Hypothesis Testing. 1 Definitions of test statistics. CB: chapter 8; section 10.3

Transcription

1 Hypothesis Testing CB: chapter 8; section 0.3 Hypothesis: statement about an unknown population parameter Examples: The average age of males in Sweden is 7. (statement about population mean) The lowest time it takes to run 30 miles is hours. (statement about population max) Stocks are more volatile than bonds. (statement about variances of stock and bond returns) In hypothesis testing, you are interested in testing between two mutually exclusive hypotheses, called the null hypothesis (denoted H 0 ) and the alternative hypothesis (denoted H ). H 0 and H are complementary hypotheses, in the following sense: If the parameter being hypothesized about is θ, and the parameter space (i.e., possible values for θ) is Θ, then the null and alternative hypotheses form a partition of Θ: H 0 : θ Θ 0 Θ H : θ Θ c 0 (the complement of Θ 0 in Θ). Examples:. H 0 : θ = 0 vs. H : θ 0. H 0 : θ 0 vs. H : θ > 0 Definitions of test statistics A test statistic, similarly to an estimator, is just some real-valued function T n T (X,..., X n ) of your data sample X,..., X n. Clearly, a test statistic is a random variable. A test is a function mapping values of the test statistic into {0, }, where 0 implies that you accept the null hypothesis H 0 reject the alternative hypothesis H. implies that you reject the null hypothesis H 0 accept the alternative hypothesis H.

2 The subset of the real line R for which the test is equal to is called the rejection (or critical ) region. The complement of the critical region (in the support of the test statistic) is the acceptance region. Example: let µ denote the (unknown) mean male age in Sweden. You want to test: H 0 : µ = 7 vs. H : µ 7 Let your test statistic be X 00, the average age of 00 randomly-drawn Swedish males. Just as with estimators, there are many different possible tests, for a given pair of hypotheses H 0, H. Consider the following four tests:. Test : ( X00 [5, 9] ). Test : ( X00 9 ) 3. Test 3: ( X00 35 ) 4. Test 4: ( X00 7 ) Which ones make the most sense? Also, there are many possible test statistics, such as: (i) med 00 (sample median); (ii) max(x,..., X 00 ) (sample maximum); (iii) mode 00 (sample mode); (iv) sin X (the sine of the first observation). In what follows, we refer to a test as a combination of both (i) a test statistic; and (ii) the mapping from realizations of the test statistic to {0, }. Next we consider some common types of tests.. Likelihood Ratio Test Let: X = X,..., X n i.i.d. f(x θ), and likelihood function L(θ X) = n i= f(x i θ). Define: the likelihood ratio test statistic for testing H 0 : θ Θ 0 vs. H : θ Θ c 0 is λ( X) sup θ Θ 0 L(θ X) sup θ Θ L(θ X). The numerator of λ( X) is the restricted likelihood function, and the denominator is the unrestricted likelihood function.

3 The support of the LR test statistic is [0, ]. Intuitively speaking, if H 0 is true (i.e., θ Θ 0 ), then λ( X) = (since the restriction of θ Θ 0 will not bind). However, if H 0 is false, then λ( X) can be small (close to zero). So an LR test should be one which rejects H 0 when λ( X) is small; for example, (λ( X) < 0.75) Example: X,..., X n i.i.d. N(θ, ) Test H 0 : θ = vs. H : θ. Then λ( X) = exp ( i (X i ) ) exp ( i (X i X n ) ) (the denominator arises because X n is the unrestricted MLE estimator for θ. Example: X,..., X n U[0, θ]. (i) Test H 0 : θ = vs. H : θ. Restricted likelihood function { ( L( X ) ) n if max(x =,..., X n ) 0 if max(x,..., X n ) >. Unrestricted likelihood function { ( L( X θ) ) n if max(x = θ,..., X n ) θ 0 if max(x,..., X n ) > θ which is maximized at θn MLE = max(x,..., X n ). ( n, Hence the denominator of the LR statistic is max(x,...,x n)) so that { 0 if λ( X) max(x,..., X n ) > = ( ) n max(x,...,x n) otherwise LR test would say: (λ( X) c). Critical region consists of two disconnected parts (graph). 3

