BlockMissingData (version 0.1.0)

MBI: Variable selection method with multiple block-wise imputation (MBI)

Description

Fit a variable selection method with multiple block-wise imputation (MBI).

Usage

MBI(
  X,
  y,
  cov_index,
  sub_index,
  miss_source,
  complete,
  lambda = NULL,
  eps1 = 0.001,
  eps2 = 1e-07,
  eps3 = 1e-08,
  max.iter = 1000,
  lambda.min = ifelse(n > p, 0.001, 0.05),
  nlam = 100,
  beta0 = NULL,
  a = 3.7,
  gamma.ebic = 0.5,
  alpha1 = 0.5^(0:12),
  h1 = 2^(-(8:30)),
  ratio = 1
)

Value

beta

Estimated coefficients matrix with length(lambda) rows and dim(X)[2] columns.

lambda

The actual sequence of lambda values used.

bic1

BIC criterion values. '0' should be ignored.

notcon

Value indicating whether the algorithm is converged or not. '0' represents convergence; otherwise non-convergence.

intercept

Intercept sequence of length length(lambda).

beta0

Estimated coefficients matrix for standardized X

Arguments

X

Design matrix for block-wise missing covariates.

y

Response vector.

cov_index

Starting indexes of covariates in data sources.

sub_index

Starting indexes of subjects in missing groups.

miss_source

Indexes of missing data sources in missing groups, respectively ('NULL' represents no missing).

complete

Logical indicator of whether there is a group of complete cases. If there is a group of complete cases, it should be the first group. 'TRUE' represents that there is a group of complete cases.

lambda

A user supplied sequence of tuning parameter in penalty. If NULL, a sequence is automatically generated.

eps1

Convergence threshold at certain stage of the algorithm. Default is 1e-3.

eps2

Convergence threshold at certain stage of the algorithm. Default is 1e-7.

eps3

Convergence threshold at certain stage of the algorithm. Default is 1e-8.

max.iter

The maximum number of iterations allowed. Default is 1000.

lambda.min

Smallest value for lambda, as a fraction of the maximum value in lambda. Default depends on the size of input.

nlam

The number of lambda values. Default is 100.

beta0

Initial value for regression coefficients. If NULL, they are initialized automatically.

a

Tuning parameter in the SCAD penalty. Default is 3.7.

gamma.ebic

Parameter in the EBIC criterion. Default is 0.5.

alpha1

A sequence of candidate values for the step size in the conjugate gradient algorithm. Default is 0.5^(0:12).

h1

A sequence of candidate values for the parameter in the numerical calculation of the first derivative of the objective function. Default is 2^(-(8:30)).

ratio

Parameter in the numerical calculation of the first derivative. Default is 1.

Author

Fei Xue and Annie Qu

Details

The function uses the penalized generalized method of moments with multiple block-wise imputation to handle block-wise missing data, commonly found in multi-source datasets.

References

Xue, F., and Qu, A. (2021) Integrating Multisource Block-Wise Missing Data in Model Selection (2021), Journal of the American Statistical Association, Vol. 116(536), 1914-1927.

Examples

library(MASS)

# Number of subjects
n <- 30

# Number of total covariates
p <- 4

# Number of missing groups of subjects
ngroup <- 2

# Number of data sources
nsource <- 2

# Starting indexes of covariates in data sources
cov_index=c(1, 3)

# Starting indexes of subjects in missing groups
sub_index=c(1, 16)

# Indexes of missing data sources in missing groups, respectively ('NULL' represents no missing)
miss_source=list(NULL, 1)

# Indicator of whether there is a group of complete cases. If there is a group of complete cases,
# it should be the first group.
complete=TRUE

# Create a block-wise missing design matrix X and response vector y
set.seed(1)
sigma=diag(1-0.4,p,p)+matrix(0.4,p,p)
X <- mvrnorm(n,rep(0,p),sigma)
beta_true <- c(2.5, 0, 3, 0)
y <- rnorm(n) + X%*%beta_true

for (i in 1:ngroup) {
  if (!is.null(miss_source[[i]])) {
    if (i==ngroup) {
      if (miss_source[[i]]==nsource) {
        X[sub_index[i]:n, cov_index[miss_source[[i]]]:p] = NA
      } else {
        X[sub_index[i]:n, cov_index[miss_source[[i]]]:(cov_index[miss_source[[i]]+1]-1)] = NA
      }
    } else {
      if (miss_source[[i]]==nsource) {
        X[sub_index[i]:(sub_index[i+1]-1), cov_index[miss_source[[i]]]:p] = NA
      } else {
        X[sub_index[i]:(sub_index[i+1]-1), cov_index[miss_source[[i]]]:
        (cov_index[miss_source[[i]]+1]-1)] = NA
      }
    }
  }
}

# Now we can use the function with this simulated data
#start.time = proc.time()
result <- MBI(X=X, y=y, cov_index=cov_index, sub_index=sub_index, miss_source=miss_source,
complete=complete, nlam = 15, eps2 = 1e-3, h1=2^(-(8:20)))
#time = proc.time() - start.time

theta=result$beta
bic1=result$bic1
best=which.min(bic1[bic1!=0])
beta_est=theta[best,]