R/glmm_final.r

In glmmLasso: Variable Selection for Generalized Linear Mixed Models by L1-Penalized Estimation

glmm_final<-function(y,X,W,k,n,q_start,K,Delta_start,s,steps=1000,family,method,overdispersion,phi,
                     nue=1,print.iter.final=FALSE,flushit,eps.final=1e-5,Q.min=1e-13,Q.max=20,Q.fac=5)
{
  
  ## Print stuff.
  ia <- if(flushit) interactive() else FALSE
  
N<-length(y)
lin<-ncol(as.matrix(X))
Eta<-cbind(X,W)%*%Delta_start

if(is.null(family$multivariate)){
  D<-family$mu.eta(Eta)
  Mu<-family$linkinv(Eta)
  SigmaInv <- 1/family$variance(Mu)
}else{
  Mu <- family$linkinv(Eta, K)
  D <- family$deriv.mat(Eta, K)
  SigmaInv <- family$SigmaInv(Mu, K)
}

if(print.iter.final)
  #     message()
{
  cat(if(ia) "\r" else NULL)
  cat("\nFinal Re-estimation Iteration  1")
  if(.Platform$OS.type != "unix" & ia) flush.console()
}

Z_alles<-cbind(X,W)

if(s==1)
{
P1<-c(rep(0,lin),rep(1/q_start,n*s))
P1<-diag(P1)
}else{
P1<-matrix(0,lin+n*s,lin+n*s)
for(jf in 1:n)
P1[(lin+(jf-1)*s+1):(lin+jf*s),(lin+(jf-1)*s+1):(lin+jf*s)]<-chol2inv(chol(q_start))
}

Delta<-matrix(0,steps,(lin+s*n))
Eta.ma<-matrix(0,steps+1,N)
Eta.ma[1,]<-Eta

Q<-list()
Q[[1]]<-q_start

l=1
opt<-steps

if(is.null(family$multivariate)){
  D <- drop(D);SigmaInv <- drop(SigmaInv)
  score_vec<-t(Z_alles)%*%((y-Mu)*D*SigmaInv)-P1%*%Delta[1,]
  F_gross<-t(Z_alles)%*%(Z_alles*D*SigmaInv*D)+P1
}else{
  score_vec<-RcppEigenProd1(Z_alles, D, SigmaInv, y, Mu)-P1%*%Delta[1,]
  W_opt <- RcppEigenProd2(D, SigmaInv)
  F_gross <- t(Z_alles)%*%(W_opt%*%Z_alles)+P1
}

InvFisher<-try(chol2inv(chol(F_gross)),silent=TRUE)
if(inherits(InvFisher, "try-error"))
InvFisher<-solve(F_gross)  

half.index<-0
solve.test<-FALSE
Delta_r<-InvFisher%*%score_vec

P1.old<-P1
######### big while loop for testing if the update leads to Fisher matrix which can be inverted
while(!solve.test)
{  

solve.test2<-FALSE  
while(!solve.test2)
{  
Delta[1,]<-Delta_start+nue*(0.5^half.index)*Delta_r

Eta<-Z_alles%*%Delta[1,]

if(is.null(family$multivariate)){
  D<-family$mu.eta(Eta)
  Mu<-family$linkinv(Eta)
  SigmaInv <- 1/family$variance(Mu)
}else{
  Mu <- family$linkinv(Eta, K)
  D <- family$deriv.mat(Eta, K)
  SigmaInv <- family$SigmaInv(Mu, K)
}

if (method=="EM")
{  
  if(is.null(family$multivariate)){
    D <- drop(D);SigmaInv <- drop(SigmaInv)
    F_gross<-t(Z_alles)%*%(Z_alles*D*SigmaInv*D)+P1.old
  }else{
    W_opt <- RcppEigenProd2(D, SigmaInv)
    F_gross <- t(Z_alles)%*%(W_opt%*%Z_alles)+P1.old
  }
    InvFisher<-try(chol2inv(chol(F_gross)),silent=TRUE)
  if(inherits(InvFisher, "try-error"))
  InvFisher<-try(solve(F_gross),silent=TRUE)  
  if(inherits(InvFisher, "try-error"))
  {
    half.index<-half.index+1  
  }else{
    solve.test2<-TRUE 
}}else{
  solve.test2<-TRUE
}}