4 (ii) Test H 0 : θ [0, ] vs. H : θ >. In this case: the restricted likelihood is { ( so sup L( X θ) = θ [0,] n max(x,...,x n)) if max(x,..., X n ) 0 otherwise. { λ( X) if max(x,..., X = n ) 0 otherwise. () (graph) So now LR test is (λ( X) c) = (max(x,..., X n ) > ).. Wald Tests Another common way to generate test statistics is to focus on statistics which are asymptotically normal distributed, under H 0 (i.e., if H 0 were true). A common situation is when the estimator for θ, call it ˆθ n, is asymptotically normal, with some asymptotic variance V (eg. MLE). Let the null be H 0 : θ = θ 0. Then, if the null were true: (ˆθ n θ 0 ) n V The quantity on the LHS is the T-test statistic. d N(0, ). () Note: in most cases, the asymptotic variance V will not be known, and will also need to be estimated. However, if we have an estimator ˆV n such that ˆV p V, then the 4

5 statement Z n (ˆθ n θ 0 ) ˆV n d N(0, ) still holds (using the plim operator and Slutsky theorems). In what follows, therefore, we assume for simplicity that we know V. We consider two cases: (i) Two-sided test: H 0 : θ = θ 0 vs. H : θ θ 0. Under H 0 : the CLT holds, and the t-stat is N(0, ) Under H : assume that the true value is some θ θ 0. Then the t-stat can be written as (ˆθ n θ 0 ) = V n (ˆθ n θ ) + V n (θ θ 0 ). V n The first term d N(0, ), but the second (non-stochastic) term diverges to or, depending on whether the true θ exceeds or is less than θ 0. Hence the t-stat diverges to or with probability. Hence, in this case, your test should be ( Z n > c), where c should be some number in the tails of the N(0, ) distribution. Multivariate version: θ is K-dimensional, and asymptotic normal, so that under H 0, we have n(θn θ 0 ) d N(0, Σ). Then we can test H 0 : θ = θ 0 vs. H : θ θ 0 using the quadratic form Test takes the form: (Z n > c). Z n n (θ n θ 0 ) Σ (θ n θ 0 ) d χ k. (ii) One-sided test: H 0 : θ θ 0 vs. H : θ > θ 0. Here the null hypothesis specifies a whole range of true θ (Θ 0 = (, θ 0 ]), whereas the t-test statistic is evaluated at just one value of θ. Just as for the two-sided test, the one-sided t-stat is evaluated at θ 0, so that Z n = ˆθ n θ 0 n V. 5

6 Under H 0 and θ < θ 0 : Z n diverges to with probability. Under H 0 and θ = θ 0 : the CLT holds, and the t-stat is N(0, ). Under H : Z n diverges to with probability. Hence, in this case, you will reject the null only for very large values of Z n. Correspondingly, your test should be (Z n > c), where c should be some number in the right tail of the N(0, ) distribution. Later, we will discuss how to choose c..3 Score test Consider a model with log-likelihood function log L(θ X) = n Let H 0 : θ = θ 0. The sample score function evaluated at θ 0 is S(θ 0 ) θ log L(θ X) θ=θ0 = n i i log f(x i θ). θ log f(x i θ) θ=θ0. Define W i = θ log f(x i θ) θ=θ0. Under the null hypothesis, S(θ 0 ) converges to d dθ E θ0 W i = f(x θ) θ=θ 0 f(x θ 0 )dx f(x θ 0 ) = f(x θ)dx θ = θ = 0 (the information inequality). Hence, (Note that V 0 V θ0 W i = E θ0 W i = E θ0 ( θ log f(x θ) θ=θ0 ) V 0. is the usual variance matrix for MLE, which is the CRLB.) Therefore, you can apply the CLT to get that, under H 0, S(θ 0 ) n V 0 d N(0, ). (If we don t know V 0, we can use some consistent estimator ˆV 0 of it.) 6

7 ( So a test of H 0 : θ = θ 0 could be formulated as right tail of the N(0, ) distribution. Multivariate version: if θ is K-dimensional: n S(θ 0) n V 0 S n ns(θ 0 ) V 0 S(θ 0 ) d χ k. ) > c, where c is in the Recall that, in the previous lecture notes, we derived the following asymptotic equality for the MLE: n (ˆθn θ 0 ) a = n n n i = n S(θ 0) V 0. log f(x i θ) i θ θ=θ0 log f(x i θ) θ=θ0 θ (3) (The notation a = means that the LHS and RHS differ by some quantity which is o p ().) Hence, the above implies that (applying the information inequality, as we did before) ) a n (ˆθn θ 0 = n S(θ 0) d N(0, ) }{{} V }{{ 0 V } 0 () () The Wald statistic is based on (), while the Score statistic is based on (). In this sense, these two tests are asymptotically equivalent. Note: the asymptotic variance of the Wald statistic () equal the reciprocal of V 0. (Later, you will also see that the Likelihood Ratio Test statistic is also asymptotically equivalent to these two.) The LR, Wald, and Score tests (the trinity of test statistics) require different models to be estimated. LR test requires both the restricted and unrestricted models to be estimated Wald test requires only the unrestricted model to be estimated. 7

