Assume we have two independent random samples $(X_i)_{i=1}^{n}$ and $(Y_i)_{i=1}^{m}$ where $X_i\sim\mathrm{Uniform}(0,\theta_1)$ and $Y_j\sim\mathrm{Uniform}(0,\theta_2)$. To test, at the significant level $\alpha$, $H_0: \theta_1 = \theta_2$ against $H_1:\theta_1\ne\theta_2$ using Likelihood Ratio Test, denoting by $Z$ the ratio of the $n$-th ($m$-th resp.) ordered statistics of two samples, i.e., $Z = \frac{X_{(n)}}{Y_{(m)}}$, we may find:

$$ \Lambda = \frac{ \left( \max{X_i} \right) ^ {n} \left( \max{Y_j} \right) ^ m }{ \left( \max \left(X_i, Y_j \right) \right) ^ {n + m} } = \begin{cases} Z ^ {-m} \quad X_{(n)} \ge Y_{(m)} \\ Z^{n} \quad X_{(n)} < Y_{(m)} \end{cases} \mathrm{.} $$

The form of the rejection region $C$ is $C = \{Z\le c_1 <1\}\cup\{Z\ge c_2 > 1\} $ by considering the cases where $X_{(n)} < Y_{(m)}$ or otherwise respectively. To further determine the constants $0 < c_1 \le 1$ and $0 < c_2 \le 1$, I am not sure whether to set:

$$ \mathrm{Pr} \left( Z \ge c_2 \vert Z \ge 1\right) = \mathrm{Pr} \left (Z \le c_1 \vert Z < 1 \right) \overset{\mathrm{set}}{=} \alpha\mathrm{,} $$

or: $$ \mathrm{Pr}\left( Z \ge c_2 \right) = \mathrm{Pr}\left( Z \le c_1 \right) \overset{\mathrm{set}}{=} \frac{\alpha}{2} \mathrm{.} $$ Since for two independent samples with size $n$ and $m$ following the same uniform disribution $\mathrm{Uniform}(0,\beta)$, with the maxima of the observations being $V$ and $W$ resepectively, we have $\mathrm{Pr}(V\le kW) = \frac{mk^n}{n+m}$ for some constant $0<k<1$. If I follow the first idea, I can get:

$$ C = \left\{ Z \le \alpha ^ {\frac{1}{n}} \right\} \cup \left\{ Z \ge \alpha ^ {-\frac{1}{m}} \right\} \mathrm{.} $$

Since $$ \begin{align} \mathrm{Pr} \left (Z \ge c_2 \vert Z \ge 1 \right) & = \frac{ \mathrm{Pr} \left( Z \ge c_2, Z \ge 1 \right) }{ \mathrm{Pr} \left ( Z \ge 1 \right)} \\\\ & = \frac{ \mathrm{Pr} \left( Y_{(m)} \le \frac{1}{c_2} X_{(n)} \right) } { \mathrm{Pr} \left( Y_{(m)} \le X_{(n)} \right) } \\\\ & = \frac{ \frac{n}{n + m} \left( \frac{1}{c_2} \right) ^ {m} } { \frac{n}{n + m} } \\\\ & = c_{2}^{-m} \\\\ & \overset{ \mathrm{set} }{=} \alpha \end{align} $$ Hence $c_2 = \alpha ^ {-\frac{1}{m}} $, and similarly $c_1 = \alpha ^ {\frac{1}{n}} $.

But if I follow the second idea, since $\frac{m}{m+n}$ and $\frac{n}{m+n}$ are both less than $1$, I get:

$$ C = \left\{ Z \le \left( \frac{n+m}{m} \frac{\alpha}{2} \right) ^ {\frac{1}{n}} \right\} \cup \left\{ Z \ge \left( \frac{n+m}{n} \frac{\alpha}{2} \right)^{-\frac{1}{m}} \right\} \mathrm{.} $$

Apparently, I have made a mistake since those two do not agree if two samples are not of the equal size. Nevertheless, I have trouble in locating where did my argument go wrong.

Thank you very much in advance.


  1. I found a good discussion in this thread on finding the joint p.d.f. of $X_{(n)}$ and $Y_{(n)}$ (and more) for this test. Nevertheless, it remains a question for me to determine the rejection region.
  2. I think this is similar to the the cases where we perform a two-sided LRT where the likelihood ratio is a step function. For example, in a two-sided $F$ test for the variances of two normal samples ($H_0: \sigma_{0}^2=\sigma_{1}^2$ against $H_1: \sigma_{0}^2\ne\sigma_{1}^2$), the standard procedure follows the second idea above. Also of interests, in DeGroot (p. 603, 4th Ed.), it is noted while deriving the two sided $F$ test from LRT that:

    Unfortunately, it is usually tedious to compute the necessary values $c_1$ and $c_2$. For this reason, people often abandon the strict likelihood ratio criterion in this case and simply let $c_1$ and $c_2$ be the $\frac{\alpha_0}{2}$ and $1 - \frac{\alpha_0}{2}$ quantiles of the appropriate $F$ distribution.

    Might it also be theoretically justifiable to derive the rejection region using the first idea?

  3. This question is also posted on Quora.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.