Statistics 550 Notes 20 Reading: Section 4.1-4.2 I. Hypothesis Testing Basic Setup (Chapter 4.1) Review Motivating Example: A graphologist claims to be able to identify the writing of a schizophrenic person from a nonschizophrenic person. The graphologist is given a set of 10 folders, each containing handwriting samples of two persons, one nonschizophrenic and the other schizophrenic. In an experiment, the graphologist made 6 correct identifications. Is there strong evidence that the graphologist is able to better identify the writing of schizophrenics better than a person who was randomly guessing would? Probability model: Let p be the probability that the graphologist successfully identifies the writing of a randomly chosen schizophrenic vs. nonschizophrenic person. A reasonable model is that the 10 trials are iid Bernoulli with probability of success p . Hypotheses: H 0 : p 0.5 H1 : p 0.5 The alternative/research hypothesis H1 should be the hypothesis we’re trying to establish. 1 Test Statistic: W ( X1 , , X n ) that is used to decide whether to accept or reject H 0 . In the motivating example, a natural test statistic is W ( X1 , , X10 ) i 1 X i , the number of successful 10 identifications of schizophrenics the graphologist makes. The observed value of this test statistic is 6. Critical region: Region of values C of the test statistic for which we reject the null hypothesis, e.g., C {6, 7,8,9,10} . Errors in Hypothesis Testing: True State of Nature Decision H1 is true H 0 is true Type I error Correct decision Reject H 0 Accept (retain) H 0 Correct decision Type II error The best critical region would make the probability of a Type I error small when H 0 is true and the probability of a Type II error small when H1 is true. But in general there is a tradeoff between these two types of errors. Size of a test: We say that a test with test statistic W ( X1 , , X n ) and critical region C is of size if max P [W ( X1 , 0 , X n ) C] . 2 The size of the test is the maximum probability (where the maximum is taken over all that are part of the null hypothesis; for the motivating example, the null hypothesis only has one value of in it) of making a Type I error when the null hypothesis is true. A test with size is said to be a test of (significance) level . Suppose we use a critical region C {6, 7,8,9,10} with the , X10 ) i 1 X i for the motivating 10 test statistic W ( X1 , example. The size of the test is Pp 0.5 (Y 6) 0.377 where Y has a binomial distribution with n=10 and probability p=0.5. Power: The power of a test at an alternative 1 is the probability of making a correct decision when is the true parameter (i.e., the probability of not making a Type II error when is the true parameter). The power of the test with test statistic , X10 ) i 1 X i and critical region C {6, 7,8,9,10} at p=0.6 is Pp 0.6 (Y 6) 0.633 and at W ( X1 , 10 p=0.7 is Pp 0.7 (Y 7) 0.850 where Y has a binomial distribution with n=10 and probability p. The power depends on the specific parameter in the alternative hypothesis that is being considered. Power function: C ( ) P [W ( X1 , , X n ) C ]; 1 . 3 Neyman-Pearson paradigm: Set the size of the test to be at most some small level, typically 0.10, 0.05 or 0.01 (most commonly 0.05) in order to protect against Type I errors. Then among tests that have this size, choose the one that has the “best” power function. In Chapter 4.2, we will define more precisely what we mean by “best” power function and derive optimal tests for certain situations. For the test statistic W ( X1 , , X10 ) i 1 X i , the critical region C {6, 7,8,9,10} has a size of 0.377; this gives too high a probability of Type I error. The critical region C {8,9,10} has a size of 0.0547, which makes the probability of a Type I error reasonably small. Using C {8,9,10} , we retain the null hypothesis for the actual experiment for which W was equal to 6. 