8 Score test requires only the restricted model to be estimated. Applicability of each test then depends on the nature of the hypotheses. For H 0 : θ = θ 0, the restricted model is trivial to estimate, and so LR or Score test might be preferred. For H 0 : θ θ 0, restricted model is a constrained maximization problem, so Wald test might be preferred. Methods of evaluating tests Consider X,..., X n i.i.d. (µ, σ ). Test statistic X n. Test H 0 : µ = vs. H : µ. Why are the following good or bad tests?. ( X n ). ( X n.) 3. ( X n [.8,.]) 4. ( X n [ 0, 30]) Test rejects too often (in fact, for every n, you reject with probability ). Test is even worse, since it rejects even when X n is close to. Test 3 is not so bad, Test 4 accepts too often. Basically, we are worried when a test is wrong. Since the test itself is a random variable, we cannot guarantee that a test is never wrong, but we can characterize how often it would be wrong. There are two types of mistakes that we are worried about: Type I error: rejecting H 0 when it is true. (This is the problem with tests and.) Type II error: Accepting H 0 when it is false. (This is the problem with test 4.) 8

9 Let T n T (X,..., X n ) denote the sample test statistic. Consider a test with rejection region R (i.e., test is (T n R)). Then: P (type I error) = P (T n R θ Θ 0 ) P (type II error) = P (T n R θ Θ c 0) Example: X, X i.i.d. Bernoulli, with probability p. Test H 0 : p = vs. H : p. Consider the test ( X +X ). Type I error: rejecting H 0 when p =. P(Type I error) = P( X +X p = ) = P( X +X = 0 p = )+ P( X +X p = ) = + = Type II error: Accepting H 0 when p. For p : X + X So P( X +X = p)=p. Graph. 0 with prob ( p) = with prob ( p)p with prob p = Power function More generally, type I and type II errors are summarized in the power function. 9

10 Definition: the power function of a hypothesis test with a rejection region R is the function of θ defined by β(θ) = P (T n R θ). Example: For above example, β(p) = P ( X +X p ) = p. Graph Power function gives the Type I error probabilities, for any singleton null hypothesis H 0 : θ = θ 0. From β(p), see that if you are worried about Type I error, then you should only use this test when your null is that p is close to (because only for p 0 close to is the power function low). β(p) gives you the Type II error probabilities, for any point alternative hypothesis So if you are worried about Type II error, then β(p) tells you that you should use this test when your alternative hypothesis postulates that p is low (close to zero). We say that this test has good power against alternative values of p close to zero. Important: power function is specific to a given test (T n R), regardless of the specific hypotheses that the test may be used for. Example: X,..., X n U[0, θ]. Test H 0 : θ vs. H : θ >. Derive β(θ) for the LR test (λ( X) < c). 0

11 Recall our earlier derivation of LR test in Eq. (). Hence, Graph. β(θ) = P (λ( X) < c θ) = P (max(x,..., X n ) > θ) = P (max(x,..., X n ) < θ) { 0 for θ = ( n θ) for θ >. In practice, researchers often concerned about Type I error (i.e., don t want to reject H 0 unless evidence overwhelming against it): conservative bias? But if this is so, then you want a test with a power function β(θ) which is low for θ Θ 0, but high elsewhere: This motivates the definition of size and level of a test. For 0 α, a test with power function β(θ) is a size α test if sup θ Θ0 β(θ) = α. For 0 α, a test with power function β(θ) is a level α test if sup θ Θ0 β(θ) α. The θ Θ 0 at which the sup is achieved is called the least favorable value of θ under the null, for this test. It is the value of θ Θ 0 for which the null holds, but which is most difficult to distinguish (in the sense of having the highest rejection probability) from any alternative parameter θ Θ 0.

