This script requires that the working directory includes the folders “results” and “manuscript”. Consider uncommenting the lines for saving results and figures. For each simulation and application, the first chunk performs the computationally intensive analysis, and the other chunks summarise the results. We pre-processed the TCGA data with R 3.6.3 (2020-02-29) on a local machine (x86_64-w64-mingw32/x64, Windows 10 x64), and analysed the data with R 3.6.1 (2019-07-05) on a virtual machine (x86_64-pc-linux-gnu, Ubuntu 16.04.6 LTS).

Introduction

#<<start>>
ellipse <- function(x,y,type=TRUE,text=NULL,a=0.5,b=0.5,border="black",col=NULL,txt.col="black",...){
    n <- max(c(length(x),length(y)))
    if(is.null(col)){col <- rep(grey(0.9),times=n)}
    if(length(col)==1){col <- rep(col,times=n)}
    if(length(x)==1){x <- rep(x,times=n)}
    if(length(y)==1){y <- rep(y,times=n)}
    if(length(text)==1){text <- rep(text,times=n)}
    if(length(border)==1){border <- rep(border,times=n)}
    for(i in seq_len(n)){
        if(type){
          angle <- seq(from=0,to=2*pi,length=100)
          xs <- x[i] + a * cos(angle)
          ys <- y[i] + b * sin(angle)
          graphics::polygon(x=xs,y=ys,col=col[i],border=border[i])
        } else {
          graphics::polygon(x=x[i]+c(-a,-a,a,a),y=y[i]+c(-b,b,b,-b),border=border[i],col=col[i])
        }
        graphics::text(labels=text[i],x=x[i],y=y[i],col=txt.col,...)
    }
}

txt <- list()
txt$y <- expression(hat(y))
txt$omega <- eval(parse(text=paste0("expression(",paste0("hat(omega)[",c(1:3,"k","m"),"]",collapse=","),")")))
txt$alpha <- eval(parse(text=paste0("expression(",paste0("hat(y)*\"|\"*alpha[",c(1:3,"k","m"),"]",collapse=","),")")))
txt$beta <- eval(parse(text=paste0("expression(",paste0("hat(beta)[",c(1:3,"j","p"),"]*\"|\"*alpha[k]",collapse=","),")")))
txt$x <- eval(parse(text=paste0("expression(",paste0("x[\"",c(1:3,"j","p"),"\"]",collapse=","),")")))
txt$dots <- expression(cdots)

pos <- list()
pos$y <- 4
pos$alpha <- c(1,2,3,5,7)
pos$x <- c(0.5,1.5,2.5,5.5,7.5)
pos$omega <- pos$y+(pos$alpha-pos$y)/2
pos$beta <- 5+(pos$x-5)/2
pos$beta[1] <- pos$beta[1] - 0.3
pos$beta[3] <- pos$beta[3] + 0.3
 
a <- b <- 0.3
#grDevices::pdf(file="manuscript/figure_NET.pdf",width=5,height=3)
graphics::plot.new()
graphics::par(mar=c(0,0,0,0),mfrow=c(1,1))
graphics::plot.window(xlim=c(0.4,7.6),ylim=c(0.8,5.2))

# omega
segments(x0=4,y0=5-a,x1=pos$alpha,y1=3+a,lwd=2,col=red)
ellipse(x=pos$omega,y=4,text=txt$omega,a=0.2,b=0.21,cex=1.2,col="white",border=red,txt.col=red,type=FALSE)

# beta
segments(x0=rep(pos$alpha,each=length(pos$x)),y0=3-a,x1=rep(pos$x,times=length(pos$alpha)),y1=1+a,lwd=2,col="grey")
segments(x0=pos$x,y0=1+a,y1=3-a,x1=5,lwd=2,col=blue)
ellipse(x=pos$beta,y=2,text=txt$beta,a=0.35,b=0.27,cex=1.2,col="white",border=blue,txt.col=blue,type=FALSE)

