evaluation of point estimators lecture 11 biostatistics
play

Evaluation of Point Estimators Lecture 11 Biostatistics 602 - - PowerPoint PPT Presentation

. . February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang February 14th, 2013 Hyun Min Kang Evaluation of Point Estimators Lecture 11 Biostatistics 602 - Statistical Inference . . Summary Cramer-Rao . Evaluation MLE Recap


  1. • The value of . where We define the induced likelihood function L by L x sup L x . . that maximize L x is called the MLE of . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . Definition . . . . . . . . . . Recap MLE 8 / 33 Evaluation Cramer-Rao . Summary Induced Likelihood Function . . . . . . . . . . . . . . . . . . . . . . . . . . • Let L ( θ | x ) be the likelihood function for a given data x 1 , · · · , x n , • and let η = τ ( θ ) be a (possibly not a one-to-one) function of θ .

  2. • The value of . Induced Likelihood Function February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang . x is called the MLE of that maximize L sup . . Definition . . Summary . . . . . . . . . . Recap MLE Cramer-Rao Evaluation 8 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . • Let L ( θ | x ) be the likelihood function for a given data x 1 , · · · , x n , • and let η = τ ( θ ) be a (possibly not a one-to-one) function of θ . We define the induced likelihood function L ∗ by L ∗ ( η | x ) = L ( θ | x ) θ ∈ τ − 1 ( η ) where τ − 1 ( η ) = { θ : τ ( θ ) = η, θ ∈ Ω } .

  3. . . February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang sup . . Definition . Induced Likelihood Function Summary . Cramer-Rao . . . . . . . . . 8 / 33 Recap MLE Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . • Let L ( θ | x ) be the likelihood function for a given data x 1 , · · · , x n , • and let η = τ ( θ ) be a (possibly not a one-to-one) function of θ . We define the induced likelihood function L ∗ by L ∗ ( η | x ) = L ( θ | x ) θ ∈ τ − 1 ( η ) where τ − 1 ( η ) = { θ : τ ( θ ) = η, θ ∈ Ω } . • The value of η that maximize L ∗ ( η | x ) is called the MLE of η = τ ( θ ) .

  4. . x L x sup L x sup sup L x sup L x L x L sup . L x L x Hence, L x L x and is the MLE of . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . . . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary . Invariance Property of MLE Theorem 7.2.10 . . . . . Proof - Using Induced Likelihood Function 9 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . . . If θ is the MLE of ˆ θ , then the MLE of η = τ ( θ ) is τ (ˆ θ ) , where τ ( θ ) is any function of θ .

  5. . L sup sup L x sup L x L x L x sup x . L x Hence, L x L x and is the MLE of . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 sup 9 / 33 . . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Invariance Property of MLE . . . . Theorem 7.2.10 Proof - Using Induced Likelihood Function . . . . . . . . . . . . . . . . . . . . . . . . . . If θ is the MLE of ˆ θ , then the MLE of η = τ ( θ ) is τ (ˆ θ ) , where τ ( θ ) is any function of θ . L ∗ (ˆ η L ∗ ( η | x ) η | x ) =

  6. . L . sup sup sup L x L x L x sup x . L x Hence, L x L x and is the MLE of . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . 9 / 33 Proof - Using Induced Likelihood Function . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary . Invariance Property of MLE Theorem 7.2.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . If θ is the MLE of ˆ θ , then the MLE of η = τ ( θ ) is τ (ˆ θ ) , where τ ( θ ) is any function of θ . L ∗ (ˆ η L ∗ ( η | x ) = sup η | x ) = L ( θ | x ) η θ ∈ τ − 1 ( η )

  7. . x . sup sup sup L x L x sup L L . x Hence, L x L x and is the MLE of . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . 9 / 33 Proof - Using Induced Likelihood Function . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Invariance Property of MLE . Theorem 7.2.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . If θ is the MLE of ˆ θ , then the MLE of η = τ ( θ ) is τ (ˆ θ ) , where τ ( θ ) is any function of θ . L ∗ (ˆ η L ∗ ( η | x ) = sup η | x ) = L ( θ | x ) η θ ∈ τ − 1 ( η ) = L ( θ | x ) θ

  8. . x . . sup sup sup L x sup L L Proof - Using Induced Likelihood Function x Hence, L x L x and is the MLE of . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . 9 / 33 . . Evaluation Cramer-Rao . Summary Invariance Property of MLE . Recap Theorem 7.2.10 MLE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . If θ is the MLE of ˆ θ , then the MLE of η = τ ( θ ) is τ (ˆ θ ) , where τ ( θ ) is any function of θ . L ∗ (ˆ η L ∗ ( η | x ) = sup η | x ) = L ( θ | x ) η θ ∈ τ − 1 ( η ) L ( θ | x ) = L (ˆ = θ | x ) θ

  9. . x . . . sup sup sup sup L Hence, L . x L x and is the MLE of . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 Proof - Using Induced Likelihood Function 9 / 33 Evaluation . Cramer-Rao MLE Recap . . . . . . . . . Summary Invariance Property of MLE . Theorem 7.2.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . If θ is the MLE of ˆ θ , then the MLE of η = τ ( θ ) is τ (ˆ θ ) , where τ ( θ ) is any function of θ . L ∗ (ˆ η L ∗ ( η | x ) = sup η | x ) = L ( θ | x ) η θ ∈ τ − 1 ( η ) L ( θ | x ) = L (ˆ = θ | x ) θ L (ˆ θ | x ) = L ( θ | x ) θ ∈ τ − 1 ( τ (ˆ θ ))

  10. . Hence, L . . . sup sup sup sup x . L x and is the MLE of . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 Proof - Using Induced Likelihood Function 9 / 33 . . MLE Recap . . . . . . . Cramer-Rao . Evaluation . Summary Invariance Property of MLE . Theorem 7.2.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . If θ is the MLE of ˆ θ , then the MLE of η = τ ( θ ) is τ (ˆ θ ) , where τ ( θ ) is any function of θ . L ∗ (ˆ η L ∗ ( η | x ) = sup η | x ) = L ( θ | x ) η θ ∈ τ − 1 ( η ) L ( θ | x ) = L (ˆ = θ | x ) θ L (ˆ L ( θ | x ) = L ∗ [ τ (ˆ θ | x ) = θ ) | x ] θ ∈ τ − 1 ( τ (ˆ θ ))

  11. . Theorem 7.2.10 February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang sup sup sup sup . . . Proof - Using Induced Likelihood Function . . . 9 / 33 . Invariance Property of MLE Recap . . . . MLE Evaluation . Cramer-Rao Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . If θ is the MLE of ˆ θ , then the MLE of η = τ ( θ ) is τ (ˆ θ ) , where τ ( θ ) is any function of θ . L ∗ (ˆ η L ∗ ( η | x ) = sup η | x ) = L ( θ | x ) η θ ∈ τ − 1 ( η ) L ( θ | x ) = L (ˆ = θ | x ) θ L (ˆ L ( θ | x ) = L ∗ [ τ (ˆ θ | x ) = θ ) | x ] θ ∈ τ − 1 ( τ (ˆ θ )) η | x ) = L ∗ [ τ (ˆ θ ) | x ] and τ (ˆ Hence, L ∗ (ˆ θ ) is the MLE of τ ( θ ) .