12 Reflecting perhaps the conservative bias, researcher often use tests of size α =0.05, or 0.0. Example: X,..., X n i.i.d. N(µ, ). Then X n N(µ, /n), and Z n (µ) n( Xn µ) N(0, ). Consider the test (Z n () > c), for the hypotheses H 0 : µ vs. H : µ >. The power function β(µ) = P ( n( X n ) > c µ) = P ( n( X n µ) > c + n( µ) µ) = Φ(c + n( µ)) where Φ( ) is the standard normal CDF. Note that β(µ) is increasing in µ. Size of test= sup µ Φ(c + n( µ)). Since β(µ) is increasing in µ, the supremum occurs at µ =, so that size is β() = Φ(c). Assume you want a test with size α. Then you want to set c such that Φ(c) = α c = Φ ( α). Graph: c is the ( α)-th quantile of the standard normal distribution. You can get these from the usual tables. For α = 0.05, then c =.96. For α = 0.05, then c =.64. Now consider the above test, with c =.96, but change the hypotheses to H 0 : µ = vs. H : µ.

13 Test still has size α = But there is something intuitively wrong about this test. You are less likely to reject when µ <. So the Type II error is very high for the alternatives µ <. We wish to rule out such tests Definition: a test with power function β(θ) is unbiased if β(θ ) β(θ ) for every pair (θ, θ ) where θ Θ c 0 and θ Θ 0. Clearly, the test above is biased for the stated hypotheses. What would be an unbiased test, with the same size α = 0.05? Definition: Let C be a class of tests for testing H 0 : θ Θ 0 vs. H : θ Θ c 0. A test in class C, with power function β(θ), is uniformly most powerful (UMP) in class C if, for every other test in class C with power function β(θ) β(θ) β(θ), for every θ Θ c 0. Often, the classes you consider are test of a given size α. Graphically, the power function for a UMP test lies above the upper envelope of power functions for all other tests in the class, for θ Θ c 0: (graph). UMP tests in special case: two simple hypotheses It can be difficult in general to see what form an UMP test for any given size α is. But in the simple case where both the null and alternative hypotheses are simple hypotheses, we can appeal to the following result. Theorem 8.3. (Neyman-Pearson Lemma): Consider testing H 0 : θ = θ 0 vs. H : θ = θ 3

14 (so both the null and alternative are singletons). The pdf or pmf corresponding to θ i is f( X θ i ), for i = 0,. The test has a rejection region R which satisfies: X R if f( X θ ) > k f( X θ 0 ) ( ) X R c if f( X θ ) < k f( X θ 0 ) and α = P ( X R θ 0 ). (4) Then: Any test satisfying (*) is a UMP test with level α. If there exists a test satisfying (*) with k > 0, then every UMP level α test is a size α test and every UMP level α test satisfies (*). In other words, for simple hypotheses, a likelihood ratio test of the sort ( f( X θ ) ) f( X θ 0 ) > k is UMP-size α (where k is chosen so that P ( f( X θ ) f( X θ 0 ) > k θ 0) = α). Note that this LR statistic is different than λ( X), which for these hypotheses is f( X θ 0 ) max[f( X θ 0 ),f( X θ )]. Example: return to -coin toss again. Test H 0 : p = vs. H : p = 3 4. The likelihood ratios for the three possible outcomes i X i = 0,, are: f(0 p = 3 4 ) f(0 p = ) = 4 f( p = 3 4 ) f( p = ) = 3 4 f( p = 3 4 ) f( p = ) = 9 4. Let l( X) = f( i X i p= 3 4 ) f( i X i p= ). Hence, there are 4 possible rejection regions, for values of l( X): 4

15 (i) ( 9, + ) (size α = 0) 4 (ii) ( 3, + ) (size α = ) 4 4 (iii) (, + ) (size α = 3) 4 4 (iv) (, + ) (size α = ). At all other values for α, no value of k satisfies (*) exactly, and so the test above are level-α tests. This is due to the discreteness of this example. Example A: X,..., X n N(µ, ). Test H 0 : µ = µ 0 vs. H : µ = µ. (Assume µ > µ 0.) By the NP-Lemma, the UMP test has rejection region characterized by Taking logs of both sides: exp ( i (X i µ ) ) exp ( i (X i µ 0 ) ) > k. (µ µ 0 ) i X i > log k + n (µ µ 0 ) X i > n i n log k + (µ µ 0 ) (µ µ 0 ) d. where d is determined such that P ( X n > d) = α, where α is the desired size. This makes intuitive sense: you reject the null when the sample mean is too large (because µ > µ 0 ). Under the null, n( X n µ 0 ) N(0, ), so for α = 0.05, you want to set d = µ / n. 5

