5
$\begingroup$

There are $X_1, X_2$ where $X_i \sim N(\mu_i,1), i=1,2$. They are independent. The question is

Find the likelihood ratio test with $H_0:(\mu_1,\mu_2)=(0,0), H_1:(\mu_1,\mu_2) \neq (0,0)$. The significance level is $\alpha (0< \alpha <1)$ and parameter space $\Omega$ is $$\Omega = \left\{ (\mu_1,\mu_2) : \mu_1 \geq 0, \mu_2 \geq 0\right\}$$

My solution is $X_1^2 \geq \chi_p^2(1)$ or $X_2^2 \geq \chi_q^2(1)$ or $X_1^2+X_2^2 \geq \chi_r^2(2)$ where $p+q+r=\alpha$. Is it right?

Detail of my solution :

Let $\mu = (\mu_1, \mu_2)^T$. Then $\hat{\mu}^{\Omega_0}=(0,0)$ and $\hat{\mu}^{\Omega} = \left(\max\{x_1, 0\}, \max\{x_2,0\}\right)$ because parameter space is not $\mathbb{R}^2$.

Then I calculated $\Lambda = 2[l(\hat{\mu}^{\Omega}) - l(\hat{\mu}^{\Omega_0})]$ to find the rejection region from $\Lambda \geq \lambda (\lambda > 0)$.

After some algebra, I got $\Lambda = x_1^2I_{(x_1>0, x_2<0)} + x_2^2I_{(x_1<0, x_2>0)} + (x_1^2+x_2^2)I_{(x_1>0, x_2>0)}$.

Under the null hypothesis, $X_i^2 \sim \chi^2(1)(i=1,2)$ so $X_1^2+X_2^2 \sim \chi^2(2)$.

Finally I got the above rejection region.

$\endgroup$
2
  • $\begingroup$ Thanks. I added the tag and detail. $\endgroup$
    – flossy
    Oct 25 '20 at 4:29
  • $\begingroup$ Can you say something about how do you get the MLE of $(\mu_1, \mu_2)$? $\endgroup$
    – Bob
    Jan 15 at 21:23
1
$\begingroup$

Your solution seems to be correct. The strange shape of your parameter space (it's not a open subset of $\mathbb R^2$) creates this ambiguity in the final result: each combination of $(p,q,r)$ gives a different LRT. Some are more powered for $\mu_1=0$, some are more powered for $\mu_2=0$ and some are more powered for $\mu_1\neq0,\mu_2\neq0$, but all of them are valid LRTs with significance level $\alpha$.

$\endgroup$
2
  • $\begingroup$ Thank you for your comment! I understand as follows: My rejection region is valid LRT but not MP(most powerful) for all hypotheses $\endgroup$
    – flossy
    Oct 26 '20 at 16:07
  • $\begingroup$ @flossy that's correct. also, there is no uniformly most powerful test in this setting, since the LRT with $p=1$ is MP if $\mu_2=0$ and the LRT with $q=1$ is MP if $\mu_1=0$. $\endgroup$
    – PedroSebe
    Oct 27 '20 at 4:00

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.