Likelihood Ratio Test, Most Powerful Test, Uniformly Most Powerful Test
A. We first review the general definition of a hypothesis test. A hypothesis test is like a lawsuit:
H 0 : the defendant is innocent versus
H a : the defendant is guilty The truth
H 0 : the defendant H a the defendant is innocent is guilty
Jury’s H 0 Right decision Type II error Decision H a Type I error Right decision
The significance level and the power of the test.
α = P(Type I error) = P(Reject H 0 | H 0 ) ← signi icance level
β = P(Type II error) = P(Fail to reject H 0 | H a )
Power = 1- β = P(Reject H 0 | H a )
B. Now we derive the likelihood ratio test for the usual two- sided hypotheses for a population mean. (It has a simple null hypothesis and a composite alternative hypothesis.)
Example 1. Please derive the likelihood ratio test for H0: μ = μ0 versus Ha: μ ≠ μ0, when the population is normal and population variance σ2 is known.
Solution: For a 2-sided test of H0: μ = μ0 versus Ha: μ ≠ μ0, when the population is normal and population variance σ2 is known, we have:
1
: 0 and :
The likelihoods are:
L L0 n 2 n 1 x 1 2 1 n 2 exp i 0 exp x i1 2 2 2 2 i1 i 0 2 2 2 2
There is no free parameter in L , thus Lˆ L .
L L n 2 n 1 x 1 2 1 n 2 exp i exp x i1 2 2 2 2 i1 i 2 2 2 2
There is only one free parameter μ in L. Now we shall find the value of μ that maximizes the log likelihood n 1 n ln L ln2 2 x 2 . 2 2 2 i1 i d ln L 1 n By solving x 0 , we have ˆ x d 2 i1 i It is easy to verify that ˆ x indeed maximizes the loglikelihood, and thus the likelihood function.
Therefore the likelihood ratio is: Lˆ L 0 ˆ L max L n 1 2 1 n 2 exp x L 2 2 2 2 i1 i 0 0 Lˆ n 1 2 1 n 2 exp x x 2 2 i1i 2 2
1 n 2 2 exp x x x 2 i1i 0 i 2 2 1 x 0 1 2 exp exp z0 2 / n 2
2
Therefore, the likelihood ratio test that will reject H0 when * is equivalent to the z-test that will reject H0 when
Z0 c , where c can be determined by the significance level α as c z / 2 .
C. MP Test, UMP Test, and the Neyman-Pearson Lemma
Now considering some one-sided tests where we have :
H0 : 0 versus H a : 0 or, H0 : 0 versus H a : 0 Given the composite null hypothesis for the second one-sided test, we need to expand our definition of the significance level as follows:
For 퐻 : 휃 ∈ 휔 versus 퐻 : 휃 ∈ 휔