In GabrielHoffman/mvIC: Multivariate Information Criteria to identify important predictors in high dimensional multivariate regression

knitr::opts_chunk$set(tidy=FALSE, 
                      cache=TRUE,                 
                      dev=c("png", "pdf"),
                      package.startup.message = FALSE,
                      message=FALSE, 
                      error=FALSE, 
                      warning=FALSE)
options(width=100)

library(TAS)
library(beam)
library(corpcor)
library(ShrinkCovMat)
library(mvtnorm)
library(CovTools)
library(mvIC)
library(reshape2)
library(ggplot2)
library(bayesSurv)

# calculate inverse correlation matrix given rho and p
calc_inv_corr = function(rho, p){
    # https://math.stackexchange.com/questions/1762990/inverse-of-a-correlation-matrix-when-all-the-correlations-are-equal
    a = 1-rho
    A = diag(a, p)
    u = rep(sqrt(rho), p)
    # Sigma = A + u %*% t(u)
    solve_A = diag(1/(1-rho), p)
    # SigInv = solve(A) - (tcrossprod(solve(A, u))) / (1 + crossprod(u, solve(A, u)))[1]
    # solve_A - (tcrossprod(solve_A %*% u)) / (1 + crossprod(u, solve_A %*% u))[1]

    solve_A - tcrossprod(u/a) / (1 + crossprod(u, u/a))[1]
}

Vary number of responses

n = 200
rho = .5 

df = lapply( c(seq(10, 200, by=10), seq(250, 500, by=100)), function(p){

    # cat(p, "\n")  

    # responses are columns
    Y = rMVNorm( n, Q = calc_inv_corr(rho, p))

    fitBeam = beam( Y, verbose=FALSE ) 
    # fitGC = gcShrink( t(Y), var=3, cor=1, plot=FALSE)
    fitStrimmer = cov.shrink( Y, lambda.var=0, verbose=FALSE)
    fitSCM = shrinkcovmat.unequal( t(Y) )

    target = diag(apply(Y, 2, var))
    fitLW = CovEst.2003LW(Y, target) 
    fitOAS = CovEst.2010OAS(Y)
    fitRBLW = CovEst.2010RBLW(Y) 

    data.frame( n, p, 
        beam            = fitBeam@alphaOpt,
        # gcShrink      = fitGC$optimalpha,
        cov.shrink      = attr(fitStrimmer, "lambda"), 
        shrinkcovmat    = fitSCM$lambdahat, 
        LW              = fitLW$delta, 
        OAS             = fitOAS$rho, 
        RBLW            = fitRBLW$rho,
        eb_cov_est      = mvIC:::eb_cov_est(t(Y))$alpha,
        estimateMVN_EB  = mvIC:::estimateMVN_EB(Y)$alpha    )
}) 
df = do.call(rbind, df)

df2 = melt(df, id.vars=c('p', 'n'))

ggplot(df2, aes(p, value, color=variable)) + geom_point() + geom_line() + theme_bw() + theme(aspect.ratio=1) + scale_color_discrete("Method") + xlab("# responses") + ylab("alpha") + ylim(0,1)

Vary number of samples

p = 5000
rho = .1

Y_all = t(rMVNorm( 1000, Q = calc_inv_corr(rho, p)))

df = lapply( c(4:9, seq(10, 100, by=10), seq(150, 500, by=100)), function(n){

    message(n, "\n")

    Y = t(Y_all[,1:n])

    fitBeam = beam( Y, verbose=FALSE ) 
    # fitGC = gcShrink( t(Y), var=3, cor=1, plot=FALSE)
    fitStrimmer = cov.shrink( Y, lambda.var=0, verbose=FALSE)
    fitSCM = shrinkcovmat.unequal( t(Y) )

    target = diag(apply(Y, 2, var))
    # fitLW = CovEst.2003LW(Y, target)
    # fitOAS = CovEst.2010OAS(Y) 
    # fitRBLW = CovEst.2010RBLW(Y) 

    data.frame( n, p, 
        beam            = fitBeam@alphaOpt,
        # gcShrink      = fitGC$optimalpha,
        cov.shrink      = attr(fitStrimmer, "lambda"), 
        shrinkcovmat    = fitSCM$lambdahat, 
        # LW                = fitLW$delta, 
        # OAS           = fitOAS$rho, 
        # RBLW          = fitRBLW$rho,
        eb_cov_est      = mvIC:::eb_cov_est(t(Y))$alpha,
        estimateMVN_EB  = mvIC:::estimateMVN_EB(Y)$alpha)
})
df = do.call(rbind, df)

df2 = melt(df, id.vars=c('p', 'n'))

ggplot(df2, aes(n, value, color=variable)) + geom_point() + geom_line() + theme_bw() + theme(aspect.ratio=1, legend.position="right") + scale_color_discrete("Method") + xlab("# samples") + ylab("alpha") + ylim(0,1)

GabrielHoffman/mvIC documentation built on Aug. 30, 2022, 7:58 p.m.