  12. . . February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang robust). 4 Heavily depends on the underlying distributional assumptions (i.e. not . 3 Not always easy to obtain; may be hard to find the global maximum. . space. 2 By definition, MLE will always fall into the range of the parameter . . 1 Optimal in some sense : We will study this later . Properties of MLE . Summary . Cramer-Rao Evaluation MLE Recap . . . . . . . . . 10 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . .

  13. . . February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang robust). 4 Heavily depends on the underlying distributional assumptions (i.e. not . 3 Not always easy to obtain; may be hard to find the global maximum. . . space. 2 By definition, MLE will always fall into the range of the parameter . . 1 Optimal in some sense : We will study this later . . Recap . . . . . . . . . MLE Properties of MLE Evaluation Cramer-Rao . Summary 10 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . .

  14. . 3 Not always easy to obtain; may be hard to find the global maximum. . . 2 By definition, MLE will always fall into the range of the parameter space. . . . . . 4 Heavily depends on the underlying distributional assumptions (i.e. not robust). Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 1 Optimal in some sense : We will study this later . . Recap . . . . . . . . . MLE Properties of MLE Evaluation Cramer-Rao . Summary 10 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . .

  15. . 3 Not always easy to obtain; may be hard to find the global maximum. . . 2 By definition, MLE will always fall into the range of the parameter space. . . . . . 4 Heavily depends on the underlying distributional assumptions (i.e. not robust). Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 1 Optimal in some sense : We will study this later . . Recap . . . . . . . . . MLE Properties of MLE Evaluation Cramer-Rao . Summary 10 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . .

  16. X n are iid samples from a distribution with mean X i is an estimator of E X Therefore X is an unbiased estimator for . . . . X . Let X n n i . The bias is Bias E . n n i X i n n i E X i . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . . . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary . Method of Evaluating Estimators Definition : Unbiasedness If the bias is equal to 0, then . Example . . is an unbiased estimator for 11 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . . . Suppose ˆ θ is an estimator for θ , then the bias of θ is defined as Bias ( θ ) = E (ˆ θ ) − θ

  17. X n are iid samples from a distribution with mean X i is an estimator of E X Therefore X is an unbiased estimator for . . . . X . Let X n n i . The bias is Bias E . n n i X i n n i E X i . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . . . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Method of Evaluating Estimators . Definition : Unbiasedness . Example . 11 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . . . Suppose ˆ θ is an estimator for θ , then the bias of θ is defined as Bias ( θ ) = E (ˆ θ ) − θ If the bias is equal to 0, then ˆ θ is an unbiased estimator for θ .

  18. E X Therefore X is an unbiased estimator for . n . . n The bias is Bias E n X i i . n n i E X i . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 Example . 11 / 33 Method of Evaluating Estimators . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary . . . Definition : Unbiasedness . . . . . . . . . . . . . . . . . . . . . . . . . . Suppose ˆ θ is an estimator for θ , then the bias of θ is defined as Bias ( θ ) = E (ˆ θ ) − θ If the bias is equal to 0, then ˆ θ is an unbiased estimator for θ . X 1 , · · · , X n are iid samples from a distribution with mean µ . Let X = 1 ∑ n i =1 X i is an estimator of µ .

  19. Therefore X is an unbiased estimator for . X i . . n E n n i n . n i E X i . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 Example . 11 / 33 . Summary Evaluation MLE Method of Evaluating Estimators Recap Cramer-Rao . . . . . Definition : Unbiasedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Suppose ˆ θ is an estimator for θ , then the bias of θ is defined as Bias ( θ ) = E (ˆ θ ) − θ If the bias is equal to 0, then ˆ θ is an unbiased estimator for θ . X 1 , · · · , X n are iid samples from a distribution with mean µ . Let X = 1 ∑ n i =1 X i is an estimator of µ . The bias is E ( X ) − µ Bias ( µ ) =

  20. Therefore X is an unbiased estimator for . X i . . n E n n n n . i E X i . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 Example . . Summary . . . . . . . . . Recap MLE Evaluation Cramer-Rao . 11 / 33 . . Definition : Unbiasedness Method of Evaluating Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . Suppose ˆ θ is an estimator for θ , then the bias of θ is defined as Bias ( θ ) = E (ˆ θ ) − θ If the bias is equal to 0, then ˆ θ is an unbiased estimator for θ . X 1 , · · · , X n are iid samples from a distribution with mean µ . Let X = 1 ∑ n i =1 X i is an estimator of µ . The bias is E ( X ) − µ Bias ( µ ) = ( ) 1 ∑ = − µ i =1

  21. Therefore X is an unbiased estimator for . X i . . n E n n n n . i E X i . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 Example . . Summary . . . . . . . . . Recap MLE Evaluation Cramer-Rao . 11 / 33 . . Definition : Unbiasedness Method of Evaluating Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . Suppose ˆ θ is an estimator for θ , then the bias of θ is defined as Bias ( θ ) = E (ˆ θ ) − θ If the bias is equal to 0, then ˆ θ is an unbiased estimator for θ . X 1 , · · · , X n are iid samples from a distribution with mean µ . Let X = 1 ∑ n i =1 X i is an estimator of µ . The bias is E ( X ) − µ Bias ( µ ) = ( ) 1 ∑ = − µ i =1

  22. Therefore X is an unbiased estimator for . n Example . . n E n X i . n n . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . . Cramer-Rao . . . . . . . . . Recap MLE Evaluation Definition : Unbiasedness 11 / 33 . Summary . Method of Evaluating Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . Suppose ˆ θ is an estimator for θ , then the bias of θ is defined as Bias ( θ ) = E (ˆ θ ) − θ If the bias is equal to 0, then ˆ θ is an unbiased estimator for θ . X 1 , · · · , X n are iid samples from a distribution with mean µ . Let X = 1 ∑ n i =1 X i is an estimator of µ . The bias is E ( X ) − µ Bias ( µ ) = ( ) 1 − µ = 1 ∑ ∑ = E ( X i ) − µ i =1 i =1