# x and y
ellipse(x=pos$x,y=1,text=txt$x,a=a,b=b,cex=1.2)
text(x=c(4,6.5),y=1,labels=txt$dots,cex=1.2)
ellipse(x=pos$alpha,y=3,text=txt$alpha,a=a+0.1,b=b,cex=1.2)
text(x=c(4,6),y=3,labels=txt$dots,cex=1.2)
ellipse(x=pos$y,y=5,text=txt$y,a=a,b=b,cex=1.2)
#grDevices::dev.off()

Simulation 1

#<<start>>
n <- 10000; p <- 500
fold <- rep(c(0,1),times=c(100,n-100))
mode <- rep(c("sparse","dense","mixed"),times=100)
family <- "gaussian"

for(id in c(1,NA,0)){
  loss <- list()
  for(i in seq_along(mode)){
    set.seed(i)
    cat(round(100*i/length(mode),digits=2)," ")
    data <- .simulate.block(n=n,p=p,mode=mode[i],family=family)
    nzero <- seq_len(10)*switch(mode[i],sparse=1,dense=10,mixed=5,stop())
    set.seed(i)
    loss[[i]] <- tryCatch(cv.starnet(y=data$y,X=data$X,family=family,foldid.ext=fold,alpha.meta=id,nzero=nzero),error=function(x) NULL)
  }
  #save(loss,mode,file=paste0("results/sim_prediction_",id,".RData"))
}

#writeLines(text=capture.output(utils::sessionInfo(),cat("\n"),
#        sessioninfo::session_info()),con="results/sim_prediction.txt")
# <<start>>
load("results/sim_prediction_1.RData")
cond <- sapply(loss,is.null)
loss <- loss[!cond]; mode <- mode[!cond]

#grDevices::pdf(file="manuscript/figure_BOX.pdf",width=5,height=2)
graphics::par(mfrow=c(1,3),mar=c(2.1,2,0.5,0.5),oma=c(0,2,0,0))
meta <- sapply(loss,function(x) x$meta)
base <- sapply(loss,function(x) x$base[c("alpha0.05","alpha0.95")])
table <- t(rbind(base,meta))

names <- c("ridge","lasso","tune","stack")
col <- ifelse(grepl(pattern="ridge|alpha|lasso",x=names),red,blue)

modes <- c("sparse","dense","mixed")
median <- winner <- pvalue <- matrix(data=NA,nrow=length(modes),ncol=length(names),
                          dimnames=list(modes,names))

for(i in modes){
  values <- table[mode==i,names]

  # information
  median[i,] <- apply(values,2,median)
  winner[i,] <- rowMeans(apply(values,1,rank)<2)
  pvalue[i,] <- apply(values,2,function(x) suppressWarnings(wilcox.test(x=values[,"stack"],y=x,paired=TRUE,alternative="two.sided")$p.value))
  
  # boxes
  graphics::plot.new()
  graphics::plot.window(xlim=c(0.4,4.6),ylim=range(values))
  y <- apply(values,2,function(x) stats::quantile(x,p=c(0.05,0.25,0.50,0.75,0.95)))
  
  # whiskers
  graphics::segments(x0=seq_len(ncol(values)),y0=y["5%",],y1=y["95%",])
  
  # boxes
  graphics::boxplot(values,cex.axis=1,col=col,medcol="white",names=rep("",times=ncol(values)),whiskcol=NA,staplecol=NA,outcol=NA,add=TRUE)
  
