3
  • 1
    $\begingroup$ What is the difference between $S$ and $s$? Are you sure it isn't a typo? $\endgroup$
    – dlnB
    Mar 14, 2019 at 21:54
  • $\begingroup$ @dlnB The difference is that $s$ is $S$ computed for a given sample, i.e. it's a constant. $S$ is not. $\endgroup$ Mar 14, 2019 at 21:57
  • $\begingroup$ $P(\bar{X}_n>c) =P(\frac{\bar{X}_n-\mu}{s/\sqrt{n-1}}>\frac{c-\mu}{s/\sqrt{n-1}})$ is correct logic. Multiplying both sides by a positive constant or subtracting a constant from both sides preserves the inequality. $\endgroup$
    – dlnB
    Mar 14, 2019 at 22:19

1 Answer 1

0
$\begingroup$

$X_1,...,X_n$ are iid $N(\mu, \sigma^2)$. The first step is to derive the distribution of $\bar{X}_n= \sum_{i=1}^n \frac{1}{n}X_i$. First notice that $\frac{1}{n}X_i \sim N(\frac{\mu}{n}, \frac{\sigma^2}{n^2})$ for $i=1,...,n.$ Then $\bar{X}_n$ is simply the sum of these iid variables $\frac{1}{n}X_i$, so $\bar{X}_n \sim nN(\frac{\mu}{n} ,\frac{\sigma^2}{n^2})=N(\mu ,\frac{\sigma^2}{n}).$ It follows that $$P(\bar{X}_n > c) = \int_{c}^{\infty} \frac{1}{\sqrt{2 \pi \sigma^2/n}} \exp\{-\frac{(z-\mu)^2}{2\sigma^2/n}\}dz. $$

This is one expression for the exact probability you describe, but requires $\sigma^2$ for calculation.

Following the $t$ approach you describe, $$P(\bar{X}_n>c) =P(\frac{\bar{X}_n-\mu}{s/\sqrt{n-1}}>\frac{c-\mu}{s/\sqrt{n-1}})=P(t(n-1) > \frac{c-\mu}{s/\sqrt{n-1}}),$$

which can simply be found by taking $1-F(\frac{c-\mu}{s/\sqrt{n-1}})$, where $F(\cdot)$ is the cdf for the $t(n-1)$ distribution.

$\endgroup$
7
  • $\begingroup$ thanks for the answer, however, I don't think it's correct... we need to use $S$ and not $s$ for the pivot to have the $t(n-1)$ distribution. $\endgroup$ Mar 14, 2019 at 22:30
  • 1
    $\begingroup$ I still don't understand what the two different notations mean. You write $s^2$ is sample variance, but use $S$ in your formula for $t(n-1)$. The "s" or "S" used in the $t$ formula is the estimated standard deviation from the sample. $\endgroup$
    – dlnB
    Mar 14, 2019 at 22:33
  • $\begingroup$ It's the same difference as the one we have between $X_i$ and $x_i$. Hum... I'm not going to be completely correct, but think of $X_i$ as a random variable and $x_i$ as an observation of the variable $X_i$. $x_i$ is a constant. $\endgroup$ Mar 14, 2019 at 22:37
  • $\begingroup$ I see your confusion. $\bar{X}$ and $S$ are indeed random variables, so let $\bar{X}_n$ and $s$ denote realized values of those random variables, calculated from your sample $X_1,...,X_n$. A sample $X_1,...,X_n$ gives us realized values $\bar{X}_n$ and $s$ and therefore a realization of $\frac{\bar{X}_n-\mu}{s/\sqrt{n-1}}$, which is a realized value from a $t(n-1)$ distribution. Therefore $P(\frac{\bar{X}_n-\mu}{s/\sqrt{n-1}}>\frac{c-\mu}{s/\sqrt{n-1}})=P(t(n-1) > \frac{c-\mu}{s/\sqrt{n-1}}).$ $\endgroup$
    – dlnB
    Mar 14, 2019 at 22:48
  • $\begingroup$ The problem is that we have only $\frac{\bar X-\mu}{S/\sqrt{n-1}}\sim t(n-1)$. We do NOT know the distribution of $\frac{\bar X-\mu}{s/\sqrt{n-1}}$ (because we do not know the population variance) $\endgroup$ Mar 14, 2019 at 23:00

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.

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