  23. Therefore X is an unbiased estimator for . n Example . . n E n X i . n n . Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . . Cramer-Rao . . . . . . . . . Recap MLE Evaluation Definition : Unbiasedness 11 / 33 . Summary . Method of Evaluating Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . Suppose ˆ θ is an estimator for θ , then the bias of θ is defined as Bias ( θ ) = E (ˆ θ ) − θ If the bias is equal to 0, then ˆ θ is an unbiased estimator for θ . X 1 , · · · , X n are iid samples from a distribution with mean µ . Let X = 1 ∑ n i =1 X i is an estimator of µ . The bias is E ( X ) − µ Bias ( µ ) = ( ) 1 − µ = 1 ∑ ∑ = E ( X i ) − µ = µ − µ = 0 i =1 i =1

  24. . n . Example . . n E n . X i n n Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . Definition : Unbiasedness Evaluation . . . . . . . . . Recap MLE . 11 / 33 Cramer-Rao . Method of Evaluating Estimators Summary . . . . . . . . . . . . . . . . . . . . . . . . . . Suppose ˆ θ is an estimator for θ , then the bias of θ is defined as Bias ( θ ) = E (ˆ θ ) − θ If the bias is equal to 0, then ˆ θ is an unbiased estimator for θ . X 1 , · · · , X n are iid samples from a distribution with mean µ . Let X = 1 ∑ n i =1 X i is an estimator of µ . The bias is E ( X ) − µ Bias ( µ ) = ( ) 1 − µ = 1 ∑ ∑ = E ( X i ) − µ = µ − µ = 0 i =1 i =1 Therefore X is an unbiased estimator for µ .

  25. . Summary February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang . than (red) is biased but more likely to be closer to the true • . (blue) is unbiased but has a chance to be very far away from • How important is unbiased? . . Cramer-Rao Evaluation MLE Recap . . . . . . . . . 12 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . .

  26. . Evaluation February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang How important is unbiased? Summary . . Cramer-Rao MLE Recap . . . . . . . . . 12 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . • ˆ θ 1 (blue) is unbiased but has a chance to be very far away from θ = 0 . • ˆ θ 2 (red) is biased but more likely to be closer to the true θ than ˆ θ 1 .

  27. . E . . . . MSE E E E E E E E E . E E E E E E E E E Var Bias Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . . . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Mean Squared Error . Definition . . . Property of MSE 13 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . Mean Squared Error (MSE) of an estimator ˆ θ is defined as MSE (ˆ θ ) = E [(ˆ θ − θ )] 2

  28. . E . . E E E E E E E E E . E E E E E Var Bias Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . Property of MSE 13 / 33 . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Mean Squared Error Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mean Squared Error (MSE) of an estimator ˆ θ is defined as MSE (ˆ θ ) = E [(ˆ θ − θ )] 2 MSE (ˆ E [(ˆ θ − E ˆ θ + E ˆ θ − θ )] 2 θ ) =

  29. . E . Property of MSE . . E E E . E E E Var Bias Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . Definition . . . . . . . . . . Recap MLE Evaluation 13 / 33 Summary Cramer-Rao . Mean Squared Error . . . . . . . . . . . . . . . . . . . . . . . . . . Mean Squared Error (MSE) of an estimator ˆ θ is defined as MSE (ˆ θ ) = E [(ˆ θ − θ )] 2 MSE (ˆ E [(ˆ θ − E ˆ θ + E ˆ θ − θ )] 2 θ ) = E [(ˆ θ − E ˆ θ ) 2 ] + E [( E ˆ θ − θ ) 2 ] + 2 E [(ˆ θ − E ˆ θ )] E [( E ˆ = θ − θ )]

  30. . Mean Squared Error February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang Bias Var . . Property of MSE . . . Definition . . Summary MLE . . . . . . . . . . Recap 13 / 33 Cramer-Rao Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . Mean Squared Error (MSE) of an estimator ˆ θ is defined as MSE (ˆ θ ) = E [(ˆ θ − θ )] 2 MSE (ˆ E [(ˆ θ − E ˆ θ + E ˆ θ − θ )] 2 θ ) = E [(ˆ θ − E ˆ θ ) 2 ] + E [( E ˆ θ − θ ) 2 ] + 2 E [(ˆ θ − E ˆ θ )] E [( E ˆ = θ − θ )] θ − θ ) 2 + 2( E ˆ E [(ˆ θ − E ˆ θ ) 2 ] + ( E ˆ θ − E ˆ θ ) E [( E ˆ θ − θ )] =

  31. . Mean Squared Error February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang . . . Property of MSE . . . Definition . 13 / 33 Summary . . . . . . . . . Recap MLE . Cramer-Rao Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . Mean Squared Error (MSE) of an estimator ˆ θ is defined as MSE (ˆ θ ) = E [(ˆ θ − θ )] 2 MSE (ˆ E [(ˆ θ − E ˆ θ + E ˆ θ − θ )] 2 θ ) = E [(ˆ θ − E ˆ θ ) 2 ] + E [( E ˆ θ − θ ) 2 ] + 2 E [(ˆ θ − E ˆ θ )] E [( E ˆ = θ − θ )] θ − θ ) 2 + 2( E ˆ E [(ˆ θ − E ˆ θ ) 2 ] + ( E ˆ θ − E ˆ θ ) E [( E ˆ θ − θ )] = Var (ˆ θ ) + Bias 2 ( θ ) =

  32. • Suppose that the true • Therefore, we cannot find an estimator that is uniformly the best in • Restrict the class of estimators, and find the ”best” estimator within . estimator can beat E X Var X n , then MSE MSE , and no . in terms of MSE when true E terms of MSE across all among all estimators the small class. Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 MSE MSE . MLE . . . . . . . . . Recap Evaluation Cramer-Rao . Summary Example i.i.d. 14 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . • X 1 , · · · , X n ∼ N ( µ, 1) • µ 1 = 1 , µ 2 = X .

  33. • Suppose that the true • Therefore, we cannot find an estimator that is uniformly the best in • Restrict the class of estimators, and find the ”best” estimator within . estimator can beat E X Var X n , then MSE MSE , and no . in terms of MSE when true . terms of MSE across all among all estimators the small class. Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 MSE 14 / 33 . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Example i.i.d. . . . . . . . . . . . . . . . . . . . . . . . . . . . • X 1 , · · · , X n ∼ N ( µ, 1) • µ 1 = 1 , µ 2 = X . µ 1 − µ ) 2 = (1 − µ ) 2 MSE (ˆ µ 1 ) = E (ˆ

  34. • Suppose that the true • Therefore, we cannot find an estimator that is uniformly the best in • Restrict the class of estimators, and find the ”best” estimator within . , then MSE MSE , and no estimator can beat in terms of MSE when true . . terms of MSE across all among all estimators the small class. Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 n 14 / 33 . Evaluation Cramer-Rao Summary Example MLE Recap . . . . . . . i.i.d. . . . . . . . . . . . . . . . . . . . . . . . . . . . . • X 1 , · · · , X n ∼ N ( µ, 1) • µ 1 = 1 , µ 2 = X . µ 1 − µ ) 2 = (1 − µ ) 2 MSE (ˆ µ 1 ) = E (ˆ E ( X − µ ) 2 = Var ( X ) = 1 MSE (ˆ µ 2 ) =

  35. • Therefore, we cannot find an estimator that is uniformly the best in • Restrict the class of estimators, and find the ”best” estimator within . Summary February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang the small class. among all estimators terms of MSE across all n . i.i.d. Example 14 / 33 . MLE Recap Cramer-Rao . . . . . . . . . Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . • X 1 , · · · , X n ∼ N ( µ, 1) • µ 1 = 1 , µ 2 = X . µ 1 − µ ) 2 = (1 − µ ) 2 MSE (ˆ µ 1 ) = E (ˆ E ( X − µ ) 2 = Var ( X ) = 1 MSE (ˆ µ 2 ) = • Suppose that the true µ = 1 , then MSE ( µ 1 ) = 0 < MSE ( µ 2 ) , and no estimator can beat µ 1 in terms of MSE when true µ = 1 .

