1
$\begingroup$

Suppose $X$ is a single observation from a $\text{Poisson}(\lambda)$ distribution. Determine the uniformly most powerful size $\alpha = .05$ test for the hypothesis test $H_0: \lambda = 5$ and $H_1: \lambda = 10$.

My thoughts: The Neyman-Pearson Lemma tells us the UMP level $\alpha$ test is the likelihood ratio test. The likelihood ratio function is $\Lambda = \frac{5^x e^{-5}}{x!}/ \frac{10^x e^{-10}}{x!} = 2^{-x}e^5$. We reject the null hypothesis if $\Lambda < k$ which is equivalent to rejecting if $x > k^*$.

Our hypothesis test will be of the form $$\phi(x) = \begin{cases} 1 \hfill &\text{if $x > k^*$} \\ \gamma \hfill &\text{if $x = k^*$} \\ 0 \hfill &\text{otherwise} \end{cases} $$ To make a size $\alpha$ test we want to solve for $\gamma$ in the following: $$\alpha = .05 = \text{E}_{\lambda=5}(\phi(x)) = P_{\lambda=5}(X>k^*) + \gamma P_{\lambda=5}(X=k^*).$$

Can someone offer some confirmation/guidance on this? I am not confident in my method.

$\endgroup$
  • $\begingroup$ where did you get that formula for $\Lambda$? $\endgroup$ – Taylor May 7 '18 at 2:53
  • 1
    $\begingroup$ I'm not sure why I used the incorrect values there. I edited my expression, so I believe that part is now correct. $\endgroup$ – antsatsui May 7 '18 at 3:17
1
$\begingroup$

In case you are still interested, your work upto this point was correct as far as I can tell.

Suppose $\lambda\in\Theta$ where $\Theta=\mathbb R^{+}$ is the parameter space.

By Neyman-Pearson lemma, a most powerful test of size $\alpha$ for testing $H_0:\lambda=\lambda_0$ against $H_1:\lambda=\lambda_1(\ne\lambda_0)$ is usually defined as

\begin{align} \phi(x)=\begin{cases}1&,\text{ if }\Lambda(x)>c\\\gamma&,\text{ if }\Lambda(x)=c\\0&,\text{ if }\Lambda(x)<c\end{cases} \end{align}

, where

$$\Lambda(x)=\frac{f_{\lambda_1}(x)}{f_{\lambda_0}(x)}$$

and $c\,(>0)$ and $\gamma\in(0,1)$ are so chosen that $$E_{\lambda=\lambda_0}\,\phi(X)=\alpha$$

Therefore in this case,

$$\Lambda(x)=e^{-5}2^x\quad,\,\text{ an increasing function of }x$$

That is, for some positive $k$, $$\Lambda(x)>c\implies x>k$$

So the test $\phi$ is of the form

\begin{align} \phi(x)=\begin{cases}1&,\text{ if }x>k\\\gamma&,\text{ if }x=k\\0&,\text{ if }x<k\end{cases} \end{align}

, where $k$ and $\gamma\in(0,1)$ are so chosen that

\begin{align} E_{\lambda=5}\,\phi(X)=0.05 \end{align}

That is, $$P(Y>k)+\gamma P(Y=k)=0.05\quad,\, Y\sim \text{Poisson}(5)$$

I used WolframAlpha to get

$$P(Y>8)\approx 0.068094\qquad\text{ and }\qquad P(Y>9)\approx 0.031838$$

So looks like $k=9$ is the appropriate value subject to the size restriction and corresponding to that, we get $$\gamma\approx 0.501075$$

Finally, the test function is

\begin{align} \phi(x)=\begin{cases}1&,\text{ if }x>9\\0.501075&,\text{ if }x=9\\0&,\text{ if }x<9\end{cases} \end{align}

Since the above test is independent of $\lambda_1=10$, $\phi$ is most powerful against all alternatives $\lambda_1\in \Theta_1$ where $\Theta_1\subset \Theta-\{\lambda_0\}$. Hence $\phi$ is uniformly most powerful of size $\alpha$ for testing $H_0$ versus $H_1$.

$\endgroup$
  • $\begingroup$ Forgot to define $f_{\lambda}$: it is the population pmf of course. $\endgroup$ – StubbornAtom Oct 13 '18 at 19:28

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.