10 P-value: For a test statistic W ( X1 , , X n ) , consider a family of critical regions {C : } each with different sizes. For the observed value of the test statistic Wobs from the sample, consider the subset of critical regions for which we would reject the null hypothesis, {C : Wobs C } . The p-value is the minimum size of the tests in the subset {C : Wobs C } , p-value = min Size(test with critical region C ) . {C :Wobs C } The p-value is the minimum significance level for which we would still reject the null hypothesis. 4 The p-value is a measure of how much evidence there is against the null hypothesis; it is the minimum significance level for which we would still reject the null hypothesis. Consider the family of critical regions Ci {i 1 X i i} for the motivating example. Since the graphologist made 6 correct identifications, we reject the null hypothesis for critical regions Ci , i 6 . The minimum size of the critical regions Ci , i 6 is for i=6 and equals 0.377. The p-values is thus 0.377. 10 Scale of evidence p-value <0.01 Evidence very strong evidence against the null hypothesis Strong evidence against the null hypothesis weak evidence against the null hypothesis little or no evidence against the null hypothesis 0.01-0.05 0.05-0.10 >0.1 Warnings: (1) A large p-value is not strong evidence in favor of H 0 . A large p-value can occur for two reasons: (i) H 0 is true or (ii) H 0 is false but the test has low power at the true alternative. 5 (2) Do not confuse the p-value with P( H 0 | Data) . The pvalue is not the probability that the null hypothesis is true. II. Testing simple versus simple hypotheses: Bayes procedures Consider testing a simple null hypothesis H 0 : 0 versus a simple alternative hypothesis H1 : 1 , i.e., under the null hypothesis X ~ P( X | 0 ) P0 and under the alternative hypothesis X ~ P( X | 1 ) P1 . Example 1: X 1 , , X n iid N ( ,1) . H 0 : 0 , H1 : 1 . Example 2: X has one of the following two distributions: P(X=x) 0 1 2 3 4 0.1 0.1 0.1 0.2 0.5 P0 P1 0.3 0.3 0.2 0.1 0.1 Bayes procedures: Consider 0-1 loss, i.e., the loss is 1 if we choose the incorrect hypothesis and 0 if we choose the correct hypothesis. Let the prior probabilities be on 0 and 1 on 1 . The posterior probability for 0 is P( X | 0 ) ( 0 ) P( 0 | X ) P( X | 0 ) ( 0 ) P( X | 1 ) (1 ) . The posterior risk for 0-1 loss is minimized by choosing the hypothesis with higher posterior probability. 6 Thus, the Bayes rule is to choose H 0 (equivalently 0 ) if P( X | 0 ) (0 ) P( X | 1 ) (1 ) and choose H1 (equivalently 1 ) otherwise. 1 For ( 0 ) (1 ) 2 , the Bayes rule is choose H 0 if P( X | 0 ) P( X | 1 ) and choose H1 otherwise. 1 Note that the Bayes risk for the prior ( 0 ) (1 ) 2 of a test is 0.5*P(Type I error)+0.5*P(Type II error). Thus, the 1 ( ) ( ) 0 1 Bayes procedure for the prior 2 minimizes the sum of the probability of a type I error and the probability of a type II error. Example 1 continued: Suppose ( X1 , , X 5 ) (1.1064, 1.1568, -0.1602, 1.0343, -0.1079), X 0.6059 . Then 5 5 ( X i 0) 2 1 i 1 0.001613 P( X | 0) exp 2 2 5 5 ( X i 1) 2 1 i 1 0.002739 P( X | 1) exp 2 2 We choose H 0 : 0 if P( X | 0) ( 0) P( X | 1) ( 1) , or equivalently P( X | 0) ( 0) 1, P( X | 1) ( 1) 7 0.001613* ( 0) which for this data is 0.002739* ( 1) 1. Writing ( 1) (1 ( 0)) , we have that we choose H 0 : 0 for priors with ( 0) 0.6294 and H1 : 1 for priors with ( 0) 0.6294 . III. Neyman-Pearson Lemma (Section 4.2) In the Neyman-Pearson paradigm, the hypotheses are not treated symmetrically: Fix 0 1 . Among tests having level (Type I error probability) , find the one that has the “best” power function. For simple vs. simple hypothesis, the best power function means the best power at H1 : 1 . Such a test is called the most powerful level test. The Neyman-Pearson lemma provides us with a most powerful level test for simple vs. simple hypotheses. Analogy: To fill up a bookshelf with books with the least cost, we should start by picking the one with the largest width/$ and continue. Similarly, to find a most powerful level test, we should start by including in the critical region those sample points that are most likely under the alternative relative to the null hypothesis and continue. Define the likelihood ratio statistic by 8 L( X , 0 , 1 ) p ( X | 1 ) p( X | 0 ) , where p ( x | ) is the probability mass function or probability density