| Type: | Package |
| Title: | Selection Threshold Optimized Empirically via Splitting |
| Version: | 0.2 |
| Date: | 2022-05-20 |
| Author: | Marinela Capanu [aut, cre], Mihai Giurcanu [aut, ctb], Colin Begg [aut], Mithat Gonen [aut] |
| Maintainer: | Marinela Capanu <capanum@mskcc.org> |
| Imports: | MASS, cvTools, glmnet, changepoint |
| Description: | Implements variable selection procedures for low to moderate size generalized linear regressions models. It includes the STOPES functions for linear regression (Capanu M, Giurcanu M, Begg C, Gonen M, Optimized variable selection via repeated data splitting, Statistics in Medicine, 2020, 19(6):2167-2184) as well as subsampling based optimization methods for generalized linear regression models (Marinela Capanu, Mihai Giurcanu, Colin B Begg, Mithat Gonen, Subsampling based variable selection for generalized linear models). |
| License: | GPL-2 |
| NeedsCompilation: | no |
| Packaged: | 2022-05-25 14:35:54 UTC; mgiurcanu |
| Repository: | CRAN |
| Date/Publication: | 2022-05-27 08:40:13 UTC |
ALASSO variable selection via cross-validation regularization parameter selection
Description
alasso.cv computes the ALASSO estimator.
Usage
alasso.cv(x, y)
Arguments
x |
n x p covariate matrix |
y |
n x 1 response vector |
Value
alasso.cv returns the ALASSO estimate
alasso |
the ALASSO estimator |
References
Hui Zou, (2006). "The adaptive LASSO and its oracle properties", JASA, 101 (476), 1418-1429
Examples
p <- 5
n <- 100
beta <- c(2, 1, 0.5, rep(0, p - 3))
x <- matrix(nrow = n, ncol = p, rnorm(n * p))
y <- rnorm(n) + crossprod(t(x), beta)
alasso.cv(x, y)
Optimization via Subsampling (OPTS)
Description
opts computes the OPTS MLE in low dimensional
case.
Usage
opts(X, Y, m, crit = "aic", prop_split = 0.5, cutoff = 0.75, ...)
Arguments
X |
n x p covariate matrix (without intercept) |
Y |
n x 1 binary response vector |
m |
number of subsamples |
crit |
information criterion to select the variables: (a) aic = minimum AIC and (b) bic = minimum BIC |
prop_split |
proportion of subsample size and sample size, default value = 0.5 |
cutoff |
cutoff used to select the variables using the stability selection criterion, default value = 0.75 |
... |
other arguments passed to the glm function, e.g., family = "binomial" |
Value
opts returns a list:
betahat |
OPTS MLE of regression parameter vector |
Jhat |
estimated set of active predictors (TRUE/FALSE) corresponding to the OPTS MLE |
SE |
standard error of OPTS MLE |
freqs |
relative frequency of selection for all variables |
Examples
require(MASS)
P = 15
N = 100
M = 20
BETA_vector = c(0.5, rep(0.5, 2), rep(0.5, 2), rep(0, P - 5))
MU_vector = numeric(P)
SIGMA_mat = diag(P)
X <- mvrnorm(N, MU_vector, Sigma = SIGMA_mat)
linearPred <- cbind(rep(1, N), X)
Y <- rbinom(N, 1, plogis(linearPred))
# OPTS-AIC MLE
opts(X, Y, 10, family = "binomial")
Threshold OPTimization via Subsampling (OPTS_TH)
Description
opts_th computes the threshold OPTS MLE in low
dimensional case.
Usage
opts_th(X, Y, m, crit = "aic", type = "binseg", prop_split = 0.5,
prop_trim = 0.2, q_tail = 0.5, ...)
Arguments
X |
n x p covariate matrix (without intercept) |
Y |
n x 1 binary response vector |
m |
number of subsamples |
crit |
information criterion to select the variables: (a) aic = minimum AIC and (b) bic = minimum BIC |
type |
method used to minimize the trimmed and averaged information criterion: (a) min = observed minimum subsampling trimmed average information, (b) sd = observed minimum using the 0.25sd rule (corresponding to OPTS-min in the paper), (c) pelt = PELT changepoint algorithm (corresponding to OPTS-PELT in the paper), (d) binseg = binary segmentation changepoint algorithm (corresponding to OPTS-BinSeg in the paper), (e) amoc = AMOC method. |
prop_split |
proportion of subsample size of the sample size; default value is 0.5 |
prop_trim |
proportion that defines the trimmed mean; default value = 0.2 |
q_tail |
quantiles for the minimum and maximum p-values across the subsample cutpoints used to define the range of cutpoints |
... |
other arguments passed to the glm function, e.g., family = "binomial" |
Value
opts_th returns a list:
betahat |
STOPES MLE of regression parameters |
SE |
SE of STOPES MLE |
Jhat |
set of active predictors (TRUE/FALSE) corresponding to STOPES MLE |
cuthat |
estimated cutpoint for variable selection |
pval |
marginal p-values from univariate fit |
cutpoits |
subsample cutpoints |
aic_mean |
mean subsample AIC |
bic_mean |
mean subsample BIC |
Examples
require(MASS)
P = 15
N = 100
M = 20
BETA_vector = c(0.5, rep(0.5, 2), rep(0.5, 2), rep(0, P - 5))
MU_vector = numeric(P)
SIGMA_mat = diag(P)
X <- mvrnorm(N, MU_vector, Sigma = SIGMA_mat)
linearPred <- cbind(rep(1, N), X)
Y <- rbinom(N, 1, plogis(linearPred))
# Threshold OPTS-BinSeg MLE
opts_th(X, Y, M, family = "binomial")
Selection of Threshold OPtimized Empirically via Splitting (STOPES)
Description
stopes computes the STOPES estimator.
Usage
stopes(x, y, m = 20, prop_split = 0.50, prop_trim = 0.20, q_tail = 0.90)
Arguments
x |
n x p covariate matrix |
y |
n x 1 response vector |
m |
number of split samples, with default value = 20 |
prop_split |
proportion of data used for training samples, default value = 0.50 |
prop_trim |
proportion of trimming, default prop_trim = 0.20 |
q_tail |
proportion of truncation samples across the split samples, default values = 0.90 |
Value
stopes returns a list with the STOPE estimates via data splitting using 0.25 method and the PELT method:
beta_stopes |
the STOPE estimate via data splitting |
J_stopes |
the set of active predictors corresponding to STOPES via data splitting |
final_cutpoints |
the final cutpoint for STOPES |
beta_pelt |
the STOPE estimate via PELT |
J_pelt |
the set of active predictors corresponding to STOPES via PELT |
final_cutpoints_PELT |
the final cutpoint for PELT |
quan_NA |
test if the vector of trimmed cutpoints has length 0, with 1 if TRUE and 0 otherwise |
Author(s)
Marinela Capanu, Mihai Giurcanu, Colin Begg, and Mithat Gonen
Examples
p <- 5
n <- 100
beta <- c(2, 1, 0.5, rep(0, p - 3))
x <- matrix(nrow = n, ncol = p, rnorm(n * p))
y <- rnorm(n) + crossprod(t(x), beta)
stopes(x, y)