This is a question from Statistical Theory I have encountered. I have almost solved it, but have some trouble interpreting the solution. Something seems weird, and I am not sure whether I am entirely correct.

X is a random variable with distribution $\frac3{\theta^3} x^2 {1}_{(0,\theta)}$. We want to test the hypotheses: $H_0: \theta = 1$ vs. $H_1: \theta = 1.1$ with significance level $\alpha$ = $P(\text{reject }H_0| H_0 \text{ is true})$. We gather $n$ = 100 observations.

Okay, we perform the likelihood ratio test: $$ \Lambda = \frac{L(x_i,\theta_1)}{L(x_i\theta_0)} = \frac{(\frac{3}{1.1^3})^n(\Pi x_i)^2 1_{(0,1.1)}}{3^n(\Pi x_i)^2 1_{(0,1)}} = \begin{cases} c_n, & \text{if $\max x_i \leq 1$} \\ +\infty, & \text{if $\max x_i \geq 1$} \\ \end{cases} $$ Where $c_n = (\frac 1{1.1^{3n}}) = c_{100} = c$

I get trouble attempting to match the hypothesis level: the test should go like this: reject $H_0$ is $\Lambda$ > $C = C(\alpha)$.

If $C < c$ we always reject $H_0$. Therefore significance level is 1. if $C > c$ we reject the null hypothesis only if $\max x_i > 1$. But under $H_0$ this never happens, therefore $\alpha = 0$

The power of the test is $\pi = 1 - P(\text{do not reject } H_0| H_1 \text{ is true}) = \text{either } 1 \text{ or some other constant, which I have calculated to be ~~$1-0.75^n$, depending on $C >< c$}$

So, first of all, can somebody please verify whether I have made any significant mistakes? (I may have gotten some number wrong, but am I conceptually correct?)

How do I calculate $C = C(\alpha)?$ What is the precise significance and power of this test for, say $\alpha = 0.01$? The fact the the test is so "$\alpha$-independent" makes me suspicious.


OK, you found the ratio correctly. Now you need to construct the critical region in the form $\Omega = \{x:\Lambda(x) > C_{\alpha}\}$, where $C_{\alpha}$ corresponds to your significance level $\alpha$. Let's vary $C$ and look what we have. If $C=0$, then our critical region consists of all possible $x$, i.e. $\Omega_1 = [0, 1.1]$. In this case the probability of 1 type error equals 1. And we have the same result for all $C \in [0,c_n)$. If $C \ge c_n$, then our critical region consists only of $x$ with $\max x_i \ge 1$, i.e. $\Omega_2 = \{ x: \max x_i \ge 1 \} $. In this case the probability of 1 type error equals 0. So, we see, that neither $\Omega_1$ nor $\Omega_2$ are not critical regions what we are looking for. But we've got through all possible values $C$. In this case a randomization process can be applied: we consider the critical function $$\varphi \left( x \right) = \left\{ \begin{gathered} 1, \text{ for } x \in \Omega_2 \\ \alpha, \text{ for }x \in \Omega_1 \setminus \Omega_2 \\ \end{gathered} \right.$$ If $x\in \Omega_2$, then $H_0$ is rejected. If $x \in \Omega_1 \setminus \Omega_2$, then $H_0$ is rejected with probability $\frac{\alpha-\alpha_0}{p_0}$, where $\alpha_0 = \Pr(\Omega_2 | H_0)=0$ and $p_0 = \Pr(\Omega_2 \setminus \Omega_1 | H_0) - \Pr(\Omega_2|H_0)=1$. Note also that you accept $H_0$ with probability 1 on $\Omega_1 \setminus \Omega_1 = \varnothing$ as $\Omega_1$ contains all possible $x$ in our case. This criterion is proved to be optimal in the sense of minimizing the probability of 2 type error.

This situation is standard when you consider hypotheses with discrete (and in some cases with continuous) distributions.

  • $\begingroup$ So, if I understood correctly, in this case I reject $H_0$ when it's obvious it's not true (i.e. I have observed something impossible under $H_0$), and otherwise I flip an $\alpha$ coin? $\endgroup$
    – Aahz
    Aug 18 '14 at 13:29
  • $\begingroup$ Yes, you are correct. $\endgroup$
    – shmaxg
    Aug 18 '14 at 13:34
  • $\begingroup$ Note also that in general case after we found the "nearest" (in terms of the ratio) $\Omega_1$ and $\Omega_2$ such that $\Pr(\Omega_2) < \alpha \le \Pr(\Omega_1)$, the critical function would be $\varphi(x) = 1$ for $x \in \Omega_2$, $\varphi(x) = (\alpha - \alpha_0)/p_0$ for $x \in \Omega_1 \setminus \Omega_2$ and $\varphi(x) = \Omega \setminus \Omega_1$, where $\Omega$ is the sample space and $\alpha_0$, $p_0$ are defined above. $\endgroup$
    – shmaxg
    Aug 18 '14 at 13:48

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.