if (method=="EM")
{
############################# Q updaten ################
Q1<-InvFisher[(lin+1):(lin+s),(lin+1):(lin+s)]+Delta[1,(lin+1):(lin+s)]%*%t(Delta[1,(lin+1):(lin+s)])
for (i in 2:n)
Q1<-Q1+InvFisher[(lin+(i-1)*s+1):(lin+i*s),(lin+(i-1)*s+1):(lin+i*s)]+Delta[1,(lin+(i-1)*s+1):(lin+i*s)]%*%t(Delta[1,(lin+(i-1)*s+1):(lin+i*s)])
Q1<-1/n*Q1
}else{
  if(is.null(family$multivariate)){
    Eta_tilde<-Eta+(y-Mu)/D
  }else{
    Eta_tilde<-Eta+solve(D)%*%(y-Mu)
  }

Betadach<-Delta[1,1:lin]

if(s==1)
{
low <- (1/Q.fac)*Q.min
upp <- Q.fac*Q.max
optim.obj<-nlminb(sqrt(q_start),likelihood_nlminb,D=D,SigmaInv=SigmaInv,X=X,X_aktuell=X,
                  Eta_tilde=Eta_tilde,n=n,Betadach=Betadach,W=W,
                  lower = low, upper=upp)  
Q1<-as.matrix(optim.obj$par)^2
}else{
q_start_vec<-c(diag(q_start),q_start[lower.tri(q_start)])
up1<-Q.fac*Q.max
upp<-rep(up1,length(q_start_vec))
low<-c(rep(0,s),rep(-up1,0.5*(s^2-s)))
kkk_vec<-c(rep(-1,s),rep(0.5,0.5*(s^2-s)))
optim.obj<-bobyqa(q_start_vec,likelihood,D=D,SigmaInv=SigmaInv,X=X,X_aktuell=X,Eta_tilde=Eta_tilde,Betadach=Betadach,W=W,n=n,s=s,k=k,lower=low,upper=upp)
Q1<-matrix(0,s,s)
Q1[lower.tri(Q1)]<-optim.obj$par[(s+1):(s*(s+1)*0.5)]
Q1<-Q1+t(Q1)
diag(Q1)<-(optim.obj$par[1:s])

#### Check for positive definitness ########
      for (ttt in 0:100)
      {
      Q1[lower.tri(Q1)]<-((0.5)^ttt)*Q1[lower.tri(Q1)]
      Q1[upper.tri(Q1)]<-((0.5)^ttt)*Q1[upper.tri(Q1)]
      Q_solvetest<-try(solve(Q1))
         if(all (eigen(Q1)$values>0) & !inherits(Q_solvetest, "try-error"))
         break
      }
}}

Q[[2]]<-Q1

if(s==1)
{
  P1<-c(rep(0,lin),rep(1/Q1,n*s))
  P1<-diag(P1)
}else{
  P1<-matrix(0,lin+n*s,lin+n*s)
  for(jf in 1:n)
    P1[(lin+(jf-1)*s+1):(lin+jf*s),(lin+(jf-1)*s+1):(lin+jf*s)]<-chol2inv(chol(Q1))
}

if(is.null(family$multivariate)){
  D <- drop(D);SigmaInv <- drop(SigmaInv)
  score_vec<-t(Z_alles)%*%((y-Mu)*D*SigmaInv)-P1%*%Delta[1,]
  F_gross<-t(Z_alles)%*%(Z_alles*D*SigmaInv*D)+P1
}else{
  score_vec<-RcppEigenProd1(Z_alles, D, SigmaInv, y, Mu)-P1%*%Delta[1,]
  W_opt <- RcppEigenProd2(D, SigmaInv)
  F_gross <- t(Z_alles)%*%(W_opt%*%Z_alles)+P1
}

InvFisher<-try(chol2inv(chol(F_gross)),silent=TRUE)

  if(inherits(InvFisher, "try-error"))
  InvFisher<-try(solve(F_gross),silent=TRUE)  

  if(inherits(InvFisher, "try-error"))
  {
    half.index<-half.index+1  
  }else{
    solve.test<-TRUE 
  }
}

