Statistics 550 Notes 9 Reading: Section 1.6.4-1.6.5, 2.1 I. Multiparameter exponential families The family of distributions of a model {P : } is said to be a k-parameter exponential family if there exist realvalued functions 1 , ,k , B of such that the pdf or pmf may be written as k p( x | ) h( x) exp{ j ( )T j ( x) B( )} j 1 By the factorization theorem, T ( X ) (T1 ( X ), a sufficient statistic. (1.1) , Tk ( X )) is 2 Example 1: X 1 , , X n is iid N ( , ) is a two-parameter exponential family. Example 2: Multinomial. Suppose we observe n independent trials where each trial can end up in one of k possible categories {1,...,k} with probabilities { p1 , , pk 1 , pk 1 p1 pk 1} . Let y1 ( x ), , yk ( x) be the number of outcomes in categories 1,...,k in the n trials. Then, 1 p( x | ) n! y1 ( x ) yk ( x ) n! y1 ( x ) p1 yk ( x ) pk n! y1 ( x ) yk ( x ) pk yk ( x ) p1 y1 ( x ) y1 ( x ) pk 1 pk yk 1 ( x ) pk n exp[ y1 ( x ) log( p1 / pk ) y k 1 ( x ) log( pk 1 / pk ) n log pk ] n! y1 ( x ) yk ( x ) exp[ y1 ( x ) log( p1 / pk ) k 1 y k 1 ( x) log( pk 1 / pk ) n log(1 exp(log i 1 The multinomial is a (k-1) parameter exponential family with (log( p1 / pk , ,log( pk 1 / pk )) , k 1 T ( x) y1 ( x), , yk 1 ( x) and A( ) n log(1 exp(i )) . i 1 Moments of Sufficient Statistics: As with the oneparameter exponential family, it is convenient to index the family by (1 , ,k ) . The analogue of Theorem 1.6.2 that calculates the moments of the sufficient statistics is Corollary 1.6.1: A E0 T ( X ) (0 ), 1 T A , (0 ) k 2 A Var0 T ( X ) (0 ) a b Example 2 continued: For the multinomial distribution, 2 pi ))] pk n pi pk nei k 1 i E[ y j ( x )] n log 1 e k 1 k 1 j i 1 1 ei 1 i 1 Cov0 [ yi ( x ), y j ( x )] i 1 pi pk npi 1 pi pk pk n nei e j k 1 n log 1 ei npi p j , i j k 1 j k 2 i 1 (1 e i ) i 1 Var0 [ yi ( x )] n log 1 ei npi (1 pi ) . 2 j i 1 2 k 1 II. Conjugate Families of Prior Distributions (Chapter 1.6.5) Consider a Bayesian model for the data in which the distribution of the data given the parameter is p ( x | ) where . A family of prior distributions for , { ( | ), } , is a conjugate family of priors to this model if the posterior distribution for , p ( | x ) , also belongs to { ( | ), } . Note the parameters of a prior distribution are often called hyperparameters. Examples: In Notes 2, we showed that the beta family of priors is conjugate to the binomial model for the data and in Notes 3, we showed that the normal family of priors is conjugate to the normal model for the data in which the variance is known. Suppose X 1 , , X n are iid from the k-parameter exponential family (1.1) so that 3 n k n p( x | ) h( xi ) exp j ( ) T j ( xi ) nB( )} (1.2) i 1 i 1 j 1 A conjugate exponential family of priors is obtained by n letting t j T j ( xi ), j 1, i 1 , k and n tk 1 be “parameters” and treating as the variable of interest. That is, let t (t1 , , tk 1 )T and (t ) k exp{ t j j ( ) tk 1 B( )}d1 j 1 d k , {(t1 , , tk 1 ) : 0 (t1 , , tk 1 ) } with integrals replaced by sums in the discrete case. We assume is nonempty. Then, Proposition 1.6.1: The (k 1) parameter exponential family given by k ( | t ) exp j ( )t j tk 1B( ) log (t ) , t (t1, j 1 , tk 1 ) (1.3) is a conjugate family of prior distributions to p ( x | ) given by (1.2). Note: We can view the prior distribution as saying that we * have tk 1 additional data points x from p ( x | ) with * sufficient statistics T1 t1 , , Tk tk , since the data ( x , x ) has a pdf/pmf that is proportional to the joint distribution of ( , x ) in the Bayesian model. 4 Proof: If p ( x | ) is given by (1.2) and the prior distribution for is given by a member of the family (1.3), then p( | x ) p( x | ) ( | t ) exp j 1 j ( ) , where s (s1 , k n i 1 T j ( xi ) t j (tk 1 n) B( ) ( | s) , sk 1 )T t1 i 1T1 ( xi ), n , tk i 1Tk ( xi ), tk 1 n n Because two probability densities that are proportional must be equal, p( | x ) ( | s) . Note that the prior parameter t is simply updated to s t a , where a = (i 1T1 ( X i ), n ,i 1Tk ( X i ), n) . n Many well known examples of conjugate families are special cases of Proposition 1.6.1. Example: Suppose X 1 , , X n are iid Bernoulli( ). The X i follow a one-parameter exponential family p( x | ) x (1 )1 x exp[ x log log(1 )] , 1 ( ) log , B( ) log(1 ), T ( x) x . with 1 From Proposition 1.6.1, a conjugate family of prior distributions is 5 ( | t1 , t2 ) exp ( )t1 t2 B( ) exp log 1 t t log(1 ) 1 2 exp t1 log (t2 t1 ) log 1 t (1 )t t 1 2 1 This is proportional to a Beta( t1 1, t2 t1 1 ) distribution. Thus, the beta family of prior distributions is a conjugate family to the Bernoulli (or binomial) model. For r and s integers, we can view a Beta(r,s) prior as saying that we have r s 2 additional data points of which r 1 are successes. In Example 2 of Notes 2, we put a Beta(53,47) prior on Shaquille O’Neal’s probability of making a free throw shot. This is saying that our prior information is equivalent to observing Shaq take 98 free throw shots and make 52 of the shots. III. Methods of Estimation: Basic Heuristics of Estimation Basic Setup: Family of possible distributions for data X { p( x | ), } . Observe data X. Point estimation: Best estimate of based on data X. We discussed the decision theoretic approach to evaluating point estimates focusing particularly on squared error as a loss function which results in mean squared error as the risk function. But how do we come up with possible estimates of . 6 Example Estimation Problems: (1) Bernoulli data: We observe X 1 , , X n iid Bernoulli ( ) (e.g., Shaq’s free throws). How do we estimate ? (2) Regression. We are interested in the mean of a response Y given covariates X 1 , , X p and assume a model E (Y | X 1 , , X p ) g ( X 1 , , X p ) , where g is a known function and is the unknown parameter vector. Example: Life insurance companies are keenly interested in predicting how long their customers will live because their premiums and profitability depend on such numbers. An actuary for one insurance company gathered data from 100 recently deceased male customers. She recorded Y=the age at death of the customer, X1 = the age at death of his mother, X2 =the age at death of his father, X3=the mean age at death of his grandmothers and X4 =the mean age at death of his grandfathers. Multiple linear regression model: E (Y | X1 , X 2 , X 3 , X 4 ) 0 1 X1 2 X 2 3 X 3 4 X 4 (3) Parameter estimation in an iid model. As part of a study to estimate the population size of the bowhead whale, Raftery and Zeh wanted to understand the distribution of whale swimming speeds. They randomly sampled the time to swim 1km of 210 whales and believe that the gamma 7 model is a reasonable model for this data. p x p 1e x p( x | p, ) ( p ) . How do we estimate p and ? (4) Hardy-Weinberg equilibrium. If gene frequencies are in equilibrium, then for a gene with two alleles, the genotypes AA, Aa and aa occur in a population with 2 2 frequencies (1 ) , 2 (1 ), respectively according to the Hardy-Weinberg law. In a sample from the Chinese population of Hong-Kong in 1937, blood types occurred with the following frequencies, where M and N are erythrocyte antigens: Blood Type M MN N Total Frequency 342 500 187 1029 We can model the observed blood types as an iid sample from a multinomial distribution with probabilities (1 )2 , 2 (1 ), 2 . How do we estimate ? Minimum contrast heuristic: Choose a contrast function ( X , ) that measures the “discrepancy” between the data X and the parameter vector . The range of the contrast function is typically taken to be the real numbers greater than or equal to zero and the smaller the value of the contrast function, the more “plausible” is based on the data X. 8 Let 0 denote the true parameter. Define the population discrepancy D( 0 , ) as the expected value of the discrepancy ( X , ) : D( 0 , ) E0 ( X , ) (1.4) In order for ( X , ) to be a valid contrast function, we require that D( 0 , ) is uniquely minimized for 0 , i.e., D( 0 , ) D( 0 , 0 ) if 0 . 