Maximum Likelihood Large Sample Theory

Transcription

1 Maximum Likelihood Large Sample Theory MIT Dr. Kempthorne Spring

2 Outline 1 Large Sample Theory of Maximum Likelihood Estimates 2

3 Asymptotic Results: Overview Asymptotic Framework Data Model : X n = (X 1, X 2,..., X n ). i.i.d. sample with pdf/pmf f (x n 1,..., x n θ) = i=1 f (x i θ) Data Realization: X n = x n = (x 1,..., x n ) Likelihood of θ (given x n ): lik(θ) = f (x 1,..., x n θ) θˆ n: MLE of θ given xn = ( x1,..., xn) { θ ˆn, n } : sequence of MLEs indexed by sample size n Results: L Consistency: θˆn θ L κ(θ)/n Asymptotic Variance: σˆ θ n = Var(ˆ θ n ) where κ(θ) is an explicit function of the pdf/pmf f ( θ). Limiting Distribution: n(θˆn θ) N(0, κ(θ)). L 3

4 Theorems for Asymptotic Results Setting: x 1,..., x n a realization of an i.i.d. sample from distribution with density/pmf f (x θ). n (θ) = i=1 ln f (x i θ) θ 0 : true value of θ ˆθ n : the MLE Theorem A Under appropriate smoothness conditions on f, the MLE θˆn is consistent, i.e., for any true value θ 0, for every E > 0, P( θˆn θ 0 > E) 0. Proof: Weak Law of Large Numbers (WLLN) 1 L n (θ) E [log f (x θ) θ 0 ] = log[f (x θ)]f (x θ 0 )dx (Note!! statement holds for every θ given any value of θ 0.) 4

5 Theorem 8.5.2A (continued) Proof (continued): The MLE θˆn maximizes 1 n (θ) Since 1 n (θ) E [log f (x θ) θ 0 ], ˆθ n is close to θ maximizing E [log f (x θ) θ 0 ] Under smoothness conditions on f (x θ), θ maximizes E [log f (x θ) θ 0 ] if θ solves d dθ (E[log f (x θ) θ 0 ]) = 0 Claim: θ = θ 0 : d d dθ (E[log f (x θ) θ 0 ]) = dθo log[f (x θ)]f { (x θ 0 )dx d = log[f (x θ)] f (x θ 0 )dx (dθ x d [f (x θ)] dθ = f (x θ) f (x θ 0 )dx d (at θ = θ 0 ) = [ dθ [f (x θ)]dx] θ=θ0 d = [ d f (x θ)dx] θ=θ0 = (1) 0. dθ 5 dθ

6 Theorems for Asymptotic Results Theorem 8.5.B Under smoothness conditions on f (x θ), L n(θˆn - θ 00 ) N(0, 1/I( ( θ 0 )) 2 where I(θ) = E log f (X θ) θ θ 2 Proof: Using the Taylor approximation to ' (θ), centered at θ 0 consider the following development: 0 = ' (θˆ) ' (θ θ θ) '' 0 ) + (ˆ (θ 0 ) ' (θ 0 ) = (θˆ θ) '' (θ 0 ) n[ 1 n ' (θ 0 )] = n(θˆ θ) 1 n [ '' (θ 0 )] L By the CLT n[ 1 ' (θ 0 )] N(0, I(θ 0 )) (Lemma A) n L By the WLLN 1 [ '' (θ 0 )] I(θ 0 ) n L Thus: n(θˆ θ) (1/I(θ 0 )) 2 N(0, I(θ 0 )) = N(0, 1/I(θ 0 )) 6

7 Theorems for Asymptotic Results Lemma A (extended). For the distribution with pdf/pmf f (x θ) define The Score Function U(X ; θ) = θ logf (X θ) The (Fisher) Information - 0 of the distribution ( ) I(θ) = E 2 log f (X θ) θ θ 2 Then under sufficient smoothness conditions on f (x θ) (a). E [U(X ; θ) θ] = 0 (b). Var[U(X ; θ) θ] = I(θ) o { = E [U(X ; θ)] 2 θ] Proof: Differentiate f (x θ)dx = 1 with respect to θ two times. Interchange the order of differentiation and integration. (a) follows from the first derivative. (b) follows from the second derivative. 7

8 Theorems for Asymptotic Results Qualifications/Extensions Results require true θ 0 to lie in interior of the parameter space. Results require that {x : f (x θ) > 0} not vary with θ. Results extend to multi-dimensional θ Vector-valued Score Function Matrix-valued (Fisher) Information 8