Eta.ma[2,]<-Eta
P1.old.temp<-P1.old

###############################################################################################################################################
################################################################### Main Iterations ###################################################################
eps<-eps.final*sqrt(length(Delta_r))

for (l in 2:steps)
{
  
  if(print.iter.final)
    #  message("Iteration ",l)
  {
    cat(if(ia) "\r" else if(l > 1) "\n" else NULL)
    cat(paste("Final Re-estimation Iteration ",l))
    if(.Platform$OS.type != "unix" & ia) flush.console()
  }
  
half.index<-0
solve.test<-FALSE

P1.old<-P1

Delta_r<-InvFisher%*%score_vec
######### big while loop for testing if the update leads to Fisher matrix which can be inverted
first.time<-FALSE
while(!solve.test)
{  

solve.test2<-FALSE  
while(!solve.test2)
{  

if(half.index>50)
{
half.index<-Inf;P1.old<-P1.old.temp
}
Delta[l,]<-Delta[l-1,]+nue*(0.5^half.index)*Delta_r

Eta<-Z_alles%*%Delta[l,]

if(is.null(family$multivariate)){
  D<-family$mu.eta(Eta)
  Mu<-family$linkinv(Eta)
  SigmaInv <- 1/family$variance(Mu)
}else{
  Mu <- family$linkinv(Eta, K)
  D <- family$deriv.mat(Eta, K)
  SigmaInv <- family$SigmaInv(Mu, K)
}

if(method=="EM")
{  
  if(is.null(family$multivariate)){
    D <- drop(D);SigmaInv <- drop(SigmaInv)
    F_gross<-t(Z_alles)%*%(Z_alles*D*SigmaInv*D)+P1.old
  }else{
    W_opt <- RcppEigenProd2(D, SigmaInv)
    F_gross <- t(Z_alles)%*%(W_opt%*%Z_alles)+P1.old
  }
  InvFisher<-try(chol2inv(chol(F_gross)),silent=TRUE)
  if(inherits(InvFisher, "try-error"))
    InvFisher<-try(solve(F_gross),silent=TRUE)  
  if(inherits(InvFisher, "try-error"))
  {
    half.index<-half.index+1  
  }else{
    if(!first.time)
    half.index.final<-half.index
    solve.test2<-TRUE 
    first.time<-TRUE
  }}else{
    if(!first.time)
    half.index.final<-half.index
    solve.test2<-TRUE 
    first.time<-TRUE
  }}

if (method=="EM")
{
############################# Q update ################
Q1<-InvFisher[(lin+1):(lin+s),(lin+1):(lin+s)]+Delta[l,(lin+1):(lin++s)]%*%t(Delta[l,(lin+1):(lin+s)])
for (i in 2:n)
Q1<-Q1+InvFisher[(lin+(i-1)*s+1):(lin+i*s),(lin+(i-1)*s+1):(lin+i*s)]+Delta[l,(lin+(i-1)*s+1):(lin+i*s)]%*%t(Delta[l,(lin+(i-1)*s+1):(lin+i*s)])

Q1<-1/n*Q1
}else{
  if(is.null(family$multivariate)){
    Eta_tilde<-Eta+(y-Mu)/D
  }else{
    Eta_tilde<-Eta+solve(D)%*%(y-Mu)
  }
  
Betadach<-Delta[l,1:lin]

if(s==1)
{
upp<-max(upp,Q.fac*sqrt(Q1))
low<-min(low,(1/Q.fac)*sqrt(Q1))
optim.obj<-nlminb(sqrt(Q1),likelihood_nlminb,D=D,SigmaInv=SigmaInv,X=X,X_aktuell=X,
                  Eta_tilde=Eta_tilde,n=n,Betadach=Betadach,W=W, 
                  lower = low, upper = upp)
Q1<-as.matrix(optim.obj$par)^2
}else{
up1<-max(up1,Q.fac*max(Q1))  
upp<-rep(up1,length(q_start_vec))
low<-c(rep(0,s),rep(-up1,0.5*(s^2-s)))
Q1_vec<-c(diag(Q1),Q1[lower.tri(Q1)])
optim.obj<-bobyqa(Q1_vec,likelihood,D=D,SigmaInv=SigmaInv,X=X,X_aktuell=X,Eta_tilde=Eta_tilde,n=n,s=s,k=k,Betadach=Betadach,W=W,lower=low,upper=upp)