16 .. Discussion of NP Lemma (From Amemiya pp. 90-9) Rather than present a proof of NP Lemma (you can find one in CB), let s consider some intuition for the NP test. In doing so, we will derive an interesting duality between Bayesian and classical approaches to hypothesis testing (the NP Lemma being a seminal result for the latter). Start by considering a Bayesian approach to this testing problem. Assume that decisionmaker incurs loss γ if he mistakenly chooses H when H 0 is true, and γ if he mistakenly chooses H 0 when H is true. Then, given data observations x, he will Reject H 0 (=accept H ) γ P (θ 0 x) < γ P (θ x) ( ) where P (θ 0 x) denotes the posterior probability of the null hypothesis given data x. In other words, this Bayesian s rejection region R 0 is given by { R 0 = x : P (θ x) P (θ 0 x) > γ }. γ If we multiply and divide the ratio of posterior probabilities by f( x), the (marginal) joint density of x, and use the laws of probability, we get: { R 0 = x : P (θ x)f( x) P (θ 0 x)f( x) > γ } γ { = x : L( x θ )P (θ ) L( x θ 0 )P (θ 0 ) > γ } γ { = x : L( x θ ) L( x θ 0 ) > γ } P (θ 0 ) γ P (θ ) c. In the above, P (θ 0 ) and P (θ ) denote the prior probabilities for, respectively, the null and alternative hypotheses. This provides some intuition for the likelihood ratio form of the NP rejection region. Next we show a duality between the Bayesian and classical approaches to finding the best test. In the case of two simple hypotheses, a best test should be such that for a given size α, it should have the smallest type error β (these are called admissible tests). For any testing scenario, the frontier (in (β, α) space) of tests is convex. (Why?) Both the Bayesian and the classical statistician want to choose a test that is on the frontier, but they might differ in the one they choose. First consider the Bayesian. She wishes to employ a test, equivalently, choose a rejection region R, to minimize expected loss min φ(r) γ P (θ 0 X R)P ( X R) + γ P (θ X R)P ( X R). R 6

17 We will show that the region R 0 defined above optimizes this problem. Consider any other region R. Recall that R 0 = (R 0 R ) (R 0 R ). Then we can rewrite φ(r 0 ) =γ P (θ 0 R 0 R )P (R 0 R ) + γ P (θ 0 R 0 R )P (R 0 R ) + γ P (θ R 0 R )P ( R 0 R ) + γ P (θ R 0 R )P ( R 0 R ). φ(r ) =γ P (θ 0 R R 0 )P (R R 0 ) + γ P (θ 0 R R 0 )P (R R 0 ) + γ P (θ R R 0 )P ( R R 0 ) + γ P (θ R R 0 )P ( R R 0 ). First and fourth terms of equations above are identical. From the definition of R 0 above, we know that φ(r 0 ) ii < φ(r ) iii. That is, for all x R 0 R 0 R, we know from (*) that γ P (θ 0 x) < γ P (θ x). Similarly, we know that φ(r 0 ) iii < φ(r ) ii, implying that φ(r 0 ) < φ(r ). Moreover, using the laws of probability, we can re-express φ(r) = γ P (θ 0 )P (R θ 0 ) + γ P (θ )P ( R θ ) η 0 α R + η β R with η 0 = γ P (θ 0 ) and η = γ P (θ ). In (β, α) space, the iso-loss curves are parallel lines with slope η 0 /η, decreasing towards the origin. Hence, the optimal test (with region R 0 ) lies at tangency of (α, β) frontier with a line with slope η 0 /η. Hence, the optimal Bayesian test (R 0 ). Now consider the classical statistician. He doesn t want to specify priors P (θ 0 ), P (θ ), so η 0, η are not fixed. But he is willing to specify a desired size α. Hence the optimal test is the one with a rejection region characterized by the slope of the line tangent at α in the (β, α) space. This is in fact the NP test. From the above calculations, we know that this slope is equivalent to c. That is, the NP test corresponds to an optimal Bayesian test with η 0 /η = c.. Extension of NP Lemma: models with monotone likelihood ratio The case covered by NP Lemma that both null and alternative hypotheses are simple is somewhat artificial. For instance, we may be interested in the one-sided hypotheses H 0 : θ θ 0 vs. H : θ θ 0, where θ is scalar. It turns out UMP for one-sided hypotheses { exists under an additional assumption on the family of density functions f( X θ) } : θ Definition: the family of densities f( X θ) has monotone likelihood ratio in T ( X) is there exists a function T ( X) such that for any pair θ < θ the densities f( X θ) 7