  # whiskers
  graphics::segments(x0=seq_len(ncol(values))-0.15,x1=seq_len(ncol(values))+0.15,y0=y["5%",])
  graphics::segments(x0=seq_len(ncol(values))-0.15,x1=seq_len(ncol(values))+0.15,y0=y["95%",])
  
  for(j in seq_len(4)){
    cond <- values[,j]>y["95%",j] | values[,j]<y["5%",j]
    points(x=rep(j,times=sum(cond)),y=values[cond,j],cex=0.7)
  }
  
  even <- seq(from=1,to=ncol(values),by=2)
  odd <- seq(from=2,to=ncol(values),by=2)
  graphics::axis(side=1,labels=colnames(values)[even],at=even)
  graphics::axis(side=1,labels=colnames(values)[odd],at=odd)
  
  graphics::points(x=1,y=median(table[mode==i,"alpha0.05"]),pch=21,col="black",bg=red,cex=1.2)
  graphics::points(x=2,y=median(table[mode==i,"alpha0.95"]),pch=21,col="black",bg=red,cex=1.2)
  
}
graphics::title(ylab=paste0(paste0(rep(" ",times=12),collapse=""),"mean squared error"),outer=TRUE,line=1)
#grDevices::dev.off()

round(median,digits=2)
round(winner,digits=2)
signif(9*pvalue,digits=2)
#<<start>>
load("results/sim_prediction_1.RData")
cond <- sapply(loss,is.null)
loss <- loss[!cond]; mode <- mode[!cond]

#grDevices::pdf(file="manuscript/figure_PHS.pdf",width=5,height=2)
graphics::par(mfrow=c(1,3),mar=c(3.3,2,0.5,0.5),oma=c(0,2,0,0))
for(i in c("sparse","dense","mixed")){
  x <- as.numeric(colnames(loss[mode==i][[1]]$extra))
  names <- list(rownames(loss[[1]]$extra),x,seq_len(sum(mode==i)))
  X <- array(unlist(lapply(loss[mode==i],function(x) x$extra)),dim=sapply(names,length),dimnames=names)
  X <- apply(X,c(1,2),median)
  lasso <- median(sapply(loss[mode==i],function(x) x$meta["lasso"]))
  stack <- median(sapply(loss[mode==i],function(x) x$meta["stack"]))
  graphics::plot.new()
  graphics::plot.window(xlim=range(x,na.rm=TRUE),ylim=range(c(lasso,stack,X)))
  graphics::box()
  graphics::axis(side=1)
  graphics::axis(side=2)
  graphics::abline(h=lasso,col="grey",lty=2)
  graphics::abline(h=stack,lty=2)
  graphics::lines(y=X["lasso",],x=x,col="grey")
  graphics::points(y=X["lasso",],x=x,pch=21,col="grey",bg="white")
  graphics::lines(y=X["stack",],x=x)
  graphics::points(y=X["stack",],x=x,pch=21,bg="white")
  graphics::title(xlab="nzero",line=2.5)
}
graphics::title(ylab=paste0(paste0(rep(" ",times=12),collapse=""),"mean squared error"),outer=TRUE,line=1)
#grDevices::dev.off()

Simulation 2

#<<start>>
n <- 10000; p <- 500
fold <- rep(c(0,1),times=c(100,n-100))

family <- "gaussian"
mode <- rep(c("sparse","dense","mixed"),times=100)

for(id in c(1,NA,0)){
  loss <- list()
  mae0 <- mae1 <- mse0 <- mse1 <- sapply(c("lasso","ridge","tune","stack"),function(x) numeric())
  sel0 <- sel1 <- TP <- FP <- TN <- FN <- sapply(c("lasso","enet","stack"),function(x) numeric())
  graphics::par(mfrow=c(1,3),mar=c(2,2,0,0))
  for(i in seq_along(mode)){
    cat(round(100*i/length(mode),digits=2)," ")
    set.seed(i)
    data <- .simulate.mode(n=n,p=p,mode=mode[i])
    
    #--- prediction ---
    nzero <- seq_len(10)*switch(mode[i],sparse=1,dense=10,mixed=5,stop("Invalid mode."))
    set.seed(i)
    loss[[i]] <- tryCatch(cv.starnet(y=data$y,X=data$X,family=family,alpha.meta=id,foldid.ext=fold,nzero=nzero),error=function(x) NULL)
    tryCatch(graphics::title(main=paste0("mode=",mode[i])),error=function(x) NULL)
    
