The Statement of the Problem:

Let $X_1,..., X_n$ be a random sample from $N(\mu, 1)$. Consider testing $$ H_0: \mu=0 \quad \text{vs.} \quad H_a: \mu=1. $$

Let the rejection region be $C = \{ \overline X > c\}$
(a) Find $c$ so that the test has size $\alpha = 0.01$.
(b) Find the power under $H_a$, i.e. find $\gamma_C(1).$
(c) Show that $\gamma_c(1) \to 1$ as $n \to \infty$. How do you interpret this?

Where I Am:

I apologize if this is really basic, but I can't figure out how to do this. I know how to test for alternative hypotheses that are state that the mean is not a value, or larger than or smaller than some value; and I know how to test for the difference between means of different samples, but this doesn't really make sense to me. If someone could shed a little light on this, I'd appreciate it.

  • $\begingroup$ is it a homework? $\endgroup$
    – Simone
    Commented Jul 30, 2015 at 2:14
  • $\begingroup$ Yes, it is. (Here's some more characters). $\endgroup$ Commented Jul 30, 2015 at 2:15
  • 1
    $\begingroup$ Testing an alternative of this form is essentially the same as a composite alternative. The question is already specifying the form of the rejection region, so you want to consider the distribution of $\bar{X}$ when $\mu = 0$ and when $\mu = 1$, and how often you will be rejecting $H_0$ for different values of $c$ in either situation. $\endgroup$
    – dsaxton
    Commented Jul 30, 2015 at 2:16
  • 1
    $\begingroup$ @Simone, it is best to ask the OP to add the SS tag & read it's wiki, rather than add the tag for them. That way, there is a better chance they will be familiar w/ our policy. If they don't add the tag, you can vote to close. $\endgroup$ Commented Jul 30, 2015 at 3:05
  • 1
    $\begingroup$ @gung Will do! Sorry. First-time poster. $\endgroup$ Commented Jul 30, 2015 at 3:06

3 Answers 3


$\bar{X}$, the sample mean, is distributed as $N(0,n^{-1})$ when $H_0$ is true and as $N(1,n^{-1})$ when $H_1$ is true. (Here, the parameters are $\mu$ and $\sigma^2$

Thus, we want $c$ to be that number for which $P\{\bar{X} > c \mid H_0\}$ equals $0.01$. But, $$P\{\bar{X} > c \mid H_0\} =Q\left(c\sqrt{n}\right)$$ where $Q(\cdot)$ is the complementary standard normal distribution function, and a quick look at tables (or resort to an R-gument) shows that $Q(2.33) \approx 0.01$. Hence,

$$c \approx \frac{2.33}{\sqrt{n}}\tag{1}$$

It follows that

$$P\{\bar{X} > c \mid H_1\} = Q\left((c-1)\sqrt{n}\right) \approx Q\left(2.33 - \sqrt{n}\right)\tag{2}$$

Since the argument of $Q$ in $(2)$ diverges to $-\infty$ as $n \to \infty$ (and so the value of $Q(\cdot)$ approaches $1$), we have part (c) as well. The interpretation is that while the threshold $c$ converges to $0$ as $n\to\infty$ (cf. $(1)$), the variance of $\bar{X}$ is converging to $0$, and thus, when $H_1$ is true, the probability that $\bar{X} \sim N(1,\sqrt{n^{-1}})$ exceeds the threshold $c$ is close to $1$. Even though that threshold is quite close to the mean $1$ in absolute terms, it is zillions of standard deviations away from the mean $1$.



(a) Under $H_0$ we have $\bar{X} \sim$ normal$(0, 1 / n)$ and so $\bar{X} \sqrt{n} \sim$ normal$(0, 1)$, and so we want to find a $c$ such that $P(\bar{X} \sqrt{n} > c \sqrt{n}) = \alpha$. It's perhaps a bit confusing that the author framed the question in this way since here $c$ will depend on $n$ in order to get proper size.

(b) You actually can't answer this without knowing $n$, because the power depends on $n$. I'm beginning to wonder if the author made a mistake.

(c) Suppose we've chosen some $z$ to use as the lower bound of a rejection region of the form $\{ \bar{X} \sqrt{n} > z \}$ (i.e., $z = c \sqrt{n}$). When $H_1$ is true we can look at the event of rejection in this way $\{ \bar{X} \sqrt{n} > z \} = \{ (\bar{X} - 1)\sqrt{n} + \sqrt{n} > z \} = \{ (\bar{X} - 1)\sqrt{n} > z - \sqrt{n} \}$. What happens to the probability of this event as $n \to \infty$?


For (b), you will reject $H_0$ if $\frac{\bar{X}-0}{\sigma/\sqrt{n}}\geq Z_{\alpha}$ (where $\sigma=\sqrt{1/n}$ in fact, it should be sample variance)

The power is the probability to reject $H_0$ if the $H_1$ is true

$\gamma_{(c)}(1)=P_{\mu}( \bar{X} \geq Z_{\alpha}\sigma/\sqrt{n} )=P_{\mu}(\frac{\bar{X}-1}{\sigma/ \sqrt{n}} \geq \frac{Z_{\alpha}\sigma/\sqrt{n}-1}{\sigma/\sqrt{n}})=P_{\mu}(\frac{\bar{X}-1}{\sigma/ \sqrt{n}}\geq Z_{\alpha}-\frac{1}{\sigma/\sqrt{n}})=1-\Phi(Z_{\alpha}-n)$

($\sigma=\sqrt{1/n}, \Phi $ is CDF of standard normal distribution)

So your power function is $1-\Phi(Z_{\alpha}-n)$. It depends on both $\alpha $ and $n$.

When $\alpha=0.01$ The power is $1-\Phi(1.64485-n)$

You also can see when $n\rightarrow\infty, \Phi(1.64485-n)\rightarrow0$ so the power will be 1


Your Answer

By clicking “Post 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.