require(graphics)
library(igraph)
library(MASS)
library(Matrix)
library(corpcor)
library(corrplot)
source("isee_all.R") ## ISEE codes
library(CorShrink)
We compare the ISEE approach with that of the standard inverse of the empirical covariance matrix, for \(n > p\) scenario.
band.mat <- function(a, p, K=1, permu=c(1:p)){
ones = rep(1,p)
Omega0 = a*ones%*%t(ones)
diag(Omega0) = rep(1,p)
Omega = 1*band(Omega0,-K,K)
Sigma = qr.solve(Omega)
Sigma = Sigma*(abs(Sigma)>1e-4)
Sigma.half=chol(Sigma)
Sigma.half = Sigma.half*(abs(Sigma.half)>1e-4)
Sigma = Sigma[permu,permu]
Omega = Omega[permu,permu]
Sigma.half = Sigma.half[permu,permu]
obj = list(Sigma=Sigma, Omega = Omega, Sigma.half = Sigma.half)
}
make.data <- function(Sigma.half, n, p, seed){
set.seed(seed)
X = matrix(rnorm(n*p),n,p)%*%Sigma.half
return(X)
}
n = 1000
p = 100
obj = band.mat(a=0.5, p, K = 1)
Sig.half = obj$Sigma.half
Ome.true = obj$Omega
X.mat = make.data(Sig.half, n, p, seed = 1000)
pcorSigma <- -as.matrix(cov2cor(obj$Omega))
diag(pcorSigma) <- rep(1, dim(pcorSigma)[1])
col2 <- c("blue", "white", "red")
corrplot(pcorSigma, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
corSigma <- cov2cor(obj$Sigma)
col2 <- c("blue", "white", "red")
corrplot(corSigma, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
pest <- - as.matrix(cov2cor(solve(cov(X.mat))))
diag(pest) <- rep(1, dim(pest)[1])
col2 <- c("blue", "white", "red")
corrplot(pest, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
regfactor = "log" # or "one", "sqrt"
npermu = 1 # or >= 2
sis.use = 0 # or 1, whether to use SIS for screening
bia.cor = 0 # or 1, whether to apply bias correction for ISEE
obj.n = isee(X.mat, regfactor, npermu, sis.use, bia.cor = 0)
pseemat <- -cov2cor(obj.n$Omega.isee)
diag(pseemat) <- rep(1, dim(pseemat)[1])
corrplot(pseemat, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
X_tilde_out <- isee.X(X.mat, bia.cor = 0, reg.fac = 1, permu = 1:ncol(X.mat),
use_slasso = TRUE)
X_tilde <- X_tilde_out$X.tilde
pStilde <- -cor(X_tilde)
diag(pStilde) = 1
corrplot(pStilde, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
pcor2 <- pCorShrinkData(X.mat, reg_type = "lm")
corrplot(pcor2, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
n = 500
p = 100
obj = band.mat(a=0.5, p, K = 1)
Sig.half = obj$Sigma.half
Ome.true = obj$Omega
X.mat = make.data(Sig.half, n, p, seed = 1000)
pcorSigma <- -as.matrix(cov2cor(obj$Omega))
diag(pcorSigma) <- rep(1, dim(pcorSigma)[1])
col2 <- c("blue", "white", "red")
corrplot(pcorSigma, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
corSigma <- cov2cor(obj$Sigma)
col2 <- c("blue", "white", "red")
corrplot(corSigma, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
pest <- - as.matrix(cov2cor(solve(cov(X.mat))))
diag(pest) <- rep(1, dim(pest)[1])
col2 <- c("blue", "white", "red")
corrplot(pest, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
regfactor = "log" # or "one", "sqrt"
npermu = 1 # or >= 2
sis.use = 0 # or 1, whether to use SIS for screening
bia.cor = 0 # or 1, whether to apply bias correction for ISEE
obj.n = isee(X.mat, regfactor, npermu, sis.use, bia.cor = 0)
pseemat <- -cov2cor(obj.n$Omega.isee)
diag(pseemat) <- rep(1, dim(pseemat)[1])
corrplot(pseemat, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
X_tilde_out <- isee.X(X.mat, bia.cor = 0, reg.fac = 1, permu = 1:ncol(X.mat),
use_slasso = TRUE)
X_tilde <- X_tilde_out$X.tilde
pStilde <- -cor(X_tilde)
diag(pStilde) = 1
corrplot(pStilde, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
pcor2 <- pCorShrinkData(X.mat, reg_type = "lm")
corrplot(pcor2, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
n = 105
p = 100
obj = band.mat(a=0.5, p, K = 1)
Sig.half = obj$Sigma.half
Ome.true = obj$Omega
X.mat = make.data(Sig.half, n, p, seed = 1000)
pcorSigma <- -as.matrix(cov2cor(obj$Omega))
diag(pcorSigma) <- rep(1, dim(pcorSigma)[1])
col2 <- c("blue", "white", "red")
corrplot(pcorSigma, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
corSigma <- cov2cor(obj$Sigma)
col2 <- c("blue", "white", "red")
corrplot(corSigma, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
pest <- - as.matrix(cov2cor(solve(cov(X.mat))))
diag(pest) <- rep(1, dim(pest)[1])
col2 <- c("blue", "white", "red")
corrplot(pest, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
regfactor = "log" # or "one", "sqrt"
npermu = 1 # or >= 2
sis.use = 0 # or 1, whether to use SIS for screening
bia.cor = 0 # or 1, whether to apply bias correction for ISEE
obj.n = isee(X.mat, regfactor, npermu, sis.use, bia.cor = 0)
pseemat <- -cov2cor(obj.n$Omega.isee)
diag(pseemat) <- rep(1, dim(pseemat)[1])
corrplot(pseemat, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
X_tilde_out <- isee.X(X.mat, bia.cor = 0, reg.fac = 1, permu = 1:ncol(X.mat),
use_slasso = TRUE)
X_tilde <- X_tilde_out$X.tilde
pStilde <- -cor(X_tilde)
diag(pStilde) = 1
corrplot(pStilde, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
pcor2 <- pCorShrinkData(X.mat, reg_type = "lm")
corrplot(pcor2, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")
pcor2 <- pCorShrinkData(X.mat, reg_type = "glmnet")
## glmnet_alpha not provided, running Lasso regression (glmnet_alpha=1). Change glmnet_alpha between 0 and 1
## for further options, alpha = 0 -> ridge, alpha = 0.5 -> elastic net
corrplot(pcor2, diag = FALSE, col = colorRampPalette(col2)(200), tl.pos = "td",
tl.col = "black", tl.cex = 0.8, rect.col = "white",
na.label.col = "white", method = "color", type = "upper")