    #--- estimation ---
    
    set.seed(i)
    model <- starnet(y=data$y[fold==0],X=data$X[fold==0,],family=family,alpha.meta=id)

    # unrestricted model
    coef <- list()
    coef$ridge <- coef(model$base$alpha0$glmnet.fit,
                           s=model$base$alpha0$lambda.min)[-1]
    coef$lasso <- coef(model$base$alpha1$glmnet.fit,
                           s=model$base$alpha1$lambda.min)[-1]
    coef$stack <- coef(model)$beta
    select <- which.min(sapply(model$base,function(x) x$cvm[x$id.min]))
    coef$tune <- coef(model$base[[select]]$glmnet.fit,
                          s=model$base[[select]]$lambda.min)[-1]
    
    for(k in names(coef)){
      # mean absolute error
      mae0[[k]][i] <- mean(abs(data$beta[data$beta==0]-coef[[k]][data$beta==0]))
      mae1[[k]][i] <- mean(abs(data$beta[data$beta!=0]-coef[[k]][data$beta!=0]))
      # mean squared error
      mse0[[k]][i] <- mean((data$beta[data$beta==0]-coef[[k]][data$beta==0])^2)
      mse1[[k]][i] <- mean((data$beta[data$beta!=0]-coef[[k]][data$beta!=0])^2)
    }
    
    # restricted model
    coef <- list()
    coef$stack <- coef(model,nzero=10)$beta
    lasso <- .cv.glmnet(y=data$y[fold==0],x=data$X[fold==0,],alpha=1,family=family,foldid=model$info$foldid,nzero=10)
    coef$lasso <- stats::coef(lasso,s="lambda.min")[-1]
    enet <- .cv.glmnet(y=data$y[fold==0],x=data$X[fold==0,],alpha=0.95,family=family,foldid=model$info$foldid,nzero=10)
    coef$enet <- stats::coef(enet,s="lambda.min")[-1]
    
    for(k in names(coef)){
      sel0[[k]][i] <- sum(coef[[k]][data$beta==0]!=0)
      sel1[[k]][i] <- sum(coef[[k]][data$beta!=0]!=0)
       # continue here
      TP[[k]][i] <- sum(coef[[k]][data$beta!=0]!=0)
      FP[[k]][i] <- sum(coef[[k]][data$beta==0]!=0)
      TN[[k]][i] <- sum(coef[[k]][data$beta==0]==0)
      FN[[k]][i] <- sum(coef[[k]][data$beta!=0]==0)
    }
    
  }
  #save(mode,mae0,mae1,mse0,mse1,sel0,sel1,TP,FP,TN,FN,loss,file=paste0("results/sim_estimation_",id,".RData"))
}

#writeLines(text=capture.output(utils::sessionInfo(),cat("\n"),
#        sessioninfo::session_info()),con="results/sim_estimation.txt")
#<<start>>
load("results/sim_estimation_1.RData")

# estimation accuracy (mean absolute error)
round(tapply(X=100*(mae0$stack-mae0$tune)/mae0$tune,INDEX=mode,FUN=median),digits=1)
round(tapply(X=100*(mae1$stack-mae1$tune)/mae1$tune,INDEX=mode,FUN=mean),digits=1)
signif(3*tapply(X=mae1$stack-mae1$tune,INDEX=mode,FUN=function(x) stats::wilcox.test(x)$p.value),digits=2)

# estimation accuracy (mean squared error)
round(tapply(X=100*(mse0$stack-mse0$tune)/mae0$tune,INDEX=mode,FUN=median),digits=1)
round(tapply(X=100*(mse1$stack-mse1$tune)/mae1$tune,INDEX=mode,FUN=mean),digits=1)
signif(3*tapply(X=mse1$stack-mse1$tune,INDEX=mode,FUN=function(x) stats::wilcox.test(x)$p.value),digits=2)