18 and f( X θ ) are distinct and the ratio f( X θ )/f( X θ) is a nondecreasing function of T ( X). That is, there exists a nondecreasing function g( ) and function T ( X) such that f( X θ )/f( X θ) = g(t ( X)). For instance: let X = x and T (x) = x, then MLR in x means that f(x θ )/f(x θ) is nondecreasing in x. Roughly speaking: larger x s are more likely under larger θ s. Theorem: Let θ be scalar, and let f( X θ) have MLR in T ( X). Then: (i) For testing H 0 : θ θ 0 vs. H : θ θ 0, there exists a UMP test φ( X), which is given by φ( X) when T ( X) > C = γ when T ( X) = C 0 when T ( X) < C where (γ, C) are chosen to satisfy E θ0 φ( X) = α. (ii) The power function β(θ) = E θ φ( X) is strictly increasing for all points θ for which 0 < β(θ) <. (iii) For all θ, the test described in part i is UMP for testing H 0 : θ θ vs. H : θ θ at level α = β(θ ). Sketch of Proof: Consider testing θ 0 vs. θ > θ 0. By NP Lemma, this depends on ratio f( X θ )/f( X θ 0 ). Given MLR condition, this ratio can be written g(t ( X)) where g( ) is nondecreasing. Then UMP test rejects when f( X θ )/f( X θ 0 ) is large, which is when T (X) is large; this is test φ( X). Since this test does not depend on θ, φ( X) is also UMP-size α for testing H 0 : θ 0 vs. H : θ > θ 0 (composite alternative). Now since MLR holds for all (θ, θ ; θ > θ ), the test φ( X) is also UMP-size E θ φ( X) for testing θ vs. θ. Hence β(θ ) β(θ ), otherwise φ( X) cannot be UMP (why? ). Furthermore, the distinctiveness of f( X θ ) and f( X θ ) rules out β(θ ) = β(θ ). Hence we get (ii). Then since the power function is monotonic increasing in θ, the UMP-size α feature of φ( X) for testing θ 0 vs. H : θ > θ 0 extends to the composite null hypothesis H 0 : θ θ 0 vs. H : θ > θ 0, which is (i). (iii) follows immediately. Example A cont d: From above, we see that for µ > µ 0, the likelihood ratio simplifies to exp ( (µ µ 0 ) i X i n (µ µ 0) ) which is increasing in X n. Hence, Lehmann and Romano, Testing Statistical Hypotheses, pg. 65. Consider the purely random test φ( X) = with probability α. 8

19 this satisfies MLR with T ( X) = X n. Using the theorem above, the one-sided T-test which rejects when X n > µ n is also UMP for size α = 0.05 for the one-sided hypotheses H 0 : µ µ 0 vs. H : µ > µ 0. Call this test. Taking this example further, we have that for the one-sided hypotheses H 0 : µ > µ 0 vs. H : µ < µ 0, the one-sided T-test which rejects when X n < µ 0.64 n will be UMP for size α = Call this test. Now consider testing H 0 : µ = µ 0 vs. H : µ µ 0. The alternative hypothesis is equivalent to µ < µ 0 or µ > µ 0. Can we find an UMP size-α test? Note that both Test and Test are size-α tests for hypotheses (*). So we consider each in turn. For any alternative point µ > µ 0, test is UMP, implying that β (µ ) is maximal among all size-α tests. For µ < µ 0, however, test is UMP, implying that β (µ ) is maximal. Furthermore, β (µ ) < α < β (µ ) from part (ii) of theorem above, so neither test can be uniformly most powerful for all µ µ 0. And indeed there is no UMP size-α test for problem (*). But note that both Test and Test are biased for the hypotheses (*). It turns out that the two-sided T-test which rejects when X n > µ / n or X n < µ 0.96/ n is UMP among size-α unbiased tests. See discussion in CB, pp ( ) 9