Q1<-matrix(0,s,s)
Q1[lower.tri(Q1)]<-optim.obj$par[(s+1):(s*(s+1)*0.5)]
Q1<-Q1+t(Q1)
diag(Q1)<-(optim.obj$par[1:s])

#### Check for positiv definitness ########
for (ttt in 0:100)
      {
      Q1[lower.tri(Q1)]<-((0.5)^ttt)*Q1[lower.tri(Q1)]
      Q1[upper.tri(Q1)]<-((0.5)^ttt)*Q1[upper.tri(Q1)]
       Q_solvetest<-try(solve(Q1))
         if(all (eigen(Q1)$values>0) & !inherits(Q_solvetest, "try-error"))
         break
      }
}}

Q[[l+1]]<-Q1

if(s==1)
{
  P1<-c(rep(0,lin),rep(1/Q1,n*s))
  P1<-diag(P1)
}else{
  P1<-matrix(0,lin+n*s,lin+n*s)
  for(jf in 1:n)
    P1[(lin+(jf-1)*s+1):(lin+jf*s),(lin+(jf-1)*s+1):(lin+jf*s)]<-chol2inv(chol(Q1))
}

if(is.null(family$multivariate)){
  D <- drop(D);SigmaInv <- drop(SigmaInv)
  score_vec<-t(Z_alles)%*%((y-Mu)*D*SigmaInv)-P1%*%Delta[l,]
  F_gross<-t(Z_alles)%*%(Z_alles*D*SigmaInv*D)+P1
}else{
  score_vec<-RcppEigenProd1(Z_alles, D, SigmaInv, y, Mu)-P1%*%Delta[l,]
  W_opt <- RcppEigenProd2(D, SigmaInv)
  F_gross <- t(Z_alles)%*%(W_opt%*%Z_alles)+P1
}

InvFisher<-try(chol2inv(chol(F_gross)),silent=TRUE)
  if(inherits(InvFisher, "try-error"))
  InvFisher<-try(solve(F_gross),silent=TRUE)  
  if(inherits(InvFisher, "try-error"))
  {
    half.index<-half.index+1  
  }else{
    solve.test<-TRUE 
  }
}

Eta.ma[l+1,]<-Eta
P1.old.temp<-P1.old

finish<-(sqrt(sum((Eta.ma[l,]-Eta.ma[l+1,])^2))/sqrt(sum((Eta.ma[l,])^2))<eps)
finish2<-(sqrt(sum((Eta.ma[l-1,]-Eta.ma[l+1,])^2))/sqrt(sum((Eta.ma[l-1,])^2))<eps)

if(finish ||  finish2) 
  break
  
}

opt<-l

if(is.null(family$multivariate)){
  W_opt <- D*SigmaInv*D
  FinalHat<-(Z_alles*sqrt(W_opt))%*%InvFisher%*%t(Z_alles*sqrt(W_opt))
}else{
  W_inv_t <- RcppEigenSpChol(W_opt)
  FinalHat <- RcppEigenProd3(W_inv_t, Z_alles, InvFisher)
}

complexity<-sum(diag(FinalHat))

if(overdispersion)
  phi<-(sum((y-Mu)^2/family$variance(Mu)))/(N-complexity)

Deltafinal<-Delta[l,]
Q_final<-Q[[l+1]]
Standard_errors<-InvFisher

## compute ranef part of loglik
if(s==1)
{
  P1.ran<-rep(1/Q_final,n*s)
  P1.ran<-diag(P1.ran)
}else{
  P1.ran<-matrix(0,n*s,n*s)
  for(jf in 1:n)
    P1.ran[((jf-1)*s+1):(jf*s),((jf-1)*s+1):(jf*s)]<-chol2inv(chol(Q_final))
}

ranef.logLik<- -0.5*t(Deltafinal[(lin+1):(lin+n*s)])%*%P1.ran%*%Deltafinal[(lin+1):(lin+n*s)]

ret.obj<-list()
ret.obj$ranef.logLik<-ranef.logLik
ret.obj$opt<-opt
ret.obj$Delta<-Deltafinal
ret.obj$Q<-Q_final
ret.obj$Standard_errors<-Standard_errors
ret.obj$phi<-phi
ret.obj$complexity<-complexity
return(ret.obj)
}