ABSLOPE: ABSLOPE
In mgerus/ABSLOPE: High-dimensional Model Selection with Missing Values

Description Usage Arguments Value Examples

View source: R/ABSLOPE.R

This function uses algorithm SAEM to fit the ABSLOPE model with missing data.

ABSLOPE(X, y, lambda, a, b, beta.start = NA, maxit = 300,
  case = "MCAR", seed = NA, print_iter = FALSE, tol_em = 1e-06,
  impute = "PCA", sigma.known = NA, sigma.init = NA, scale = FALSE,
  method_na = "lineq")

`X`	Design matrix with missingness n p*
`y`	Response vector n 1*
`lambda`	Penalization coefficient
`a`	Coefficient for sparsity prior
`b`	Coefficient for sparsity prior
`beta.start`	Initialization for τ
`maxit`	Maximum number of iteration. The default is maxit = 300.
`case`	Missing mechanism
`seed`	An integer as a seed set for the radom generator.
`print_iter`	If TRUE, miss.saem will print the estimated parameters in each iteration of SAEM.
`tol_em`	The tolerance to stop SAEM. The default is tol_em = 1e-6.
`impute`	Imputation method for the initilization. The default is impute = 'PCA'.
`sigma.known`	If TRUE, σ is known.
`sigma.init`	Value for known σ.
`scale`	If TRUE, perform scaling in each iteration.
`method_na`	Method for dealing with missing values. Default is method_na = 'lineq'.

A list with components

`X.sim`	Simulated dataset in the last iteration.
`beta`	Estiamated β.
`seqbeta`	Sequence of β estimated in each iteration.
`beta.new`	Estiamated β * model selection indicator γ.
`gamma`	Simulated model selection indicator γ in the last iteration.
`gamma.avg`	Average γ during the iterations.
`seqgamma`	Sequence of γ in each iteration.
`pi.binom`	Probability for selecting each variable.
`theta`	Simulated sparsity θ in the last iteration.
`seqtheta`	Sequence of θ in each iteration.
`sigma`	Estimated σ.
`seqsigma`	Sequence of σ estimated in each iteration.
`mu`	Estimated μ.
`Sigma`	Estimated Σ.
`c`	Simulated inverse of average signal magnitude c in the last iteration.
`seqc`	Sequence of c in each iteration.

# Generate dataset
n=50
p=50
nspr=10 # non-sparsity
p.miss=0.1 # percentage of missingness
mu = rep(0,p)
Sigma=toeplitz(0^(0:(p-1)))
signallevel=3 # signal strength
sigma=1 # noise
amplitude = signallevel*sqrt(2*log(p))  # signal amplitude
# Design matrix
X <- matrix(rnorm(n*p), nrow=n)%*%chol(Sigma) + matrix(rep(mu,n), nrow=n, byrow = TRUE)
X = scale(X)/sqrt(n) #scale

# Coefficient and response vectors
nonzero = sample(p, nspr)
beta = amplitude * (1:p %in% nonzero)
y = X %*% beta +  sigma*rnorm(n)

# Missing values
X.obs <- X
p.miss <- 0.10
patterns <- runif(n*p)<p.miss #missing completely at random
X.obs[patterns] <- NA

# ABSLOPE
lambda = create_lambda_bhq(ncol(X),fdr=0.10)
list.SLOB = ABSLOPE(X, y, lambda, a=2/p, b=1-2/p)