vennLasso (version 0.1)

vennLasso: Fitting vennLasso models

Description

Fitting vennLasso models

Usage

vennLasso(x, y, groups, family = c("gaussian", "binomial"), nlambda = 100L,
  lambda = NULL, lambda.min.ratio = NULL, lambda.fused = NULL,
  penalty.factor = NULL, group.weights = NULL, adaptive.lasso = FALSE,
  adaptive.fused = FALSE, gamma = 1, standardize = FALSE,
  intercept = TRUE, one.intercept = FALSE, compute.se = FALSE,
  conf.int = NULL, rho = NULL, dynamic.rho = TRUE, maxit = 500L,
  abs.tol = 1e-05, rel.tol = 1e-05, irls.tol = 1e-05, irls.maxit = 100L,
  model.matrix = FALSE, ...)

Arguments

x

input matrix of dimension nobs by nvars. Each row is an observation, each column corresponds to a covariate

y

numeric response vector of length nobs

groups

A list of length equal to the number of groups containing vectors of integers indicating the variable IDs for each group. For example, groups = list(c(1,2), c(2,3), c(3,4,5)) specifies that Group 1 contains variables 1 and 2, Group 2 contains variables 2 and 3, and Group 3 contains variables 3, 4, and 5. Can also be a matrix of 0s and 1s with the number of columns equal to the number of groups and the number of rows equal to the number of variables. A value of 1 in row i and column j indicates that variable i is in group j and 0 indicates that variable i is not in group j.

family

"gaussian" for least squares problems, "binomial" for binary response, and "coxph" for time-to-event outcomes (not yet available)

nlambda

The number of lambda values. Default is 100.

lambda

A user-specified sequence of lambda values. Left unspecified, the a sequence of lambda values is automatically computed, ranging uniformly on the log scale over the relevant range of lambda values.

lambda.min.ratio

Smallest value for lambda, as a fraction of lambda.max, the (data derived) entry value (i.e. the smallest value for which all parameter estimates are zero). The default depends on the sample size nobs relative to the number of variables nvars. If nobs > nvars, the default is 0.0001, close to zero. If nobs < nvars, the default is 0.01. A very small value of lambda.min.ratio can lead to a saturated fit when nobs < nvars.

lambda.fused

tuning parameter for fused lasso penalty

penalty.factor

vector of weights to be multiplied to the tuning parameter for the group lasso penalty. A vector of length equal to the number of groups

group.weights

A vector of values representing multiplicative factors by which each group's penalty is to be multiplied. Often, this is a function (such as the square root) of the number of predictors in each group. The default is to use the square root of group size for the group selection methods.

adaptive.lasso

Flag indicating whether or not to use adaptive lasso weights. If set to TRUE and group.weights is unspecified, then this will override group.weights. If a vector is supplied to group.weights, then the adaptive.lasso weights will be multiplied by the group.weights vector

adaptive.fused

Flag indicating whether or not to use adaptive fused lasso weights.

gamma

power to raise the MLE estimated weights by for the adaptive lasso. defaults to 1

standardize

Should the data be standardized? Defaults to FALSE.

intercept

Should an intercept be fit? Defaults to TRUE

one.intercept

Should a single intercept be fit for all subpopulations instead of one for each subpopulation? Defaults to FALSE.

compute.se

Should standard errors be computed? If TRUE, then models are re-fit with no penalization and the standard errors are computed from the refit models. These standard errors are only theoretically valid for the adaptive lasso (when adaptive.lasso is set to TRUE)

conf.int

level for confidence intervals. Defaults to NULL (no confidence intervals). Should be a value between 0 and 1. If confidence intervals are to be computed, compute.se will be automatically set to TRUE

rho

ADMM parameter. must be a strictly positive value. By default, an appropriate value is automatically chosen

dynamic.rho

TRUE/FALSE indicating whether or not the rho value should be updated throughout the course of the ADMM iterations

maxit

integer. Maximum number of ADMM iterations. Default is 500.

abs.tol

absolute convergence tolerance for ADMM iterations for the relative dual and primal residuals. Default is 10^{-5}, which is typically adequate.

rel.tol

relative convergence tolerance for ADMM iterations for the relative dual and primal residuals. Default is 10^{-5}, which is typically adequate.

irls.tol

convergence tolerance for IRLS iterations. Only used if family != "gaussian". Default is 10^-5.

irls.maxit

integer. Maximum number of IRLS iterations. Only used if family != "gaussian". Default is 100.

model.matrix

logical flag. Should the design matrix used be returned?

...

not used

Value

An object with S3 class "vennLasso"

Examples

# NOT RUN {
library(Matrix)

# first simulate heterogeneous data using
# genHierSparseData
set.seed(123)
dat.sim <- genHierSparseData(ncats = 2, nvars = 25,
                             nobs = 200, 
                             hier.sparsity.param = 0.5,
                             prop.zero.vars = 0.5,
                             family = "gaussian")

x          <- dat.sim$x
conditions <- dat.sim$group.ind
y          <- dat.sim$y

true.beta.mat <- dat.sim$beta.mat

fit <- vennLasso(x = x, y = y, groups = conditions)

(true.coef <- true.beta.mat[match(dimnames(fit$beta)[[1]], rownames(true.beta.mat)),])
round(fit$beta[,,21], 2)

## fit adaptive version and compute confidence intervals
afit <- vennLasso(x = x, y = y, groups = conditions, conf.int = 0.95, adaptive.lasso = TRUE)

(true.coef <- true.beta.mat[match(dimnames(fit$beta)[[1]], rownames(true.beta.mat)),])[,1:10]
round(afit$beta[,1:10,28], 2)
round(afit$lower.ci[,1:10,28], 2)
round(afit$upper.ci[,1:10,28], 2)

aic.idx <- which.min(afit$aic)
bic.idx <- which.min(afit$bic)

# actual coverage
# actual coverage
mean(true.coef[afit$beta[,-1,aic.idx] != 0] >= 
             afit$lower.ci[,-1,aic.idx][afit$beta[,-1,aic.idx] != 0] &
         true.coef[afit$beta[,-1,aic.idx] != 0] <= 
             afit$upper.ci[,-1,aic.idx][afit$beta[,-1,aic.idx] != 0])

(covered <- true.coef >= afit$lower.ci[,-1,aic.idx] & true.coef <= afit$upper.ci[,-1,aic.idx])
mean(covered)


# logistic regression example
set.seed(123)
dat.sim <- genHierSparseData(ncats = 2, nvars = 25,
                             nobs = 200, 
                             hier.sparsity.param = 0.5,
                             prop.zero.vars = 0.5,
                             family = "binomial",
                             effect.size.max = 0.5) # don't make any 
                                                    # coefficients too big

x           <- dat.sim$x
conditions  <- dat.sim$group.ind
y           <- dat.sim$y
true.beta.b <- dat.sim$beta.mat

bfit <- vennLasso(x = x, y = y, groups = conditions, family = "binomial")

(true.coef.b <- -true.beta.b[match(dimnames(fit$beta)[[1]], rownames(true.beta.b)),])
round(bfit$beta[,,20], 2)

# }

Run the code above in your browser using DataCamp Workspace