Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

A family of pdfs or pmfs ${g(t|theta):thetainTheta}$ for a univariate random variable $T$ with real-valued parameter $theta$ has a monotone likelihood ratio (MLR) if, for every $theta_2>theta_1$, $g(t|theta_2)/g(t|theta_1)$ is a monotone (nonincreasing or nondecreasing) function of $t$ on ${t:g(t|theta_1)>0;text{or};g(t|theta_2)>0}$. Note that $c/0$ is defined as $infty$ if $0< c$.
Consider testing $H_0:thetaleq theta_0$ versus $H_1:theta>theta_0$. Suppose that $T$ is a sufficient statistic for $theta$ and the family of pdfs or pmfs ${g(t|theta):thetainTheta}$ of $T$ has an MLR. Then for any $t_0$, the test that rejects $H_0$ if and only if $T >t_0$ is a UMP level $alpha$ test, where $alpha=P_{theta_0}(T >t_0)$.
Example 1
To better understand the theorem, consider a single observation, $X$, from $mathrm{n}(theta,1)$, and test the following hypotheses: $$H_0:thetaleq theta_0quadmathrm{versus}quad H_1:theta>theta_0.$$ Then $theta_1>theta_0$, and the likelihood ratio test statistics would be $$lambda(x)=frac{f(x|theta_1)}{f(x|theta_0)}.$$ And we say that the null hypothesis is rejected if $lambda(x)>k$. To see if the distribution of the sample has MLR property, we simplify the above equation as follows: begin{aligned} lambda(x)&=frac{frac{1}{sqrt{2pi}}expleft[-frac{(x-theta_1)^2}{2}right]}{frac{1}{sqrt{2pi}}expleft[-frac{(x-theta_0)^2}{2}right]}\ &=exp left[-frac{x^2-2xtheta_1+theta_1^2}{2}+frac{x^2-2xtheta_0+theta_0^2}{2}right]\ &=expleft[frac{2xtheta_1-theta_1^2-2xtheta_0+theta_0^2}{2}right]\ &=expleft[frac{2x(theta_1-theta_0)-(theta_1^2-theta_0^2)}{2}right]\ &=expleft[x(theta_1-theta_0)right]timesexpleft[-frac{theta_1^2-theta_0^2}{2}right] end{aligned} which is increasing as a function of $x$, since $theta_1>theta_0$.
 Figure 1. Normal Densities with $mu=1,2$.
By illustration, consider Figure 1. The plot of the likelihood ratio of these models is monotone increasing as seen in Figure 2, where rejecting $H_0$ if $lambda(x)>k$ is equivalent to rejecting it if $Tgeq t_0$.
 Figure 2. Likelihood Ratio of the Normal Densities.
And by factorization theorem the likelihood ratio test statistic can be written as a function of the sufficient statistics since the term, $h(x)$ will be cancelled out. That is, $$lambda(t)=frac{g(t|theta_1)}{g(t|theta_0)}.$$ And by Karlin-Rubin theorem, the rejection region $R={t:t>t_0}$ is a uniformly most powerful level-$alpha$ test. Where $t_0$ satisfies the following: begin{aligned} mathrm{P}(T>t_0|theta_0)&=mathrm{P}(Tin R|theta_0)\ alpha&=1-mathrm{P}(Xleq t_0|theta_0)\ 1-alpha&=int_{-infty}^{t_0}frac{1}{sqrt{2pi}}expleft[-frac{(x-theta_0)^2}{2}right]operatorname{d}x end{aligned} Hence the quantile of the $1-alpha$ probability, which is $z_{alpha}$ is equal to $t_0$, that is $z_{alpha}=t_0$, and thus we reject $H_0$ if $T>z_{alpha}$.

Example 2
Now consider testing the hypotheses, $H_0:thetageq theta_0$ versus $H_1:theta< theta_0$ using the sample $X$ (single observation) from Beta($theta$, 2), and to be more specific let $theta_0=4$ and $theta_1=3$. Can we apply Karlin-Rubin? Of course! Visually, we have something like in Figure 3.
 Figure 3. Beta Densities Under Different Parameters.
Note that for this test, $theta_1 k$ if and only if $T < t_0$. Where $t_0$ satisfies the following equations: begin{aligned} mathrm{P}(T < t_0|theta_0)&=mathrm{P}(X < t_0|theta_0)\ alpha&=int_{0}^{t_0}frac{Gamma(theta_0+2)}{Gamma(theta_0)Gamma(2)}x^{theta_0-1}(1-x)^{2-1}operatorname{d}x\ alpha&=int_{0}^{t_0}frac{Gamma(6)}{Gamma(4)Gamma(2)}x^{3}(1-x)operatorname{d}x. end{aligned} Hence the quantile of the $alpha$ probability, $x_{alpha}=t_0$. And thus we reject $H_0$ if $T < x_{alpha}$.
 Figure 4. Likelihood Ratio of the Beta Densities.

### Reference

1. Casella, G. and Berger, R.L. (2001). Statistical Inference. Thomson Learning, Inc.