README.md

In chenchi0526/SIsMiss:

SIsMiss

SIsMiss is an R package for variable selection and statistical inference under shadow variable assumption for a linear regression with missing subjects in response. SIsMiss is applicable to linear regression with regularization and without regularization. Robust inference will be obtained for both scenario of missing at random (MAR) and missing not at random (MNAR). The estimates and standard error for the unknown regression coefficients will be returned, along with optional confidence intervals.

The underlying estimating method is based on conditional likelihood discussed in Zhao and Chen (2020).

Installation

You can install SIsMiss from github with:

# install.packages("devtools")
devtools::install_github("chenchi0526/SIsMiss")

Example

For simplicity, consider a linear regression with no missing subjects.

rm(list = ls())
library(SIsMiss)
n <- 50
p <- 8
beta <- c(3, 0, 1.5, 0, 2, rep(0, p-5))
gamma <- 3
u <- matrix(rnorm(n*p), ncol = p, nrow = n)
z <- rnorm(n, 0, 1)
y <- u %*% beta + gamma*z + rnorm(n)

Unregularized linear regression

For unregularized linear regression, the standard error can be estimated via asymptotic theory or perturbation method.

When estimating standard error via asymptotic theory, the symmetric asymptotic confidence interval will be returned.

SIsMiss(y, z, u, regularize = FALSE, cov.names = NULL,
        se.method = "asymp", CI.alpha = 0.05,
        M = NULL, seed_num = NULL)

When estimating standard error via perturbation, the lower bound and upper bound for confidence interval are the α/2-th quantile and 1-α/2-th quantile for the samples of perturbated estimates.

SIsMiss(y, z, u, regularize = FALSE, cov.names = NULL,
        se.method = "perturb", CI.alpha = 0.05,
        M = 200, seed_num = 123)

Regularized linear regression

For regularized linear regression, the adaptive LASSO penalty is considered where the tuning parameter is determined by BIC. The standard error of coefficients is estimated via perturbation method only.

SIsMiss(y, z, u, regularize = TRUE, cov.names = NULL,
        se.method = "perturb", CI.alpha = 0.05,
        M = 200, seed_num = 123)

chenchi0526/SIsMiss documentation built on Dec. 8, 2020, 2:35 a.m.