  36. • Restrict the class of estimators, and find the ”best” estimator within . . February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang the small class. n . i.i.d. Example Summary 14 / 33 Cramer-Rao Evaluation MLE . . . . . . . . . Recap . . . . . . . . . . . . . . . . . . . . . . . . . . • X 1 , · · · , X n ∼ N ( µ, 1) • µ 1 = 1 , µ 2 = X . µ 1 − µ ) 2 = (1 − µ ) 2 MSE (ˆ µ 1 ) = E (ˆ E ( X − µ ) 2 = Var ( X ) = 1 MSE (ˆ µ 2 ) = • Suppose that the true µ = 1 , then MSE ( µ 1 ) = 0 < MSE ( µ 2 ) , and no estimator can beat µ 1 in terms of MSE when true µ = 1 . • Therefore, we cannot find an estimator that is uniformly the best in terms of MSE across all θ ∈ Ω among all estimators

  37. . Cramer-Rao February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang the small class. n . i.i.d. Example Summary . 14 / 33 Evaluation MLE Recap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . • X 1 , · · · , X n ∼ N ( µ, 1) • µ 1 = 1 , µ 2 = X . µ 1 − µ ) 2 = (1 − µ ) 2 MSE (ˆ µ 1 ) = E (ˆ E ( X − µ ) 2 = Var ( X ) = 1 MSE (ˆ µ 2 ) = • Suppose that the true µ = 1 , then MSE ( µ 1 ) = 0 < MSE ( µ 2 ) , and no estimator can beat µ 1 in terms of MSE when true µ = 1 . • Therefore, we cannot find an estimator that is uniformly the best in terms of MSE across all θ ∈ Ω among all estimators • Restrict the class of estimators, and find the ”best” estimator within

  38. • Find the lower bound of variances of any unbiased estimator of • If W is an unbiased estimator of . . for all , where W is any other unbiased estimator of (minimum variance). . How to find the Best Unbiased Estimator . . . . . . . X , say B . and satisfies Var W X B , then W is the best unbiased estimator. Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 Var W X 2 and Var W . . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Uniformly Minimum Variance Unbiased Estimator . . . (unbiased) for all X 1 E W . . Definition . 15 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . W ∗ ( X ) is the best unbiased estimator , or uniformly minimum variance unbiased estimator (UMVUE) of τ ( θ ) if,

  39. • Find the lower bound of variances of any unbiased estimator of • If W is an unbiased estimator of . , where W is any other unbiased estimator of (minimum variance). . How to find the Best Unbiased Estimator . . . . . . . . Var W X , say B . and satisfies Var W X B , then W is the best unbiased estimator. Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 for all X . 2 and Var W . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Uniformly Minimum Variance Unbiased Estimator . Definition . . . . . . 15 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . W ∗ ( X ) is the best unbiased estimator , or uniformly minimum variance unbiased estimator (UMVUE) of τ ( θ ) if, 1 E [ W ∗ ( X ) | θ ] = τ ( θ ) for all θ (unbiased)

  40. • Find the lower bound of variances of any unbiased estimator of • If W is an unbiased estimator of . How to find the Best Unbiased Estimator . . . . . . . . say B , . . and satisfies Var W X B , then W is the best unbiased estimator. Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . . Summary . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Uniformly Minimum Variance Unbiased Estimator . Definition . . . . 15 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . W ∗ ( X ) is the best unbiased estimator , or uniformly minimum variance unbiased estimator (UMVUE) of τ ( θ ) if, 1 E [ W ∗ ( X ) | θ ] = τ ( θ ) for all θ (unbiased) 2 and Var [ W ∗ ( X ) | θ ] ≤ Var [ W ( X ) | θ ] for all θ , where W is any other unbiased estimator of τ ( θ ) (minimum variance).

  41. • If W is an unbiased estimator of . . . . . How to find the Best Unbiased Estimator . . and satisfies . Var W X B , then W is the best unbiased estimator. Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . . Cramer-Rao . . . . . . . . . Recap MLE Evaluation . Summary Uniformly Minimum Variance Unbiased Estimator . Definition 15 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . W ∗ ( X ) is the best unbiased estimator , or uniformly minimum variance unbiased estimator (UMVUE) of τ ( θ ) if, 1 E [ W ∗ ( X ) | θ ] = τ ( θ ) for all θ (unbiased) 2 and Var [ W ∗ ( X ) | θ ] ≤ Var [ W ( X ) | θ ] for all θ , where W is any other unbiased estimator of τ ( θ ) (minimum variance). • Find the lower bound of variances of any unbiased estimator of τ ( θ ) , say B ( θ ) .

  42. . Definition February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang . . How to find the Best Unbiased Estimator . . . . . . . . 15 / 33 . . Summary . Cramer-Rao . Evaluation . MLE . Uniformly Minimum Variance Unbiased Estimator Recap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . W ∗ ( X ) is the best unbiased estimator , or uniformly minimum variance unbiased estimator (UMVUE) of τ ( θ ) if, 1 E [ W ∗ ( X ) | θ ] = τ ( θ ) for all θ (unbiased) 2 and Var [ W ∗ ( X ) | θ ] ≤ Var [ W ( X ) | θ ] for all θ , where W is any other unbiased estimator of τ ( θ ) (minimum variance). • Find the lower bound of variances of any unbiased estimator of τ ( θ ) , say B ( θ ) . • If W ∗ is an unbiased estimator of τ ( θ ) and satisfies Var [ W ∗ ( X ) | θ ] = B ( θ ) , then W ∗ is the best unbiased estimator.

