Inference for fixed effects and variance components 1/16 Linear mixed models In general, a linear mixed model may be represented as Y = X β + Zu + ε, where I Y is an n × 1 vector of response; I X is an n × p design matrix; I β is a p × 1 vector of “fixed” unknown parameter values; I Z is an n × q model matrix of known constants; I u is a q × 1 random vector; I ε is an n × 1 random error. 2/16 Linear mixed models We typically assume that E(ε) = 0, Var(ε) = R, E(u) = 0, Var(u) = G. As a result, Var(Y ) = Σ(γ) = ZGZ T + R. 3/16 Generalized LSE I For any estimable function C T β, the unique BLUE is T β = C T (X T Σ−1 X )− X T Σ−1 Y . Cd I It can be shown that T β ∼ N(C T β, C T (X T Σ−1 X )− C). Cd I The inference for C T β can be built based on the asymptotic normality. 4/16 Example: random blocks (penicillin production) I Comparison of four processes for producing penicillin. I Four processes A, B, C and D, levels of a “fixed” effect treatment. I Random sample of five batches of raw material, corn steep liquor. I Split each batch into four parts. Run each process on one part, and randomize the order in which the processes are run with each batch. 5/16 Example: Penicillin production Recall that the model is Yij = µ + αi + γj + eij i = 1, · · · , a j = 1, · · · , b, where Yij is the yield for the i-th process applied to the j-th batch, γj ’s are independent and identically distributed (IID) random batch effects, eij ’s are IID random errors. Moreover, γj ’s are independent of eij ’s. 6/16 Example continued I Assume that we would like to estimate and construct confidence intervals for µ + αi . I An unbiased estimate for µ + αi is Ȳi· = b−1 Pb j=1 Yij (is this BLUE?). I It is easy to know that Var(Ȳi· ) = (σγ2 + σ 2 )/b. Based on the ANOVA estimates of the variance components, we have o 1n1 \ (MSB − MSE) + MSE Var( Ȳi· ) = b a 1 a−1 = MSB + MSE. ab ab 7/16 Example continued I A 1 − α approximate confidence interval for µ + αi is \ Ȳi· ± zα/2 {Var( Ȳi· )}1/2 . I In small sample case, one might wish to use t-quantile to \ replace standard normal quantile. Since Var( Ȳ ) is a linear i· combination of mean squares, it is not immediately clear what degree of freedom should be used. 8/16 Cochran-Satterthwaite (CS) approximation Suppose MS1 , MS2 , · · · , MSk are independent random variables such that dfi MSi ∼ χ2dfi for i = 1, · · · , k . E(MSi ) Consider a linear combination of MSi , which is S 2 = a1 MS1 + a2 MS2 + · · · + ak MSk . We would like to approximate the distribution of S 2 by a skewed chi-square distribution bχ2ν . 9/16 CS approximation The idea of CS approximation is to match the first two moments of bχ2ν with the first two moments of S 2 . Then solving these two equations to estimate the unknown parameters b and ν in the skewed chi-square distribution. 10/16 Moments of S 2 It is easy to see the following E(S 2 ) = k X ai E(MSi ); i=1 Var(S 2 ) = = k X i=1 k X Var(MSi )ai2 E 2 (MSi ) 2 a × (2dfi ) (dfi )2 i i=1 k X =2 i=1 ai2 E 2 (MSi ) . dfi 11/16 Moments of skewed chi-square bχ2ν The first two moments of bχ2ν are E(bχ2ν ) = bν; Var(bχ2ν ) = 2b2 ν. Matching with the moments of S 2 , we have the following two estimating equations: bν = k X ai E(MSi ); i=1 2 2b ν = 2 k X i=1 ai2 E 2 (MSi ) . dfi 12/16 CS approximation The solution of the above estimating equations are b= k X ai2 i=1 2 2 E 2 (MSi ) /E(S 2 ) dfi ν = E (S )/ k X ai2 i=1 E 2 (MSi ) . dfi Therefore, we can estimate b and ν by b̂ = k X i=1 ai2 MSi2 2 /S dfi ν̂ = (S 2 )2 / k X i=1 ai2 MSi2 . dfi 13/16 Example: penicillin production A t-type confidence interval for µ + αi is \ Ȳi· ± tν̂ {Var( Ȳi· )}1/2 . Here ν is approximate by the CS approximation. Namely, n 1 MSB 2 o a − 1 2 MSE 2 2 \ ν̂ = {Var( Ȳi· )}2 / + . ab b − 1 ab (a − 1)(b − 1) 14/16 Inference for variance components Suppose that a variance component parameter γ can be estimated by a linear combination of mean squares, say S 2 = a1 MS1 + a2 MS2 + · · · + ak MSk . Then an approximate 1 − α confidence interval for γ is ! ν̂S 2 ν̂S 2 , , χ2ν̂,α/2 χ2ν̂,1−α/2 where ν̂ is the approximated degree of freedom using the CS method. 15/16 Example: penicillin production Suppose we want to construct a 1 − α confidence interval for σγ2 . Recall that σ̂γ2 = 1 (MSB − MSE). a Then the estimated degree of freedom through the CS approximation is ν̂ = {σ̂γ2 }2 / n 1 MSB 2 o 1 MSE 2 + . a2 b − 1 a2 (a − 1)(b − 1) Thus a 1 − α confidence interval for σγ2 is ν̂ σ̂γ2 ν̂ σ̂γ2 , χ2ν̂,α/2 χ2ν̂,1−α/2 ! . 16/16