9 Outline 1 Large Sample Theory of Maximum Likelihood Estimates 9

10 Confidence Intervals Confidence Interval for a Normal Mean X 1, X 2,..., X n i.i.d. N(µ, σ 0 2 ), unknown mean µ (known σ 0 2 ) Parameter Estimate: X (sample mean) X = 1 n1 X i n i=1 E [X ] = µ Var[X ] = σ 2 = σ 2 0 /n X A 95% confidence interval for µ is a random interval, calculated from the data, that contains µ with probability 0.95, no matter what the value of the true µ. 10

11 Confidence Interval for a Normal Mean Sampling distribution of X X N(µ, σ 2 ) = N(µ, σ 2 0 /n) X X µ = Z = N(0, 1) (standard normal) σ X For any 0 < α < 1, (e.g. α = 0.05) Define z(α/2) : P(Z > z(α/2)) = α/2 By symmetry of the standard normal distribution P ( z(α/2) < Z < +z(α/2)) = 1 α i.e., ( x X µ P z(α/2) < < +z(α/2) = 1 α σ X Re-express interval of Z as interval of µ: o { P X z(α/2)σ X < µ < X + z(α/2)σ X = 1 α The interval given by [X ± z(α/2) σ ] is the X 100(1 α)% confidence interval for µ 11

12 Confidence Interval for a Normal Mean Important Properties/Qualifications The confidence interval is random. The parameter µ is not random. 100(1 α)%: the confidence-level of the confidence interval is the probability that the random interval contains the fixed parameter µ (the coverage probability of the confidence interval) Given a data realization of (X 1,..., X n ) = (x 1,..., x n ) µ is either inside the confidence interval or not. The confidence level scales the reliability of a sequence of confidence intervals constructed in this way. The confidence interval quantifies the uncertainty of the parameter estimate. 12

13 Confidence Intervals for Normal Distribution Parameters Normal Distribution with unknown mean (µ) and variance σ 2. X 1,..., X n i.i.d. N(µ, σ 2 ), with MLEs: µˆ = X n1 σˆ2 = 1 (X i X ) 2 n i=1 Confidence interval for µ based on the T -Statistic X µ T = t n 1 S/ n where t n 1 is Student s t distribution with (n 1) degrees of 1 n n 1 i=1 freedom and S 2 = (X i X ) 2. Define t n 1 (α/2) : P(t n 1 > t n 1 (α/2)) = α/2 By symmetry of Student s t distribution P [ t n 1 (α/2) < T < +t n 1 (α/2)] = 1 α [ X µ i.e., P t n 1 (α/2) < < +t n 1 (α/2) = 1 α S n 13

14 Confidence Interval For Normal Mean Re-express [ interval of T as interval ] of µ: X µ P t n 1 < < +t n 1 = 1 α S n = P X t n 1 (α/2)s/ n < µ < X + t n 1 (α/2)s/ n = 1 α The interval given by [X ± t n 1 (α/2) S/ n] is the 100(1 α)% confidence interval for µ Properties of confidence intervals for µ Center is X = µˆmle Width proportional to ˆσ µˆmle = S/ n (random!) 14

15 Confidence Interval For Normal Variance Normal Distribution with unknown mean (µ) and variance σ 2. X 1,..., X n i.i.d. N(µ, σ 2 ), with MLEs: µˆ = X n1 σˆ2 = 1 (X i X ) 2 n i=1 Confidence interval for σ 2 based on the sampling distribution of the MLE ˆσ 2. nσˆ2 Ω = χ 2 σ 2 n 1. where χ 2 n 1 is Chi-squared distribution with (n 1) d.f. Define χ 2 n 1 (α ) : P(χ 2 n 1 > χ n 2 1 (α )) = α Using α = o α/2 and α = (1 α/2), { P +χ2 n 1 (1 α/2) < Ω < +χ ( 2 n 1 (α/2) = x 1 α nσˆ2 i.e., P +χ2 n 1 (1 α/2) < < +χ 2 σ 2 n 1(α/2) = 1 α 15