# selection accuracy (true/false positives)
round(mean(TP$stack),digits=1); round(mean(TP$lasso),digits=1)
round(mean(FP$stack),digits=1); round(mean(FP$lasso),digits=1)

# selection accuracy (precision)
round(tapply(X=TP$stack/(TP$stack+FP$stack),INDEX=mode,FUN=mean),digits=3)
round(tapply(X=TP$lasso/(TP$lasso+FP$lasso),INDEX=mode,FUN=mean),digits=3)
round(tapply(X=TP$enet/(TP$enet+FP$enet),INDEX=mode,FUN=mean),digits=3)

## selection accuracy (true/false positives)
#round(tapply(X=FP$stack,INDEX=mode,FUN=mean),digits=1)
#round(tapply(X=FP$lasso,INDEX=mode,FUN=mean),digits=1)
#round(tapply(X=TP$stack,INDEX=mode,FUN=mean),digits=1)
#round(tapply(X=TP$lasso,INDEX=mode,FUN=mean),digits=1)

## selection accuracy (recall)
#round(tapply(X=TP$stack/(TP$stack+FN$stack),INDEX=mode,FUN=mean),digits=3)
#round(tapply(X=TP$lasso/(TP$lasso+FN$lasso),INDEX=mode,FUN=mean),digits=3)
#round(tapply(X=TP$enet/(TP$enet+FN$enet),INDEX=mode,FUN=mean),digits=3)

## prediction accuracy
#cond <- sapply(loss,is.null)
#loss <- loss[!cond]; mode <- mode[!cond]
#stack <- sapply(loss,function(x) x$meta["stack"])
#tune <- sapply(loss,function(x) x$meta["tune"])
#tapply(100*(stack-tune)/tune,mode,median)

Application 1

# <<start>>
names <- c("colon","leukaemia",paste0("SRBCT",seq_len(4)))
y <- X <- loss <- sapply(names,function(x) list(),simplify=FALSE)
data(Colon,package="plsgenomics")
y$colon <- Colon$Y-1
X$colon <- Colon$X # 62 x 2000
data(leukemia,package="plsgenomics")
y$leukaemia <- leukemia$Y-1
X$leukaemia <- leukemia$X # 38 x 3051
data(SRBCT,package="plsgenomics")
for(i in seq_len(4)){
  y[[paste0("SRBCT",i)]] <- 1*(SRBCT$Y==i)
  X[[paste0("SRBCT",i)]] <- SRBCT$X # 83 x 2308
}
n0 <- vapply(X=y,FUN=function(x) sum(x==0),FUN.VALUE=numeric(1))
n1 <- vapply(X=y,FUN=function(x) sum(x==1),FUN.VALUE=numeric(1))

nzero <- c(seq(from=2,to=20,by=2),Inf)
for(id in c(1,NA,0)){
  for(k in seq_along(names)){
    cat("---",names[k],"---","\n")
    for(seed in seq_len(11)){
      cat("---",seed,"---","\n")
      set.seed(seed)
      loss[[k]][[seed]] <- tryCatch(cv.starnet(y=y[[k]],X=X[[k]],family="binomial",nzero=nzero,alpha.meta=id),error=function(x) NULL)
    }
  }
  #save(loss,n0,n1,file=paste0("results/app_standard_",id,".RData"))
}