20 3 Large-sample properties of tests In practice, we use large-sample theory that is, LLN s and CLT s in order to determine the approximate critical regions for the most common test statistics. Why? Because finite-sample properties can be difficult to determine: Example: X,..., X n i.i.d. Bernoulli with prob. p. Want to test H 0 : p vs. H : p >, using the test stat X n = n i X i. n is finite. The exact finite-sample distribution for X n is the distribution of times n a B(n, p) random variable, which is: 0 with prob ( ) n 0 ( p) n with prob ( ) n n p( p) n with prob ( ) n n p ( p) n with prob p n Assume your test is of the following form: ( X n > c), where the critical value c is to be determined such that the size sup p P ( X n > c p) = α, for some specified α. 3 This equation is difficult to solve for! On the other hand, by the CLT, we know that n(x n p) d N(0, ). Hence, consider p( p) the T-test statistic Z n n ( ) Xn / = n ( Xn 4 ). Under any p in the null hypothesis, P (Z n > Φ ( α)) α, for n large enough. (In fact, this equation holds with equality for p =, and holds with strict inequality for p <.) Corresponding to this test, you can derive the asymptotic power function, which is β a (p) lim n P (Z n > c), for c = Φ ( α): (Graph) 3 Indeed, the Clopper-Pearson (934) confidence intervals for p are based on inverting this exact finite-sample test. 0

21 Note that the asymptotic power function is equal to at all values under the alternative. This is the notion for consistency for a test: that it has asymptotic power under every alternative. Note also that asymptotic power (rejection probability) is zero under every p of the null, except p =. (skip) Accordingly, we see that asymptotic power, vs. fixed alternatives, is not a sufficiently discerning criterion for distinguishing between tests. We can deal with this by considering local alternatives of the sort p = + h/ n. Under additional smoothness assumptions on the distributional convergence of n( X n p) p( p) around p =, we can obtain asymptotic power functions under these local alternatives. 3. Likelihood Ratio Test Statistic: asymptotic distribution Theorem 0.3. (Wilks Theorem): For testing H 0 : θ = θ 0 vs. H : θ θ 0, if X,..., X n i.i.d. f(x θ), and f(x θ) satisfies the regularity conditions in Section Then under H 0, as n : log λ( X) d χ. Note: χ denotes a random variable from the Chi-squared distribution with degree of freedom. By Lemma 5.3. in CB, if Z N(0, ), then Z χ. Clearly, χ random variables only have positive support. Proof: Assume null holds. Use Taylor-series expansion of log-likelihood function around the MLE estimator ˆθ n : log f(x i θ 0 ) = i i + = i log f(x i ˆθ n ) + i i θ log f(x i θ) θ=ˆθn (θ 0 ˆθ n ) θ log f(x i θ) θ=ˆθn (θ 0 ˆθ n ) +... log f(x i ˆθ n ) + i θ log f(x i θ) θ=ˆθn (θ 0 ˆθ n ) +... where (i) the second term disappeared because the MLE ˆθ n sets the first-order condition i log f(x θ i θ) θ=ˆθn = 0; (ii) the remainder term is o p (). This is a secondorder Taylor expansion. (5)

22 Rewrite the above as: ( ) f(x i θ 0 ) log = i f(x i ˆθ n ) n θ log f(x i θ) [ n(θ0 θ=ˆθn ˆθ ] n ) + op (). i Now n i θ log f(x i θ) θ=ˆθn p Eθ0 θ log f(x i θ 0 ) = V 0 (θ 0 ), where V 0 (θ 0 ) denotes the asymptotic variance of the MLE estimator (and is equal to the familiar information bound /I(θ 0 )). Finally, we note that variable: n(ˆθ n θ 0) V0 (θ 0 ) d N(0, ). Hence, by the definition of the χ random log λ( X) d χ. In the multivariate case (θ being k-dimensional), the above says that log(λ( X)) a = n(θ 0 θ n )I(θ 0 ) (θ 0 θ n ) χ k. Hence, LR-statistic is asymptotically equivalent to Wald and Score tests. Example: X,..., X N i.i.d. Bernoulli with prob. p. Test H 0 : p = vs. H : p. Let Y n denote the number of s. λ( X) = For test with asymptotic size α: ( n ) ( ) yn ( n yn y n ) ( n ) ( yn ) yn ( n yn ) n yn. y n n n α = P (λ( X) c) (c < ) = P ( log λ( X) log c) = P (χ log c) = F χ ( log c) ( c = exp ) F χ ( α). x For instance, for α = 0.05, then F χ ( α) = 3.84; α = 0.0, then F x χ ( α) = x.706.