This preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: Again returning to the GLMM with random intercept, an importance sampler can be constructed by using φ ( u i ; μ i , τ 2 i ), the normal density approximation to the posterior density f ( u i  y i ). That is, the representation from p.196 of the integral we’re after, E { f ( y i  b i ) } = Z + ∞∞ φ ( u i ; μ i , τ 2 i ) ( φ ( u i ) Q j f ( y ij  √ ψu i ) φ ( u i ; μ i , τ 2 i ) ) du i suggests drawing a sample ( u * 1 , . . . , u * R ) from φ ( u i ; μ i , τ 2 i ) and using the approximation E { f ( y i  b i ) } ≈ 1 R R X r ( φ ( u * r ) Q j f ( y ij  √ ψu * r ) φ ( u * r ; μ i , τ 2 i ) ) • This method is akin to AGQ, which can be viewed as a determinis tic version of importance sampling. Both methods require an extra computation of μ i , τ 2 i , the mean and variance of the normal approx imation to the posterior distribution. • Importance sampling is implemented in PROC NLMIXED. 201 3. Analytic Approximation of the Likelihood: • Most of the methods that fall into this category are based upon taking a Laplace approximation to the integral involved in the GLMM loglikelihood. For a unidimensional integral, the Laplace approximation can be written as Z + ∞∞ exp { f ( x ) } dx ≈ Z + ∞∞ exp { f (˜ x ) ( x ˜ x ) 2 / (2 σ 2 ) } dx = Z + ∞∞ exp { f (˜ x ) } √ 2 πσφ ( x ; ˜ x, σ 2 ) dx = exp { f (˜ x ) } √ 2 πσ, where φ ( x ; ˜ x, σ 2 ) is a normal density with mean ˜ x and variance σ 2 , ˜ x is the model of f ( x ) and hence of exp { f ( x ) } , and σ 2 = ∂ 2 f ( x ) ∂x 2 fl fl fl fl x =˜ x ¶ 1 . • The approximation in the first line here is obtained by approximating f ( x ) by a second order Taylor expansion around its mode. You’ll notice that the first derivative term f (˜ x ) drops out here because it is evaluated at the mode, at which point f ( x ) = 0. 202 For the GLMM with random intercept, the integrand in the likeli hood contribution for the i th cluster (which plays the role of exp { f ( x ) } ) is φ ( b i ; 0 , ψ ) Y j f ( y ij  b i ) = exp log φ ( b i ; 0 , ψ ) Y j f ( y ij  b i ) • As mentioned before, this quantity is proportional to the posterior distribution f ( b i  y i ). Therefore, in the Laplace approximation, the quantity playing the role of ˜ x is the posterior mode ˜ b i = arg max b i φ ( b i ; 0 , ψ ) Y j f ( y ij  b i ) and the curvature (inverse negative Hessian) of the integrand, σ 2 i , plays the role of σ 2 . Thus, the Laplace approximation to the loglikelihood contribution from the i th subject becomes log f ( y i ; β , ψ ) ≈ log( √ 2 πσ i ) + log { φ ( ˜ b i ; 0 , ψ ) } + X j log f ( y ij  ˜ b i ) = log( σ i / p ψ ) ˜ b 2 i / (2 ψ ) + X j log f ( y ij  ˜ b i ) . ( ** ) • This approximation is good whenever the posterior density of b i is approximately normal. This occurs for large cluster sizes....
View
Full Document
 Fall '08
 Staff
 Conditional Probability, Laplace, Laplace approximation

Click to edit the document details