0 is the minimizer of D(0 , ) . Although we don’t know D( 0 , ) , the contrast function ( X , ) is an unbiased estimate of D( 0 , ) (see (1.4)). The minimum contrast heuristic is to estimate by minimizing ( X , ) , i.e., ˆ min ( X , ) . Example 1: Suppose X 1 , , X n iid Bernoulli (p), 0 p 1 . The following is an example of a contrast functions and an associated estimate: n 2 ( X , p ) ( X p ) i “Least Squares”: . i 1 D( p0 , p) E p0 [ i 1 ( X i p) 2 ] n np0 2npp0 np 2 We have 9 D( p0 , p) 2np0 2np p and it can be verified by the second derivative test that arg min p D( p0 , p) p0 n 2 ( X , p ) ( X p ) i Thus, is a valid contrast function. i 1 The associated estimate is n pˆ arg min p ( X , p) arg min p i 1 ( X i p) 2 arg min p p 2 p i 1 X i n 2 n i 1 Xi n The following is an example of a function that is not a contrast function: n ( X , p) ( X i p) 4 i 1 D( p0 , p) E p0 [ i 1 ( X i p) 4 ] n E p0 [ i 1 X i4 4 X i3 p 6 X i2 p 2 4 X i p 3 p 4 ] n p0 4 p0 p 6 p0 p 2 4 p0 p 3 p 4 For p0 0.7 , we find that D( p0 , p) is maximized at about p=0.57 10 Least Squares methods for estimating regression can be viewed as a minimum contrast estimates (Example 2.1.1). Estimating Equation Heuristic: Suppose is d-dimensional. Consider a d-dimensional function ( X , ) and define V ( 0 ) E0 ( X , ) . 11 Suppose V (0 ) 0 has 0 as its unique solution for 0 . We do not know V (0 but ( X , ) is an unbiased estimate of V (0 . The estimating equation heuristic is to estimate by solving ( X , ) 0 , i.e., ( X ,ˆ) 0 . ( X , ) is called an estimating equation. Method of Moments: Suppose X 1 , , X n iid from { p( x | ), } where is d-dimensional. Let 1 ( ), , d ( ) denote the first d-moments of the population we are sampling from (assuming that they exist), j ( ) E ( X j ), 1 j d Define the jth sample moment ˆ j by 1 n ˆ j i 1 X i j , 1 j d . n The function ( X , ) (ˆ1 1 ( ), , ˆ d d ( )) is an estimating equation for which V (θ0 0 E ( X , ) ( E ˆ1 1 ( ), , E ˆ d d ( )) 0 For many models, V (θ0 0 for all 0 . Suppose ( 1 ( ), , d ( )) is a 1-1 continuous d d function from to . Then the estimating equation 12 estimate of based on ( X , ) is the ˆ that solves ( X ,ˆ) 0 , i.e., ˆ (ˆ) 0, j 1, , d . j j Example 2: X 1 , 1 ( ) , X n iid Uniform (0, ) . 2 The method of moments estimator solves, ˆ X 0, 2 i.e., ˆ 2X . Example 3: X 1 , 1 ( ) , X n iid N ( , 2 ) 2 ( ) 2 2 The method of moments estimator solves, X 0 n 2 X i i 1 n 2 2 0 i 1 X i2 n 2 Thus, ˆ X and ˆ n X2 n 2 ( X X ) i i 1 n Large sample motivation for method of moments: 13 . A reasonable requirement for a point estimator is that it should converge to the true parameter value as we collect more and more information. Suppose X 1 , , X n iid. A point estimator h(X1,...,Xn) of a parameter q( ) is P consistent if h(X1,...,Xn) q( ) as n for all . Definition of convergence in probability (A.14.1, page P 466). h(X1,...,Xn) q( ) means that for all 0 , lim P[| h( X 1 ,..., X n ) q( ) | ] 0 . n Under certain regularity conditions, the method of moments estimator is consistent. We give a proof for a special case Let g ( ) ( 1 ( ), , d ( )) . By the assumptions in formulating the method of moments, g is a 1-1 continuous d d function from to . The method of moments estimator solves g (ˆ) (ˆ1 , , ˆ d ) 0 . d When the g’s range is , then ˆ g 1 (ˆ1 , , ˆ d ) . We prove the method of moments 1 1 estimator is consistent when ˆ g (ˆ , , ˆ ) and g is 1 d continuous. Sketch of Proof: The method of moments estimator solves 14 ˆ j j (ˆ) 0, j 1, , d . By the law of large numbers, P ( ˆ1 , , ˆ d ) ( 1 ( ), , d ( )) . By the open mapping theorem (A.14.8, page 467), since g 1 is assumed to be continuous, ˆ g 1 ( ˆ1 , P , ˆ d ) g 1 ( 1 ( ), , d ( )) Comments on method of moments: (1) Instead of using the first d moments, we could use higher order moments instead, leading to different estimating equations. But the method of moments estimator may be altered by which moments we choose. Example: X 1 , , X n iid Poisson( ). The first moment is 1 ( ) E ( X ) . Thus, the method of moments estimator based on the first moment is ˆ X . We could also consider using the second moment to form a method of moments estimator. 2 ( ) E ( X 2 ) 2 . The method of moments estimator based on the second moment solves 1 n 2 ˆ ˆ2 Xi n i 1 Solving this equation (by taking the positive root), we find that 15 1/ 2 1 1 1 n ˆ i 1 X i2 . 2 4 n The two method of moments estimators are different. For example, for the data > rpois(10,1) [1] 2 3 0 1 2 1 3 1 2 1, the method of moments estimator based on the first moment is 1.1 and the method of moments estimator based on the second moment is 1.096872. (2) The method of moments does not use all the information that is available. X 1 , , X n iid Uniform (0, ) . The method of moments estimator based on the first moment is ˆ 2X . If 2 X max X i , we know that max X ˆ i 16