adapt_glmnet {adaptMT}

Adaptive P-value Thresholding with L1/L2 Penalized Generalized Linear Models

Description

adapt_glmnet is a wrapper of adapt that fits pi(x) and mu(x) by glmnet from glmnet package.

Usage

adapt_glmnet(x, pvals, piargs = list(), muargs = list(),
  dist = beta_family(), s0 = rep(0.45, length(pvals)), alphas = seq(0.01,
  1, 0.01), ...)

Arguments

x

covariates (i.e. side-information). Should be compatible to models. See Details

pvals

a vector of values in [0, 1]. P-values

piargs

a list. Other arguments passed to glmnet for fitting pi(x)

muargs

a list. Other arguments passed to glmnet for fitting mu(x)

dist

an object of class "gen_exp_family". beta_family() as default

s0

a vector of values in [0, 0.5). Initial threshold.

alphas

a vector of values in (0, 1). Target FDR levels.

...

other arguments passed to adapt (except models)

Details

adapt_glmnet by default implements LASSO on x with lambda selected by cross-validation. Specify in piargs and muargs if ridge or elastic-net penalty is needed.

See Also

adapt, adapt_glm, adapt_gam, glmnet

Examples

# Generate a 100-dim covariate x
set.seed(0)
m <- 100
n <- 1000
x <- matrix(runif(n * m), n, m)

# Generate the parameters from a conditional two-group
# logistic-Gamma GLM  where pi(x) and mu(x) are both
# linear in x. pi(x) has an intercept so that the average
# of pi(x) is 0.3
inv_logit <- function(x) {exp(x) / (1 + exp(x))}
pi1 <- 0.3
beta.pi <- c(3, 3, rep(0, m-2))
beta0.pi <- uniroot(function(b){
    mean(inv_logit(x %*% beta.pi + b)) - pi1
}, c(-100, 100))$root
pi <- inv_logit(x %*% beta.pi + beta0.pi)
beta.mu <- c(2, 2, rep(0, m-2))
beta0.mu <- 0
mu <- pmax(1, x %*% beta.mu + beta0.mu)

# Generate p-values
H0 <- as.logical(ifelse(runif(n) < pi, 1, 0))
y <- ifelse(H0, rexp(n, 1/mu), rexp(n, 1))
pvals <- exp(-y)

# Run adapt_glmnet
res <- adapt_glmnet(x, pvals, s0 = rep(0.15, n), nfits = 5)