#writeLines(text=capture.output(utils::sessionInfo(),cat("\n"),
#        sessioninfo::session_info()),con="results/app_standard.txt")
#<<start>>
load("results/app_standard_1.RData")
loss <- lapply(loss,function(x) x[-2]) # error at id=1, seed=2, leukaemia

median <- list()
for(i in names(loss)){
  for(j in c("meta","base")){
    median[[i]][[j]] <- apply(sapply(loss[[i]],function(x) x[[j]]),1,median)
  }
  list <- lapply(loss[[i]],function(x) x$extra)
  array <- array(data=unlist(list),dim=c(3,11,10),dimnames=list(rownames(list[[1]]),colnames(list[[1]]),seq_len(10)))
  median[[i]]$extra <- apply(X=array,MARGIN=1:2,FUN=median)
}

meta <- t(sapply(median,function(x) x$meta[c("ridge","lasso","tune","stack")]))
post <- sapply(median,function(x) x$extra["stack","Inf"])

table <- cbind("\\#0"=n0,"\\#1"=n1,format(meta,digits=1,nsmall=2)," "=format(post,digits=1,nsmall=2))
index <- cbind(seq_len(nrow(meta)),apply(meta,1,which.min)+2)
table[index] <- paste0("\\underline{",table[index],"}") 
colnames(table) <- paste0("\\text{",colnames(table),"}")
rownames(table)[3:6] <- paste0("\\textsc{",tolower(rownames(table)[3:6]),"}")
rownames(table) <- paste0("\\text{",tolower(rownames(table)),"}")
table[,c(1,2)] <- paste0("\\textcolor{gray}{",table[,c(1,2)],"}")
table[,ncol(table)] <- paste0("\\textcolor{gray}{(",table[,ncol(table)],")}")
xtable <- xtable::xtable(table,align=c("l|rrccccc"),digits=c(NA,0,0,2,2,2,2,2))
xtable::print.xtable(xtable,sanitize.text.function=function(x) x)
#<<start>>
#<<sta_table>>
#grDevices::pdf(file="manuscript/figure_STA.pdf",width=5,height=2)
graphics::par(mfrow=c(1,3),mar=c(3.3,2,0.5,0.5),oma=c(0,2,0,0))
median$SRBCT <- list()
median$SRBCT$meta <- rowMeans(sapply(median[paste0("SRBCT",1:4)],
                                     function(x) x$meta))
median$SRBCT$base <- rowMeans(sapply(median[paste0("SRBCT",1:4)],
                                     function(x) x$base))
for(i in c("colon","leukaemia","SRBCT")){
  alpha <- as.numeric(substring(names(median[[i]]$base),first=6))
  base <- median[[i]]$base
  meta <- median[[i]]$meta
  graphics::plot.new()
  graphics::plot.window(xlim=range(alpha),ylim=range(c(base,meta[c("tune","ridge","lasso","stack")])))
  graphics::axis(side=1)
  graphics::axis(side=2)
  graphics::box()
  graphics::title(xlab=expression(alpha),line=2.5)
  graphics::abline(h=meta["tune"],lty=2,col="grey")
  graphics::abline(h=meta["stack"],lty=2)
  graphics::arrows(x0=0,y0=meta["tune"],y1=meta["stack"],length=0.05,lwd=2)
  graphics::points(x=alpha,y=base,pch=21,col="black",bg="white")
  graphics::points(x=0,y=meta["ridge"],pch=16)
  graphics::points(x=1,y=meta["lasso"],pch=16)

}
graphics::title(ylab=paste0(paste0(rep(" ",times=16),collapse=""),"logistic deviance"),outer=TRUE,line=1)
#grDevices::dev.off()

sapply(median,function(x) names(which.min(x$base)))

