function to generate data with hierarchical sparsity

function to generate data with hierarchical sparsity

genHierSparseData(
  ncats,
  nvars,
  nobs,
  nobs.test = 100,
  hier.sparsity.param = 0.5,
  avg.hier.zeros = NULL,
  prop.zero.vars = 0.5,
  effect.size.max = 0.5,
  misspecification.prop = 0,
  family = c("gaussian", "binomial", "coxph"),
  sd = 1,
  snr = NULL,
  beta = NULL,
  tau = 10,
  covar = 0
)

Arguments

ncats	number of categories to stratify on
nvars	number of variables
nobs	number of observations per strata to simulate
nobs.test	number of independent test observations per strata to simulate
hier.sparsity.param	parameter between 0 and 1 which determines how much hierarchical sparsity there is. To achieve a desired total level of sparsity among the variables with hierarchical sparsity, this parameter can be estimated using the function 'estimate.hier.sparsity.param'
avg.hier.zeros	desired percent of zero variables among the variables with hierarchical zero patterns. If this is specified, it will override the given hier.sparsity.param value and estimate it. This takes a while
prop.zero.vars	proportion of all variables that will be zero across all strata
effect.size.max	maximum magnitude of the true effect sizes
misspecification.prop	proportion of variables with hierarchical missingness misspecified
family	family for the response variable
sd	standard devation for gaussian simulations
snr	signal-to-noise ratio (only used for family = "gaussian")
beta	a matrix of true beta values. If given, then no beta will be created and data will be simulated from the given beta
tau	rate parameter for rexp() for generating time-to-event outcomes
covar	scalar, pairwise covariance term for covariates

Examples

set.seed(123)

dat.sim <- genHierSparseData(ncats = 3, nvars = 100, nobs = 200)

# estimate hier.sparsity.param for 0.15 total proportion of nonzero variables
# among vars with hierarchical zero patterns
if (FALSE) {
hsp <- estimate.hier.sparsity.param(ncats = 3, nvars = 50, avg.hier.zeros = 0.15, nsims = 100)
}
# the above results in the following value
hsp <- 0.6270698

# check that this does indeed achieve the desired level of sparsity
mean(replicate(50, mean(genHierSparseBeta(ncats = 3,
                       nvars = 50, hier.sparsity.param = hsp) != 0)  ))
#> [1] 0.1527

dat.sim2 <- genHierSparseData(ncats = 3, nvars = 100, nobs = 200, hier.sparsity.param = hsp)

sparseBeta <- genHierSparseBeta(ncats = 3, nvars = 100, hier.sparsity.param = hsp)

## generate data with already generated beta
dat.sim3 <- genHierSparseData(ncats = 3, nvars = 100, nobs = 200, beta = sparseBeta)




## complete example:
## 50% sparsity:
hsp <- 0.2626451

dat.sim <- genHierSparseData(ncats = 3, nvars = 25,
                             nobs = 150, nobs.test = 1000,
                             hier.sparsity.param = hsp,
                             prop.zero.vars = 0.5,
                             effect.size.max = 0.25,
                             family = "gaussian")

x        <- dat.sim$x
x.test   <- dat.sim$x.test
y        <- dat.sim$y
y.test   <- dat.sim$y.test
grp      <- dat.sim$group.ind
grp.test <- dat.sim$group.ind.test

fit.adapt <- cv.vennLasso(x, y,
                          grp,
                          adaptive.lasso = TRUE,
                          nlambda        = 25,
                          family         = "gaussian",
                          abs.tol        = 1e-5,
                          rel.tol        = 1e-5,
                          maxit          = 1000,
                          irls.maxit     = 15L,
                          gamma          = 0.2,
                          standardize    = FALSE,
                          intercept      = TRUE,
                          nfolds         = 3,
                          model.matrix   = TRUE)

preds.a <- predict(fit.adapt$vennLasso.fit, x.test, grp.test, s = fit.adapt$lambda.min,
                   type = 'response')