regsurv: Regularized parametric survival modeling

In jeroenhoogland/regsurv: Regularized parametric survival modeling

Regularized parametric survival modeling

Description

Regularized parametric survival modeling

Usage

regsurv(
  prep,
  penpars,
  l1l2,
  groups = NULL,
  lambda.grid = NULL,
  lambda.init = -6,
  lambda.incr = 0.5,
  force.nnhazards = TRUE,
  print = FALSE,
  maxit = 100L,
  feastol = 1e-08,
  ...
)

Arguments

prep	an object of class survprep
penpars	numeric vector indicating penalized (1) and unpenalized (0) parameters. The order should follow the order the survprep model matrix
l1l2	numeric vector indicating lasso (1 or TRUE) or ridge (0 or FALSE) penalty per parameter; order as for penpars
groups	list of groups of parameters for a group lasso penalty (e.g. a group of parameters related to a dummy coded multinomial variable or to spline basis columns). If specified, penpars and l1l2 should also be specified at the group level.
lambda.grid	a grid of lambda values may be specified manually here. Default behavior is to fit all models from lambda equal to exp(-6) (default) up to the moment where all lasso penalized parameters have disappeared and all the absolute value of all ridge penalized parameters is < 1e-2.
lambda.init	logarithm of the smallest lambda penalty in the grid. Defaults to -6. Only used when lambda.grid is NULL
lambda.incr	lambda increments taken from lambda.init (on the log scale); defaults to 0.5
force.nnhazards	if TRUE, forces non-negative hazards / monotone non-decreasing cumulative hazards (only applicable for log cumulative hazard models)
print	if TRUE, prints progress (a line for each lambda for which the model was optimized)
maxit	(integer) the maximum number of iterations for the (ecos) solver, default 100L
feastol	the tolerance on the primal and dual residual, default 1e-8
...	other parameters (only used by internal functions)

Value

an object of class regsurv

optimal	TRUE when the optimization converged to an optimal value for each lambda
lambda.grid	grid of lambda values
obj.value	objective function values per lambda
betahat	matrix of model coefficients with different parameters in rows and a column per lambda value
betahat.scaled	matrix of model scaled coefficients with different parameters in rows and a column per lambda value. These scaled coefficients go with the scaled model matrices in prep
num.iters	number of iterations needed to reach the optimal solution (per lambda)
solve.times	solve times per lambda
which.param	includes 3 lists: the first with baseline model parameter indices, the second with main effect parameter indices, and the third with parameter indices for time-varying effect parameters
penpars	1 for parameters that were penalized, 0 otherwise
l1l2	l1l2 as used to fit the model
force.nnhazards	force.nnhazards as used to fit the model
survprep.id	id that matches the id of survprep used to fit this model

Examples

prep <- survprep(tte=simdata$eventtime,
                 delta=simdata$status,
                 X=as.matrix(simdata[ ,grep("x", names(simdata))]),
                 model.scale="loghazard",
                 time.scale="time",
                 spline.type="rcs",
                 ntimebasis=4,
                 qpoints=9)
# e.g. penalize all parameters but the intercept
# note that prep$which.param includes 3 lists: the first with baseline model parameters, the
# second with main effect parameters, and the third with parameters for time-varying effects
penpars <- c(rep(TRUE, length(prep$which.param[[1]])),
             rep(TRUE, length(unlist(prep$which.param[2:3]))))
penpars[1] <- FALSE # do not penalize the intercept parameter

# and use ridge for baseline parameters and lasso for all other parameters
l1l2 <- c(rep(0, length(unlist(prep$which.param[1]))),
          rep(1, length(unlist(prep$which.param[2:3]))))

# fit model over the default lambda grid
mod <- regsurv(prep, penpars, l1l2, print=TRUE)
plot(mod)

jeroenhoogland/regsurv documentation built on March 20, 2023, 3:37 a.m.