  43. d E h x h x f X x f X x log f X x . . For h x and h x W x , if the differentiation and integrations are interchangeable, i.e. d d d x x d x . h x d x Then, a lower bound of Var W X is Var W X E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 2 Var W X . . . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Cramer-Rao inequality . Theorem 7.3.9 : Cramer-Rao Theorem . . is an estimator satisfying . . 1 E W X 16 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . Let X 1 , · · · , X n be a sample with joint pdf/pmf of f X ( x | θ ) . Suppose W ( X )

  44. d E h x h x f X x f X x log f X x . For h x and h x W x , if the differentiation and integrations are interchangeable, i.e. d d d x x d x 2 Var W X h x d x Then, a lower bound of Var W X is Var W X E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . . . . . . . . . . . . . Recap MLE Evaluation Cramer-Rao Summary Cramer-Rao inequality . Theorem 7.3.9 : Cramer-Rao Theorem . . is an estimator satisfying . . 16 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . Let X 1 , · · · , X n be a sample with joint pdf/pmf of f X ( x | θ ) . Suppose W ( X ) 1 E [ W ( X ) | θ ] = τ ( θ ) , ∀ θ ∈ Ω .

  45. d E h x h x f X x f X x log f X x . d x and h x W x , if the differentiation and integrations are interchangeable, i.e. d d d x h x x . d x Then, a lower bound of Var W X is Var W X E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 For h x . . Summary . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Cramer-Rao inequality . Theorem 7.3.9 : Cramer-Rao Theorem . . is an estimator satisfying . . 16 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . Let X 1 , · · · , X n be a sample with joint pdf/pmf of f X ( x | θ ) . Suppose W ( X ) 1 E [ W ( X ) | θ ] = τ ( θ ) , ∀ θ ∈ Ω . 2 Var [ W ( X ) | θ ] < ∞ .

  46. d E h x h x f X x f X x log f X x . d x . interchangeable, i.e. d d d x h x x . d x Then, a lower bound of Var W X is Var W X E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . 16 / 33 . . . . . . . . . . . Recap MLE Evaluation . Cramer-Rao Summary Cramer-Rao inequality . Theorem 7.3.9 : Cramer-Rao Theorem . . is an estimator satisfying . . . . . . . . . . . . . . . . . . . . . . . . . . Let X 1 , · · · , X n be a sample with joint pdf/pmf of f X ( x | θ ) . Suppose W ( X ) 1 E [ W ( X ) | θ ] = τ ( θ ) , ∀ θ ∈ Ω . 2 Var [ W ( X ) | θ ] < ∞ . For h ( x ) = 1 and h ( x ) = W ( x ) , if the differentiation and integrations are

  47. log f X x . d . . . . interchangeable, i.e. d Then, a lower bound of Var W X . is Var W X E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 is an estimator satisfying . . Evaluation . . . . . . . . . Recap MLE Theorem 7.3.9 : Cramer-Rao Theorem 16 / 33 . Cramer-Rao inequality Summary . Cramer-Rao . . . . . . . . . . . . . . . . . . . . . . . . . . Let X 1 , · · · , X n be a sample with joint pdf/pmf of f X ( x | θ ) . Suppose W ( X ) 1 E [ W ( X ) | θ ] = τ ( θ ) , ∀ θ ∈ Ω . 2 Var [ W ( X ) | θ ] < ∞ . For h ( x ) = 1 and h ( x ) = W ( x ) , if the differentiation and integrations are ∫ ∫ h ( x ) ∂ d θ E [ h ( x ) | θ ] = h ( x ) f X ( x | θ ) d x = ∂θ f X ( x | θ ) d x d θ x ∈X x ∈X

  48. . Theorem 7.3.9 : Cramer-Rao Theorem February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang E d d interchangeable, i.e. . . . . is an estimator satisfying . . . . MLE . . . . Cramer-Rao inequality . . . . . Recap 16 / 33 Evaluation . Summary Cramer-Rao . . . . . . . . . . . . . . . . . . . . . . . . . . Let X 1 , · · · , X n be a sample with joint pdf/pmf of f X ( x | θ ) . Suppose W ( X ) 1 E [ W ( X ) | θ ] = τ ( θ ) , ∀ θ ∈ Ω . 2 Var [ W ( X ) | θ ] < ∞ . For h ( x ) = 1 and h ( x ) = W ( x ) , if the differentiation and integrations are ∫ ∫ h ( x ) ∂ d θ E [ h ( x ) | θ ] = h ( x ) f X ( x | θ ) d x = ∂θ f X ( x | θ ) d x d θ x ∈X x ∈X Then, a lower bound of Var [ W ( X ) | θ ] is [ τ ′ ( θ )] 2 Var [ W ( X )] ≥ [ { ∂ ∂θ log f X ( x | θ ) } 2 ]

  49. log f X X log f X X log f X X log f X X log f X X log f X X log f X X February 14th, 2013 Using Var X Var W X Cov W X Var , EX EX Biostatistics 602 - Lecture 11 Var W X Var E E Hyun Min Kang Var . . Cov W X . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Proving Cramer-Rao Theorem (1/4) By Cauchy-Schwarz inequality, Replacing X and Y , 17 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . [ Cov ( X , Y )] 2 ≤ Var ( X ) Var ( Y )

  50. log f X X log f X X log f X X log f X X log f X X . EX Cov W X Var Using Var X EX Var , . E E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 Var W X 17 / 33 . . . . . . MLE Evaluation Cramer-Rao . Summary Proving Cramer-Rao Theorem (1/4) By Cauchy-Schwarz inequality, Replacing X and Y , . . . Recap . . . . . . . . . . . . . . . . . . . . . . . . . . [ Cov ( X , Y )] 2 ≤ Var ( X ) Var ( Y ) [ ∂ ] 2 [ Cov { W ( X ) , ∂ ] ∂θ log f X ( X | θ ) } ≤ Var [ W ( X )] Var ∂θ log f X ( X | θ )

  51. log f X X log f X X log f X X . Var Using Var X EX EX , E Var Replacing X and Y , E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . 17 / 33 Summary By Cauchy-Schwarz inequality, . . . . . . . . . Recap Proving Cramer-Rao Theorem (1/4) MLE Evaluation Cramer-Rao . . . . . . . . . . . . . . . . . . . . . . . . . . . [ Cov ( X , Y )] 2 ≤ Var ( X ) Var ( Y ) [ ∂ ] 2 [ Cov { W ( X ) , ∂ ] ∂θ log f X ( X | θ ) } ≤ Var [ W ( X )] Var ∂θ log f X ( X | θ ) ] 2 [ Cov { W ( X ) , ∂ ∂θ log f X ( X | θ ) } ≥ Var [ W ( X )] [ ∂ ] ∂θ log f X ( X | θ )