Application 2

#<<start>>
type <- c("ACC","BLCA","BRCA","CESC","CHOL","COAD","DLBC","ESCA","GBM","HNSC","KICH","KIRC","KIRP","LAML","LGG","LIHC","LUAD","LUSC","MESO","OV","PAAD","PCPG","PRAD","READ","SARC","SKCM","STAD","TGCT","THCA","THYM","UCEC","UCS","UVM")
y <- x <- sapply(type,function(x) NULL)
for(i in seq_along(type)){
  cat(rep(c("-",type[i],"-"),times=c(10,1,10)),"\n")
  data <- curatedTCGAData::curatedTCGAData(diseaseCode=type[i],assays= "RNASeq2GeneNorm",dry.run=FALSE)
  data <- MultiAssayExperiment::mergeReplicates(data)
  X <- t(SummarizedExperiment::assay(data))
  Y <- TCGAutils::TCGAbiospec(rownames(X))$sample_definition
  print(table(Y))
  cond <- Y %in% c("Solid Tissue Normal","Primary Solid Tumor")
  Y <- Y[cond]; X <- X[cond,]
  Y <- 1*(Y=="Primary Solid Tumor")
  sd <- apply(X,2,sd)
  X <- X[,sd>=sort(sd,decreasing=TRUE)[2000]]
  X <- scale(X)
  y[[i]] <- Y; x[[i]] <- X
}
print(object.size(x),units="Mb")
#save(y,x,type,file=paste0("results/tcga_data.RData"))

#writeLines(text=capture.output(utils::sessionInfo(),cat("\n"),
#        sessioninfo::session_info()),con="results/app_processing.txt")

# cross-validation

load("results/tcga_data.RData")
type <- names(y)
nzero <- c(seq(from=2,to=20,by=2),Inf)
for(id in c(1,NA,0)){
  loss <- sapply(type,function(x) NULL)
  for(i in seq_along(type)){
    cat(rep(c("-",type[i],"-"),times=c(10,1,10)),"\n")
    if(sum(y[[i]]==0)<5|sum(y[[i]]==1)<5){next}
    set.seed(1)
    loss[[i]] <- tryCatch(cv.starnet(y=y[[i]],X=x[[i]],family="binomial",alpha.meta=id,nzero=nzero),error=function(x) NA)
  }
  #save(loss,file=paste0("results/app_extension_",id,".RData"))
}

#writeLines(text=capture.output(utils::sessionInfo(),cat("\n"),
#        sessioninfo::session_info()),con="results/app_extension.txt")
#<<start>>
load("results/tcga_data.RData")
load("results/app_extension_1.RData")
n0 <- sapply(y,function(x) sum(x==0))
n1 <- sapply(y,function(x) sum(x==1))
cond <- sapply(loss,function(x) !is.null(x)&!all(is.na(x)))
all(1*((n0>=5)&(n1>=5))==1*cond)
meta <- t(sapply(loss[cond],function(x) x$meta[c("ridge","lasso","tune","stack")]))

post <- sapply(loss[cond],function(x) x$extra["stack","Inf"])
table <- cbind("\\#0"=n0[cond],"\\#1"=n1[cond],format(meta,digits=1,nsmall=3)," "=format(post,digits=1,nsmall=3))
index <- cbind(seq_len(nrow(meta)),apply(meta,1,which.min)+2)
table[index] <- paste0("\\underline{",table[index],"}") 
colnames(table) <- paste0("\\text{",colnames(table),"}")
rownames(table) <- paste0("\\text{\\textsc{",tolower(rownames(table)),"}}")
table[,c(1,2)] <- paste0("\\textcolor{gray}{",table[,c(1,2)],"}")
table[,ncol(table)] <- paste0("\\textcolor{gray}{(",table[,ncol(table)],")}")
xtable <- xtable::xtable(table,align=c("l|rrccccc"),digits=c(NA,0,0,2,2,2,2,2))
xtable::print.xtable(xtable,sanitize.text.function=function(x) x)

sum(meta[,"lasso"]<meta[,"ridge"]); nrow(meta)
sum(meta[,"stack"]<meta[,"tune"]); nrow(meta)
mean(100*(meta[,"stack"]-meta[,"tune"])/meta[,"tune"])
wilcox.test(meta[,"stack"]-meta[,"tune"])$p.value
round((post-meta[,"stack"])/meta[,"stack"],digits=2)