View PDF Version - RePub - Erasmus Universiteit Rotterdam
View PDF Version - RePub - Erasmus Universiteit Rotterdam
View PDF Version - RePub - Erasmus Universiteit Rotterdam
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Discriminant analysis using a MLMM with a normal mixture 145<br />
groups. The methodology is illustrated on the Dutch PBC Study data in Section<br />
’Application to PBC data’. We have also extended the R (R Development Core<br />
Team13 ) package mixAK (Komárek14 ) to apply the proposed methods in practice.<br />
The use of the package is briefly explained in the Appendix.<br />
A multivariate linear mixed model with normal mixture<br />
in the random effects distribution<br />
In the multivariate linear mixed model (MLMM, Morrell et al. 7 ), a standard linear<br />
mixed model is first specified for the r-th marker. That is,<br />
Y i,r = Xi,rαr + Zi,rbi,r + εi,r (i =1,...,N, r =1,...,R), (1)<br />
where Xi,r is a ni,r×pr covariate matrix for fixed effects and Zi,r is a ni,r×qr covariate<br />
matrix for random effects in a model for marker r. Further, αr =(αr,1,...,αr,pr ) ′<br />
isavectoroffixedeffectsformarkerr, andbi,r =(bi,r,1,...,bi,r,qr ) ′ is a vector of<br />
random effects for marker r specific for the i-th subject. For computational convenience<br />
of the approach outlined in Section ’Estimation’, a hierarchically centered<br />
parametrization of the LMM will be used here, i.e. E(bi,r)=βr =(βr,1,...,βr,qr ) ′<br />
with matrices (Xi,r, Zi,r) of full column rank. That is, the vector (α ′ r , β ′ r ) ′ is<br />
a vector of fixed effects in a classical sense. In the remainder of the paper we let<br />
α =(α ′ 1 ,...,α′ R )′ and β =(β ′ 1,...,β ′ R) ′ be the vectors of fixed effects and means<br />
of random effects for all considered markers, respectively. Further, let p = �R r=1 pr<br />
be the length of the vector α and q = �R r=1 qr be the length of the vector β also<br />
equal to the total dimension of random effects. Finally, εi,r =(εi,r,1,...,εi,r,ni,r )′<br />
is the vector of random errors for the measurements of the r-thmarkeronthe<br />
i-th subject. The errors are assumed to be mutually independent and normally distributed.<br />
However, we allow the residual variances corresponding to different markers<br />
to differ. Hence, for εi =(ε ′ i,1 ,...,ε′ i,R )′ we assume εi ∼N(0, Σi), where Σi is<br />
a block-diagonal matrix with each diagonal block being equal to σ 2 r Ini,r , where σ2 r<br />
is the residual variance of the r-th marker. Note that the MLMM (1) can now be<br />
written as a standard LMM as<br />
where Xi is a ni ×p block-diagonal matrix with matrices Xi,1,...,Xi,R on the diagonal<br />
and similarly Zi is a ni × q block-diagonal matrix with matrices Zi,1,...,Zi,R on the<br />
diagonal.<br />
Y i = Xiα + Zibi + εi (i =1,...,N), (2)