  52. . . February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang E Var Var . Replacing X and Y , By Cauchy-Schwarz inequality, Proving Cramer-Rao Theorem (1/4) Summary 17 / 33 Cramer-Rao . . . . . . . . . Recap MLE Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . [ Cov ( X , Y )] 2 ≤ Var ( X ) Var ( Y ) [ ∂ ] 2 [ Cov { W ( X ) , ∂ ] ∂θ log f X ( X | θ ) } ≤ Var [ W ( X )] Var ∂θ log f X ( X | θ ) ] 2 [ Cov { W ( X ) , ∂ ∂θ log f X ( X | θ ) } ≥ Var [ W ( X )] [ ∂ ] ∂θ log f X ( X | θ ) Using Var ( X ) = EX 2 − ( EX ) 2 , [ ∂ [{ ∂ [ ∂ } 2 ] ] 2 ] ∂θ log f X ( X | θ ) ∂θ log f X ( X | θ ) − E ∂θ log f X ( X | θ ) =

  53. f X x f X x f X x f X x f X x log f X X log f X X . d x x d x d d d x x . (by assumption) d d Var E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 x 18 / 33 Cramer-Rao E MLE Evaluation . . Summary . . . . . Proving Cramer-Rao Theorem (2/4) . . . Recap . . . . . . . . . . . . . . . . . . . . . . . . . . [ ∂ [ ∂ ] ∫ ] ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f X ( x | θ ) d x x ∈X

  54. f X x f X x log f X X log f X X . d x d x d d x d (by assumption) . d Var E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 x 18 / 33 Proving Cramer-Rao Theorem (2/4) Recap . Summary E Evaluation MLE Cramer-Rao . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . [ ∂ [ ∂ ] ∫ ] ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f X ( x | θ ) d x x ∈X ∂ ∂θ f X ( x | θ ) ∫ = f X ( x | θ ) f X ( x | θ ) d x x ∈X

  55. f X x log f X X log f X X . d d x d x (by assumption) Var d . E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 d 18 / 33 . . Summary Proving Cramer-Rao Theorem (2/4) Evaluation MLE Recap . . . . . Cramer-Rao . . E . . . . . . . . . . . . . . . . . . . . . . . . . . . [ ∂ [ ∂ ] ∫ ] ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f X ( x | θ ) d x x ∈X ∂ ∂θ f X ( x | θ ) ∫ = f X ( x | θ ) f X ( x | θ ) d x x ∈X ∫ ∂ = ∂θ f X ( x | θ ) d x x ∈X

  56. log f X X log f X X . Proving Cramer-Rao Theorem (2/4) February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang E Var d d (by assumption) d . E 18 / 33 Summary Cramer-Rao MLE Recap . . . . . . . . . . Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . [ ∂ [ ∂ ] ∫ ] ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f X ( x | θ ) d x x ∈X ∂ ∂θ f X ( x | θ ) ∫ = f X ( x | θ ) f X ( x | θ ) d x x ∈X ∫ ∂ = ∂θ f X ( x | θ ) d x x ∈X ∫ = f X ( x | θ ) d x d θ x ∈X

  57. log f X X log f X X . Proving Cramer-Rao Theorem (2/4) February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang E Var d (by assumption) d . E 18 / 33 Summary . Evaluation Recap . . . . . . . . . MLE Cramer-Rao . . . . . . . . . . . . . . . . . . . . . . . . . . [ ∂ [ ∂ ] ∫ ] ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f X ( x | θ ) d x x ∈X ∂ ∂θ f X ( x | θ ) ∫ = f X ( x | θ ) f X ( x | θ ) d x x ∈X ∫ ∂ = ∂θ f X ( x | θ ) d x x ∈X ∫ = f X ( x | θ ) d x d θ x ∈X = d θ 1 = 0

  58. . Summary February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang E Var d (by assumption) d . E Proving Cramer-Rao Theorem (2/4) 18 / 33 . Cramer-Rao Recap . . . . . . . . . MLE Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . [ ∂ [ ∂ ] ∫ ] ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f X ( x | θ ) d x x ∈X ∂ ∂θ f X ( x | θ ) ∫ = f X ( x | θ ) f X ( x | θ ) d x x ∈X ∫ ∂ = ∂θ f X ( x | θ ) d x x ∈X ∫ = f X ( x | θ ) d x d θ x ∈X = d θ 1 = 0 [ ∂ [{ ∂ } 2 ] ] ∂θ log f X ( X | θ ) = ∂θ log f X ( X | θ )

  59. log f X X log f X X log f X X log f X x f X x f X x W x f X x d E W X f x February 14th, 2013 f x d x Biostatistics 602 - Lecture 11 x W x f x Hyun Min Kang d x d W x W x d d x (by assumption) d d x . x . . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Proving Cramer-Rao Theorem (3/4) Cov E W X E W X E E W X 19 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . [ ] W ( X ) , ∂ ∂θ log f X ( X | θ )

  60. log f X X log f X x f X x f X x W x f X x d E W X Hyun Min Kang f x x W x February 14th, 2013 f x d x x W x Biostatistics 602 - Lecture 11 f x x d x d W x E W X d d x (by assumption) d d . 19 / 33 . Cov . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Summary Proving Cramer-Rao Theorem (3/4) E . . . . . . . . . . . . . . . . . . . . . . . . . . [ ] W ( X ) , ∂ ∂θ log f X ( X | θ ) [ ∂ [ ] ] W ( X ) · ∂ = ∂θ log f X ( X | θ ) − E [ W ( X )] E ∂θ log f X ( X | θ )

  61. log f X x f X x f X x W x f X x d E W X x x W x f x d x x W x f x f x d x . W x . d d x (by assumption) d d d Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 E 19 / 33 Recap Evaluation Summary . . . . . . E Cramer-Rao Proving Cramer-Rao Theorem (3/4) Cov . . . . MLE . . . . . . . . . . . . . . . . . . . . . . . . . . [ ] W ( X ) , ∂ ∂θ log f X ( X | θ ) [ ∂ [ ] ] W ( X ) · ∂ = ∂θ log f X ( X | θ ) − E [ W ( X )] E ∂θ log f X ( X | θ ) [ ] W ( X ) · ∂ = ∂θ log f X ( X | θ )

  62. f X x f X x W x f X x d E W X . x W x f x f x d x x W x d d . x (by assumption) d d d Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 E 19 / 33 Proving Cramer-Rao Theorem (3/4) . Cov . Cramer-Rao Evaluation MLE Summary . . . . . Recap . E . . . . . . . . . . . . . . . . . . . . . . . . . . . . [ ] W ( X ) , ∂ ∂θ log f X ( X | θ ) [ ∂ [ ] ] W ( X ) · ∂ = ∂θ log f X ( X | θ ) − E [ W ( X )] E ∂θ log f X ( X | θ ) [ ] W ( X ) · ∂ ∫ W ( x ) ∂ = ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f ( x | θ ) d x x ∈X

  63. f X x W x f X x d E W X . x W x d d x (by assumption) E d d d Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . E 19 / 33 Proving Cramer-Rao Theorem (3/4) . Cramer-Rao Cov Evaluation MLE Recap . . . . . . . . . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . [ ] W ( X ) , ∂ ∂θ log f X ( X | θ ) [ ∂ [ ] ] W ( X ) · ∂ = ∂θ log f X ( X | θ ) − E [ W ( X )] E ∂θ log f X ( X | θ ) [ ] W ( X ) · ∂ ∫ W ( x ) ∂ = ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f ( x | θ ) d x x ∈X ∂ ∂θ f X ( x | θ ) ∫ = W ( x ) f ( x | θ ) d x f ( x | θ ) x ∈X

  64. W x f X x d E W X . E February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang d d d (by assumption) x d d . E 19 / 33 Cov . Evaluation MLE Proving Cramer-Rao Theorem (3/4) Recap . . . . . . . . . Summary Cramer-Rao . . . . . . . . . . . . . . . . . . . . . . . . . . [ ] W ( X ) , ∂ ∂θ log f X ( X | θ ) [ ∂ [ ] ] W ( X ) · ∂ = ∂θ log f X ( X | θ ) − E [ W ( X )] E ∂θ log f X ( X | θ ) [ ] W ( X ) · ∂ ∫ W ( x ) ∂ = ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f ( x | θ ) d x x ∈X ∂ ∂θ f X ( x | θ ) ∫ ∫ W ( x ) ∂ = W ( x ) f ( x | θ ) d x = ∂θ f X ( x | θ ) f ( x | θ ) x ∈X x ∈X

  65. d E W X . Proving Cramer-Rao Theorem (3/4) February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang d d d (by assumption) d . E E Cov 19 / 33 Summary . . Recap . . . . . . . . Evaluation MLE Cramer-Rao . . . . . . . . . . . . . . . . . . . . . . . . . . [ ] W ( X ) , ∂ ∂θ log f X ( X | θ ) [ ∂ [ ] ] W ( X ) · ∂ = ∂θ log f X ( X | θ ) − E [ W ( X )] E ∂θ log f X ( X | θ ) [ ] W ( X ) · ∂ ∫ W ( x ) ∂ = ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f ( x | θ ) d x x ∈X ∂ ∂θ f X ( x | θ ) ∫ ∫ W ( x ) ∂ = W ( x ) f ( x | θ ) d x = ∂θ f X ( x | θ ) f ( x | θ ) x ∈X x ∈X ∫ W ( x ) f X ( x | θ ) = d θ x ∈X

  66. . Summary February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang d (by assumption) d . E E Cov Proving Cramer-Rao Theorem (3/4) 19 / 33 . Cramer-Rao . . . . . . . . MLE . Recap Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . [ ] W ( X ) , ∂ ∂θ log f X ( X | θ ) [ ∂ [ ] ] W ( X ) · ∂ = ∂θ log f X ( X | θ ) − E [ W ( X )] E ∂θ log f X ( X | θ ) [ ] W ( X ) · ∂ ∫ W ( x ) ∂ = ∂θ log f X ( X | θ ) = ∂θ log f X ( x | θ ) f ( x | θ ) d x x ∈X ∂ ∂θ f X ( x | θ ) ∫ ∫ W ( x ) ∂ = W ( x ) f ( x | θ ) d x = ∂θ f X ( x | θ ) f ( x | θ ) x ∈X x ∈X ∫ W ( x ) f X ( x | θ ) = d θ x ∈X d θτ ( θ ) = τ ′ ( θ ) = d θ E [ W ( X )] = d

  67. log f X X log f X X log f X X log f X X . E Cov W X Therefore, Cramer-Rao lower bound is Var W X Var Cov W X Var E Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . 20 / 33 From the previous results Proving Cramer-Rao Theorem (4/4) . . Summary . . . . . . . . Cramer-Rao Recap Evaluation MLE . . . . . . . . . . . . . . . . . . . . . . . . . . [ ∂ [{ ∂ } 2 ] ] ∂θ log f X ( X | θ ) = ∂θ log f X ( X | θ )

  68. log f X X . Proving Cramer-Rao Theorem (4/4) February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang E Var Therefore, Cramer-Rao lower bound is Cov . E Var From the previous results 20 / 33 Summary MLE Cramer-Rao . . . . . . . . . Evaluation . Recap . . . . . . . . . . . . . . . . . . . . . . . . . . [ ∂ [{ ∂ } 2 ] ] ∂θ log f X ( X | θ ) = ∂θ log f X ( X | θ ) [ ] W ( X ) , ∂ τ ′ ( θ ) ∂θ log f X ( X | θ ) = ] 2 Cov { W ( X ) , ∂ [ ∂θ log f X ( X | θ ) } Var [ W ( X )] ≥ [ ∂ ] ∂θ log f X ( X | θ )

  69. . Summary February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang E Var Therefore, Cramer-Rao lower bound is . E Var From the previous results Proving Cramer-Rao Theorem (4/4) Cov . MLE . . . . . . . . . Cramer-Rao Recap 20 / 33 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . [ ∂ [{ ∂ } 2 ] ] ∂θ log f X ( X | θ ) = ∂θ log f X ( X | θ ) [ ] W ( X ) , ∂ τ ′ ( θ ) ∂θ log f X ( X | θ ) = ] 2 Cov { W ( X ) , ∂ [ ∂θ log f X ( X | θ ) } Var [ W ( X )] ≥ [ ∂ ] ∂θ log f X ( X | θ ) [ τ ′ ( θ )] 2 = { ∂ [ ∂θ log f X ( X | θ ) } 2 ]

  70. log f X X log f X X log f X X . . . Proof . . . . . . Var W X . We need to show that E nE Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 nE 21 / 33 . . . . . . . . Recap MLE Evaluation in the above Cramer-Rao theorem hold, then the lower-bound of Cramer-Rao . Summary Cramer-Rao bound in iid case . Corollary 7.3.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . If X 1 , · · · , X n are iid samples from pdf/pmf f X ( x | θ ) , and the assumptions Var [ W ( X ) | θ ] becomes

  71. log f X X log f X X . . . Proof . . . . . . . . We need to show that E nE Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 nE 21 / 33 in the above Cramer-Rao theorem hold, then the lower-bound of Summary . . . . . . . . . Recap MLE Evaluation Cramer-Rao . Cramer-Rao bound in iid case . Corollary 7.3.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . If X 1 , · · · , X n are iid samples from pdf/pmf f X ( x | θ ) , and the assumptions Var [ W ( X ) | θ ] becomes [ τ ′ ( θ )] 2 Var [ W ( X )] ≥ { ∂ [ ∂θ log f X ( X | θ ) } 2 ]

  72. . . . in the above Cramer-Rao theorem hold, then the lower-bound of . nE . Proof . Corollary 7.3.10 We need to show that E nE Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 . 21 / 33 . Evaluation . . . . . . . . . Recap Cramer-Rao bound in iid case MLE Cramer-Rao Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . If X 1 , · · · , X n are iid samples from pdf/pmf f X ( x | θ ) , and the assumptions Var [ W ( X ) | θ ] becomes [ τ ′ ( θ )] 2 Var [ W ( X )] ≥ { ∂ [ ∂θ log f X ( X | θ ) } 2 ] [{ ∂ [{ ∂ } 2 ] } 2 ] ∂θ log f X ( X | θ ) = ∂θ log f X ( X | θ )