function of the data X. The statistic L takes on the value when p( X | 1 ) 0, p( X | 1 ) 0 and by convention equals 0 when both p( X | 1 ) 0, p( X | 1 ) 0 . We can describe a test by a test function ( x ) . When ( x ) 1 , we always reject H 0 . When ( x ) c , 0 c 1, we conduct a Bernoulli trial and reject with probability c (thus we allow for randomized tests) When ( x ) 0 , we always accept H 0 . We call k a likelihood ratio test if 1 if L( x,0 ,1 ) k k ( x ) c if L( x,0 ,1 ) k 0 if L( x, , ) k 0 1 Theorem 4.2.1 (Neyman-Pearson Lemma): Consider testing H 0 : 0 ( P( X | 0 ) P0 ) vs. H1 : 1 ( P( X | 1 ) P1 ) (a) If 0 and k is a size likelihood ratio test, then k is a most powerful level test 9 (b) For each 0 1 , there exists a most powerful size likelihood ratio test. (c) If is a most powerful level test, then it must be a level likelihood ratio test except perhaps on a set A satisfying P0 ( X A) P1 ( X A) 0 Example 1: X 1 , , X n iid N ( ,1) . H 0 : 0 , H1 : 1 . The likelihood ratio statistic is n 1 1 2 exp ( xi 1) 2 2 L( X , 0 , 1 ) i 1 n 1 1 2 exp xi 2 2 i 1 1 exp x i 2 i 1 n exp n / 2 i 1 xi n Rejecting the null hypothesis for large values of L( X ,0 ,1 ) is equivalent to rejecting the null hypothesis for large values of n i 1 Xi . What should the cutoff be? The distribution of n i1 X i under the null hypothesis is N (0, n) so the most powerful level tests rejects for n i 1 Xi n normal. (1 ) where is the CDF of a standard 10 Example 2: P0 0 0.1 1 0.1 P(X=x) 2 3 0.1 0.2 P1 0.3 0.3 0.2 0.1 0.1 3 2 0.5 0.2 L( x,0 ,1 ) 3 4 0.5 The most powerful level 0.2 test rejects if and only if X=0 or 1. There are multiple most powerful level 0.1 tests, e.g., 1) reject the null hypothesis if and only if X=0; 2) reject the null hypothesis if and only if X=0; 3) flip a coin to decide whether to reject the null hypothesis when X=0 or X=1. Proof of Neyman-Pearson Lemma: (a) We prove the result here for continuous random variables X. The proof for discrete random variables follows by replacing integrals with sums. * Let be the test function of any other level test besides k . Because * is level , EP0 * . We want to show * that k ( x )dP1 ( x ) ( x )dP1 ( x ) . * ( ( x ) ( x ))( p1 ( x ) kp0 ( x )) dx and show We examine k * that (k ( x ) ( x ))( p1 ( x ) kp0 ( x )) dx 0 . From this, we conclude that * * ( ( x ) ( x )) p ( x ) ( ( x ) ( x ))kp0 ( x )) dx k 1 k 11 The latter integral is 0 because * ( x ) p ( x ) d x , k 0 ( x) p0 ( x)dx * Hence, we conclude that (k ( x ) ( x )) p1 ( x ) 0 as desired. * To show that (k ( x ) ( x ))( p1 ( x ) kp0 ( x )) dx 0 , let S { x : k ( x ) * ( x )} S { x : k ( x ) * ( x )} S 0 { x : k ( x ) * ( x )} Suppose x S . This implies k ( x ) 0 which implies that p1 ( x ) kp0 ( x ) . Thus, S (k ( x) * ( x))( p1 ( x) kp0 ( x))dx 0 . Also, similarly, * ( ( x ) ( x))( p1 ( x) kp0 ( x))dx 0 and k S S 0 (k ( x) * ( x))( p1 ( x) kp0 ( x))dx = 0 0 (since p1 ( x) kp0 ( x) for x S ). Thus, * ( ( x ) ( x ))( p1 ( x ) kp0 ( x )) dx 0 and this shows that k * ( ( x ) ( x )) p1 ( x ) 0 as argued above. k 12 (b) Let (c) P0 [ p1 ( x) cp0 ( x)] 1 F0 (c) where F0 is the p1 ( x ) cdf of p ( x ) under P0 . By the properties of CDFs, (c ) is 0 nonincreasing in c and right continuous. By the right continuity of (c ) , there exists c0 such that (c0 ) (c0 ) . So define p1 ( x ) 1 if c0 p ( x ) 0 (c0 ) p1 ( x ) ( x) if c0 ( c ) ( c ) p ( x ) 0 0 0 p ( x) if 1 c0 0 p0 ( x ) Then, (c0 ) p ( x) p1 ( x ) EP0 ( x ) P0 ( 1 c0 ) P ( c0 ) 0 p0 ( x ) (c0 ) (c0 ) p0 ( x ) (c0 ) (c0 ) [ (c0 ) (c0 )] (c0 ) (c0 ) So we can take k to be c0 . * (c) Let be the test function for any most powerful level test. By parts (a) and (b), a likelihood ratio test k with * size can be found that is most powerful. Since and k are both most powerful, it follows that 13 ( ( x) k * ( x )) p1 ( x ) 0 (1.1) Following the proof in part (a), (1.1) implies that ( ( x) k * ( x ))( p1 ( x ) kp0 ( x )) dx 0 * which can be the case if and only if ( x ) ( x ) 0 when p1 ( x ) kp0 ( x ) (i.e., L( X ,0 ,1 ) k ) and * ( x ) ( x ) 1 when p1 ( x) kp0 ( x) (i.e., L( X ,0 ,1 ) k ) except on a set A satisfying P0 ( X A) P1 ( X A) 0 . 14