16 Confidence Interval for Normal Variance Re-express interval of Ω as interval of σ 2 : P[+χn 1 2 (1 α/2) < Ω < +χ 2 n 1 (α/2)] = 1 α nσˆ2 P[+χn 2 1(1 α/2) < < +χ 2 n 1(α/2)] = 1 α σ 2 nσˆ2 nσˆ2 P[ < σ 2 < ] = 1 α χ 2 (α/2) χ 2 (1 α/2) n 1 n 1 The 100(1 α)% confidence interval for σ 2 is given by nσˆ2 nσˆ2 [ < σ 2 < ] χn 2 1(α/2) χn 2 1(1 α/2) Properties of confidence interval for σ 2 Asymmetrical about the MLE ˆσ 2. Width proportional to ˆσ 2 (random!) 100(1 α)% a confidence interval for σ a immediate: n n [ˆσ( 2 ) < σ < σˆ( )] (α/2) (1 α/2) χ n 1 χn

17 Confidence Intervals For Normal Distribution Parameters Important Features of Normal Distribution Case Required use of exact sampling distributions of the MLEs µˆ and ˆσ 2. Construction of each confidence interval based on pivotal quantity: a function of the data and the parameters whose distribution does not involve any unknown parameters. Examples of pivotals X µ T = t n 1 S/ n nσˆ2 Ω = χ 2 σ 2 n 1 17

18 Confidence Intervals Based On Large Sample Theory Asymptotic Framework (Re-cap) Data Model : X n = (X 1, X 2,..., X n ). i.i.d. sample with pdf/pmf f (x n 1,..., x n θ) = i=1 f (x i θ) Likelihood of θ given X n = x n = (x 1,..., x n ): lik(θ) = f (x 1,..., x n θ) θˆ n = θˆ n (x n ): MLE of θ given x n = (x 1,..., x n ) {θˆn, n }: sequence of MLEs indexed by sample size n Results (subject to sufficient smoothness conditions on f ) L Consistency: θˆn θ L 1 Asymptotic Variance: σˆ θ n = Var(θˆn ) ni(θ) o { 2 - ) d d where I(θ) = E[ dθ [log f (x θ)] ] = E [ 2 [log f (x θ)] ] dθ 2 Limiting Distribution: ni(θ)(θˆn θ) L N(0, 1). MIT Maximum LikelihoodLarge Sample Theory 18

19 Confidence Intervals Based on Large Sample Theory Large-Sample Confidence Interval Exploit the limiting pivotal quantity L Z n = ni(θ)(θˆn θ) N(0, 1). I.e. P( z(α/2) < Z n < +z(α/2)) 1 α P( z(α/2) < ni(θ)(θˆn θ) < +z(α/2)) 1 α P( z(α/2) < ni(θˆn)(θˆn θ) < +z(α/2)) 1 α Note (!): I(θˆn) substituted for I(θ) Re-express interval of Z n as interval of θ: P(θˆn z(α/2) 1 < θ < θˆn + z(α/2) 1 ) 1 α ni(θˆ) ni(θˆ) The interval given by [θˆn ± z(α/2) 1 ] is the ni(θˆ) 100(1 α)% confidence interval (large sample) for θ 19

20 Large-Sample Confidence Intervals Example 8.5.B. Poisson Distribution X 1,..., X n i.i.d. Poisson(λ) f (x λ) = λx x! e λ n (λ) = i=1[x i ln(λ) λ ln(x!)] MLE λˆ d (λ) n1 x i λˆ solves: = [ 1] = 0; λˆ = X. dλ λ i=1 d 2 X i 1 I(λ) = E [ log f (x λ)] = E [ ] = dλ 2 λ 2 λ ˆλ λ L Z n = ni(λˆ)(λˆ λ) = N(0, 1) ˆλ/n Large-Sample Confidence Interval for λ Approximate 100(1 α)% confidence interval for λ X [λˆ ± σˆλˆ] = [X ± z(α/2) n ] 20

21 Confidence Interval for Poisson Parameter Deaths By Horse Kick in Prussian Army (Bortkiewicz, 1898) Annual Counts of fatalities in 10 corps of Prussian cavalry over a period of 20 years. n = 200 corps-years worth of data. Annual Fatalities Observed Model X 1,..., X 200 as i.i.d. Poisson(λ). ˆ 122 λ MLE = X = 200 = 0.61 ˆ σˆˆ λ MLE = λ MLE /n =.0552 For an 95% confidence interval, z(α/2) = 1.96 giving 0.61 ± (1.96)(.0552) = [.5018,.7182] 21

