I'm reading through Larry Wasserman Book "All of Statistics". Currently I'm in chapter hypothesis testing, the chapter can be found here. I would like to prove theorem 10.6, I quote:

Suppose the true value of $\theta$ is $\theta_* \neq \theta_0$. The power $\beta(\theta_*)$ - the probability of correctly rejecting the null hypothesis - is given (approximately) by $$ 1- \Phi(\frac{\theta_0-\theta_*}{\hat{se}}+z_{\alpha/2})+\Phi(\frac{\theta_0-\theta_*}{\hat{se}}-z_{\alpha/2})$$

We assume that under $H_0: \theta = \theta_0$ the quantity $\frac{\hat{\theta}-\theta_0}{\hat{se}}$ is asymptotically standard normal. We reject $H_0$ if $|W|>z_{\alpha/2}$, where $W:=\frac{\hat{\theta}-\theta_0}{\hat{se}}$

My work so far:

By definition: $$\beta(\theta_*)= P(|W|>z_{\alpha/2})=P(z_{\alpha/2}<W<-z_{\alpha/2})=P(z_{\alpha/2}<\frac{\hat{\theta}-\theta_*}{\hat{se}}<-z_{\alpha/2})=P(z_{\alpha/2}<\frac{\hat{\theta}-\theta_0}{\hat{se}}+\frac{\theta_0-\theta_*}{\hat{se}}<-z_{\alpha/2})=P(z_{\alpha/2}-\frac{\theta_0-\theta_*}{\hat{se}}< Z <-z_{\alpha/2}-\frac{\theta_0-\theta_*}{\hat{se}})=P(z_{\alpha/2}+\frac{\theta_0-\theta_*}{\hat{se}}< Z <-z_{\alpha/2}+\frac{\theta_0-\theta_*}{\hat{se}})$$ where $Z$ is a standard normal and we used that $Z$ is symmetric. The last expression is equal $$P(z_{\alpha/2}+\frac{\theta_0-\theta_*}{\hat{se}}< Z <-z_{\alpha/2}+\frac{\theta_0-\theta_*}{\hat{se}})\approx\Phi(-z_{\alpha/2}+\frac{\theta_0-\theta_*}{\hat{se}})-\Phi(z_{\alpha/2}+\frac{\theta_0-\theta_*}{\hat{se}})$$

however I'm somehow missing the $+1$. Where is my mistake?


I find the way you write compound inequalities to be remarkably and disturbingly strange. If I were to write $$3 < x < -4$$ when what I really mean is $$x > 3 \quad \text{or} \quad x < -4,$$ the former expression would technically mean that $x$ has no solution: there are no real numbers that are simultaneously greater than $3$ and less than $-4$.

In the same vein, then, when you go from $$\Pr[|W| > z_{\alpha/2}] = \Pr[z_{\alpha/2} < W < -z_{\alpha/2}],$$ I can only assume that what you really mean is to write $$\Pr[|W| > z_{\alpha/2}] = 1 - \Pr[|W| \le z_{\alpha/2}] = 1 - \Pr[-z_{\alpha/2} \le W \le z_{\alpha/2}].$$ I think that you will find that this will fix your problems.

  • $\begingroup$ thanks for your help. Agree that notation was terrible and was the reason why I didnt derive the correct result. One question. If you go through it you end with $P(-z_{\alpha/2}+\frac{\theta_0-\theta_*}{\hat{\theta}}<Z<z_{\alpha/2}+\frac{\theta_0-\theta_*}{\hat{\theta}})$ Then I multiply all equations with $-1$ to get the right one and use symmetry of $Z$. But why is it desired to do that? $\endgroup$ – math Oct 14 '15 at 19:01

I am fortunate to have come across this thread while revising! Just to complete the proof, here's my working.

I am using Wasserman's notation of $z_{\alpha}$ where $z_{\alpha} = \Phi^{-1} (1-\alpha)$, which means that $z_{\alpha}$ is positive for $\alpha < 0.5$.

Continuing from math and heropup's answer:

$$\beta (\theta_*) = 1 - \mathbb{P}\left(-z_{\alpha/2} < \frac{\hat{\theta} - \theta_*}{\hat{se}} < z_{\alpha/2} \right) $$ $$= 1 - \mathbb{P}\left(-z_{\alpha/2} < \frac{\hat{\theta} - \theta_0}{\hat{se}} + \frac{\theta_0 - \theta_*}{\hat{se}} < z_{\alpha/2} \right)$$

Using the assumption of asymptotic normality, $\frac{\hat{\theta} - \theta_0}{\hat{se}} \sim N(0,1)$

Denote the standard normal as $Z$

$$\beta (\theta_*) = 1 - \mathbb{P}\left(-z_{\alpha/2} < Z + \frac{\theta_0 - \theta_*}{\hat{se}} < z_{\alpha/2} \right) $$ $$ = 1 - \mathbb{P}\left(-z_{\alpha/2} - \frac{\theta_0 - \theta_*}{\hat{se}} < Z < z_{\alpha/2} - \frac{\theta_0 - \theta_*}{\hat{se}} \right)$$

$$ = 1 - \Phi\left(z_{\alpha/2} - \frac{\theta_0 - \theta_*}{\hat{se}}\right) + \Phi\left( -z_{\alpha/2} - \frac{\theta_0 - \theta_*}{\hat{se}} \right)$$

By some additional manipulation using the symmetry property of the distribution (I guess for aesthetics?):

$$= 1 - \Phi\left( \frac{\theta_0 - \theta_*}{\hat{se}} + z_{\alpha/2}\right) + \Phi\left(\frac{\theta_0 - \theta_*}{\hat{se}} - z_{\alpha/2}\right)$$


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.