  73. log f X X i log f X X i log f X X i log f X X i log f X X j . E n i E n i n E . i i j Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 n 22 / 33 Proving Corollary 7.3.10 Recap Summary E Evaluation MLE Cramer-Rao . . . . . . . . E . . . . . . . . . . . . . . . . . . . . . . . . . . . . [{ ∂  } 2  } 2 ] { ∂ ∏ ∂θ log f X ( X | θ ) f X ( X i | θ ) =   ∂θ log i =1

  74. log f X X i log f X X i log f X X i log f X X j . . E n E n i i E n E i j Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 n 22 / 33 Summary . . Cramer-Rao E Evaluation MLE Recap Proving Corollary 7.3.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . [{ ∂  } 2  } 2 ] { ∂ ∏ ∂θ log f X ( X | θ ) f X ( X i | θ ) =   ∂θ log i =1  } 2  { ∂ ∑ = log f X ( X i | θ )   ∂θ i =1

  75. log f X X i log f X X i log f X X j . . n E E n i i E j Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 n E 22 / 33 Summary Cramer-Rao Evaluation MLE Proving Corollary 7.3.10 Recap . . . . . . . E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . [{ ∂  } 2  } 2 ] { ∂ ∏ ∂θ log f X ( X | θ ) f X ( X i | θ ) =   ∂θ log i =1  } 2  { ∂ ∑ = log f X ( X i | θ )   ∂θ i =1  } 2  { n ∂ ∑ = ∂θ log f X ( X i | θ )   i =1

  76. . Summary February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang E E n . E n E E Proving Corollary 7.3.10 22 / 33 . Cramer-Rao Recap . . . MLE . . . . . . Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . [{ ∂  } 2  } 2 ] { ∂ ∏ ∂θ log f X ( X | θ ) f X ( X i | θ ) =   ∂θ log i =1  } 2  { ∂ ∑ = log f X ( X i | θ )   ∂θ i =1  } 2  { n ∂ ∑ = ∂θ log f X ( X i | θ )   i =1 { ∂ } 2 + [∑ n ∂θ log f X ( X i | θ ) = i =1 ] ∂θ log f X ( X i | θ ) ∂ ∂ ∑ ∂θ log f X ( X j | θ ) i ̸ = j

  77. log f X X log f X X i log f X X i log f X X . E E E n i i n . E nE Hyun Min Kang Biostatistics 602 - Lecture 11 February 14th, 2013 E 23 / 33 . MLE Evaluation Cramer-Rao . . . E Summary . Proving Corollary 7.3.10 Recap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Because X 1 , · · · , X n are independent,   ∂θ log f X ( X i | θ ) ∂ ∂ ∑ ∂θ log f X ( X j | θ )  i ̸ = j [ ∂ [ ∂ ] ] ∑ ∂θ log f X ( X i | θ ) ∂θ log f X ( X j | θ ) = = 0 i ̸ = j

  78. log f X X i log f X X . E February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang nE E i n E E E . E Proving Corollary 7.3.10 . . . . . . . . . . Recap MLE Evaluation Cramer-Rao 23 / 33 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . Because X 1 , · · · , X n are independent,   ∂θ log f X ( X i | θ ) ∂ ∂ ∑ ∂θ log f X ( X j | θ )  i ̸ = j [ ∂ [ ∂ ] ] ∑ ∂θ log f X ( X i | θ ) ∂θ log f X ( X j | θ ) = = 0 i ̸ = j [{ ∂ { ∂ } 2 ] [ n } 2 ] ∑ ∂θ log f X ( X | θ ) ∂θ log f X ( X i | θ ) = i =1

  79. log f X X . Proving Corollary 7.3.10 February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang nE E n E E E . E E 23 / 33 Summary . Recap . . . . . . . Evaluation MLE . Cramer-Rao . . . . . . . . . . . . . . . . . . . . . . . . . . . Because X 1 , · · · , X n are independent,   ∂θ log f X ( X i | θ ) ∂ ∂ ∑ ∂θ log f X ( X j | θ )  i ̸ = j [ ∂ [ ∂ ] ] ∑ ∂θ log f X ( X i | θ ) ∂θ log f X ( X j | θ ) = = 0 i ̸ = j [{ ∂ { ∂ } 2 ] [ n } 2 ] ∑ ∂θ log f X ( X | θ ) ∂θ log f X ( X i | θ ) = i =1 [{ ∂ } 2 ] ∑ = ∂θ log f X ( X i | θ ) i =1

  80. . Summary February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang nE E n E E . E E Proving Corollary 7.3.10 E . Evaluation . . . . . . . . . Recap Cramer-Rao MLE 23 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . Because X 1 , · · · , X n are independent,   ∂θ log f X ( X i | θ ) ∂ ∂ ∑ ∂θ log f X ( X j | θ )  i ̸ = j [ ∂ [ ∂ ] ] ∑ ∂θ log f X ( X i | θ ) ∂θ log f X ( X j | θ ) = = 0 i ̸ = j [{ ∂ { ∂ } 2 ] [ n } 2 ] ∑ ∂θ log f X ( X | θ ) ∂θ log f X ( X i | θ ) = i =1 [{ ∂ } 2 ] ∑ = ∂θ log f X ( X i | θ ) i =1 [{ ∂ } 2 ] = ∂θ log f X ( X | θ )

  81. log f X X . . February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang . and Because nE Var W X Remark from Corollary 7.3.10 Summary . Cramer-Rao Evaluation MLE Recap . . . . . . . . . 24 / 33 . . . . . . . . . . . . . . . . . . . . . . . . . . In iid case, Cramer-Rao lower bound for an unbiased estimator of θ is

  82. . Cramer-Rao February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang . and Because nE . Summary . Remark from Corollary 7.3.10 24 / 33 Recap MLE . . . . Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . In iid case, Cramer-Rao lower bound for an unbiased estimator of θ is 1 Var [ W ( X )] ≥ { ∂ [ ∂θ log f X ( X | θ ) } 2 ]

  83. . Evaluation February 14th, 2013 Biostatistics 602 - Lecture 11 Hyun Min Kang nE Remark from Corollary 7.3.10 . . Cramer-Rao Summary 24 / 33 . MLE Recap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . In iid case, Cramer-Rao lower bound for an unbiased estimator of θ is 1 Var [ W ( X )] ≥ { ∂ [ ∂θ log f X ( X | θ ) } 2 ] Because τ ( θ ) = θ and τ ′ ( θ ) = 1 .

Recommend


More recommend