22 Confidence Interval for Multinomial Parameter Definition: Multinomial Distribution W 1,..., W n are iid Multinomial(1, probs = (p 1,..., p m )) r.v.s The sample space of each W i is W = {1, 2,..., m}, a set of m distinct outcomes. P(W i = k) = p k, k = 1, 2,..., m. Define Count Statistics from Multinomial Sample: n X k = i=1 1(W i = k), (sum of indicators of outcome k), k = 1,..., m X = (X 1,..., X m ) follows a multinomial distribution n!. m x f j (x1,..., x n p 1,..., p m ) = m j=1 p j j=1 x i! where (p 1,..., p m ) is the vector of cell probabilities with m m i=1 p i = 1 and n = j=1 x i is the total count. Note: for m = 2, the W i are Bernoulli(p 1 ) X 1 is Binomial(n, p 1 ) and X 2 n X 1 22

23 MLEs of Multinomial Parameter Maximum Likelihood Estimation for Multinomial Likelihood function of counts lik(p 1,..., p m ) = log[f (x 1,..., x m p 1,..., p m )] m m = log(n!) j=1 log(x j!) + j=1 x j log(p j ) Note: Likelihood function of Multinomial Sample w 1,...,, w n lik (p 1,..., p m ) = log[f (w 1,..., w n p 1,..., p m )] n m = i=1[ j=1 log(p j ) 1(W i = j)] m = j=1 x j log(p j ) Maximum Likelihood Estimate (MLE) of (p 1,..., p m ) maximizes lik(p 1,..., p m ) (with x 1,..., x m fixed!) Maximum achieved when differential is zero m Constraint: j=1 p j = 1 Apply method of Lagrange multipliers Solution: ˆp j = x j /n, j = 1,..., m. Note: if any x j = 0, then ˆp j = 0 solved as limit 23

24 MLEs of Multinomial Parameter Example A Hardy-Weinberg Equilibrium Equilibrium frequency of genotypes: AA, Aa, and aa P(a) = θ and P(A) = 1 θ Equilibrium probabilities of genotypes: (1 θ) 2, 2(θ)(1 θ), and θ 2. Multinomial Data: (X 1, X 2, X 3 ) corresponding to counts of AA, Aa, and aa in a sample of size n. Sample Data Genotype AA Aa aa Total Count X 1 X 2 X 3 n Frequency

25 Hardy-Weinberg Equilibrium Maximum-Likelihood Estimation of θ (X 1, X 2, X 3 ) Multinomial(n, p = ((1 θ) 2, 2θ(1 θ), θ 2 )) Log Likelihood for θ (θ) = log(f (x 1, x 2, x 3 p 1 (θ), p 2 (θ), p 3 (θ))) n! = log( p 1 (θ) x 1 p 2 (θ) x 2 p 3 (θ) x 3 x 1!x 2!x 3! ) = x 1 log((1 θ) 2 ) + x 2 log(2θ(1 θ)) +x 3 log(θ 2 ) + (non-θ terms) = (2x 1 + x 2 )log(1 θ) + (2x 3 + x 2 )log(θ) + (non-θ terms) First Differential of log likelihood: (2x 1 + x 2 ) (2x 3 + x 2 ) ' (θ) = + 1 θ θ ˆ 2x 3 + x 2 2x 3 + x 2 = θ = = = x 1 + 2x 2 + 2x 3 2n 25

26 Asymptotic variance of MLE θˆ: 1 Var(θˆ) E [ '' (θ)] Second Differential of log likelihood: d (2x 1 + x 2 ) (2x 3 + x 2 ) '' (θ) = [ + ] dθ 1 θ θ (2x 1 + x 2 ) (2x 3 + x 2 ) = (1 θ) 2 θ 2 Each of the X i are Binomial(n, p i (θ)) so E [X 1 ] = np 1 (θ) = n(1 θ) 2 E [X 2 ] = np 2 (θ) = n2θ(1 θ) E [X 3 ] = np 3 (θ) = nθ 2 E [ '' 2n (θ)] = θ(1 θ) a θˆ(1 θˆ).4247(1.4247) σˆ θˆ = = = n

27 Hardy-Weinberg Model Approximate 95% Confidence Interval for θ ˆ Interval : θ ± z(α/2) σˆˆ, with θ ˆθ = Interval : σˆθˆ = z(α/2) = 1.96 (with α = ) ± 1.96 (0.0109) = [0.4033,.4461] Note (!!) : Bootstrap simulation in R of θˆ the RMSE (θˆ) = is virtually equal to ˆσ θˆ 27

28 MIT OpenCourseWare Statistics for Applications Spring 2015 For information about citing these materials or our Terms of Use, visit: