surv.simulate {BFI}

Generate survival data with predefined censoring rates for proportional hazards models

Description

surv.simulate simulates one or multiple (right-censored) survival datasets for proportional hazards models by simultaneously incorporating a baseline hazard function from three different survival distributions (exponential, Weibull and Gompertz), a random censoring time generated from a uniform distribution with an known/unknown upper limit, and a set of baseline covariates. When the upper limit of the uniform censoring time distribution is unknown, surv.simulate can be used separately to obtain the upper limit with a predefined censoring rate.

Usage

surv.simulate(L = 1, Z, beta, a, b, u1 = 0, u2, cen_rate,
              gen_data_from = c("exp", "weibul", "gomp"),
              only_u2 = FALSE, n.rep = 100, Trace = FALSE)

Arguments

L

the number of datasets to be generated. Default is L = 1.

Z

a list of L design matrices of dimension n_\ell \times p, where n_\ell is the number of samples observed for the \ell^{\text{th}} dataset and p is the number of covariables. When L = 1, Z can be a matrix.

beta

the vector of the (true) coefficients values, with a length of p (the number of covariates).

a

scale parameter, which should be non-negative. See ‘Details’ for the form of the cumulative hazard that can be used.

b

shape/location parameter, which should be non-negative. It is not used when gen_data_from = “exp”. See ‘Details’ for the form of the cumulative hazard that can be used.

u1

a known non-negative lower limit of the uniform distribution for generating random censoring time. Default is u1 = 0. If cen_rate is not equal to 0, then u1 does not need to be defined.

u2

an non-negative upper limit of the uniform random censoring time distribution. The upper limit can be unknown (u2 = NULL, the default), or predefined. When this argument is assumed to be unknown, u2 = NULL, it is calculated by the algorithm within surv.simulate(). However, if the argument u2 is known, the censoring rate cannot be predefined (meaning there is no control over it) and is calculated based on the generated dataset. See ‘Details’ and ‘References’.

cen_rate

a value representing the proportion of observations in the simulated survival data that are censored. The range of this argument is from 0 to 1. When the upper limit is known, cen_rate can nor be predefined. If there is no censoring (cen_rate = 0), the lower (u1) and upper (u2) limits of the uniform distribution do not need to be specified.

gen_data_from

a description of the distribution from which the time to event is generated. This is a character string and can be exponential (“exp”), Weibull (“weibul”), or Gompertz (“gomp”). Can be abbreviated. By default, the exponential distribution is used.

only_u2

logical flag for calculating only the upper limit of the uniform censoring time distribution. If only_u2 = TRUE, the dataset(s) are not generated. If only_u2 = TRUE, the arguments Z and u2 do not need to be specified, and cen_rate should not be set to 0. Default is only_u2 = FALSE.

n.rep

a scalar specifying the number of iterations. This argument is exclusively used in the case of the Gompertz distribution. Default is 100.

Trace

logical flag indicating whether the output of the desired u2 and the censoring proportion for different datasets should be produced for each iteration. It works gen_data_from = “gomp”.

Details

surv.simulate function generates L simulated right-censored survival datasets from exponential, Weibull, or Gompertz distributions, incorporating the covariates, Z, distributed according to a multivariate normal distribution, with censoring time generated from a uniform distribution Uniform(u1, u2), where u1 is known but u2 can be either known or unknown.

surv.simulate() can also be used to calculate the unknown upper limit of the uniform distribution, u2, with a predefined censoring rate. To do this, set u2 = NULL and only_u2 = TRUE. In this case, the datasets are not generated; only u2 is.

surv.simulate() uses a root-finding algorithm to select the censoring parameter that achieves predefined censoring rates in the simulated survival data.

When gen_data_from = “exp”:

• the cumulative baseline hazard function is considered as \Lambda_0=a t,

• the event time for the \ell^{\text{th}} dataset, T_\ell, is computed by - log(u) \ exp(- Z_\ell \boldsymbol{\beta}) / a, where u follows a standard uniform distribution;

For gen_data_from = “weibul”:

• the cumulative hazard function is as \Lambda_0=a t ^ b,

• the event time is computed by T_\ell= (- log(u) \ exp(- Z_\ell \boldsymbol{\beta}) / a)^{1/b}, where u follows a standard uniform distribution;

For gen_data_from = “gomp”:

• the cumulative hazard function is as \Lambda_0=a (exp(b t) - 1) / b,

• the event time is computed by T_\ell= \log(1- log(u) \ exp(- Z_\ell \boldsymbol{\beta}) b / a) / b, where u follows a standard uniform distribution;

Finally the survival time is obtained by \tilde{T}_\ell=\min\{T_\ell , C_\ell \}.

The function will be updated for gen_data_from = “gomp”.

Value

surv.simulate returns a list containing the following components:

D

a list of L data frames, with dimension n_\ell \times (p+2). The first and second columns, named time and status, contain the simulated survival time and the censoring indicator, respectively, where 0 means censored and 1 means uncensored;

censor_propor

the vector of censoring proportions in the simulated datasets D, containing L elements;

u1

the lower limit of the uniform distribution used to generate random censoring times with a predefined censoring rate. Sometimes this output is less than the value entered by the user, as it is adjusted to achieve the desired amount of censoring rate;

u2

the upper limit of the uniform distribution used to generate random censoring times. If u2 = NULL, this output will be the estimated upper limit necessary to achieve the desired censoring rate across the L datasets.

Author(s)

Hassan Pazira
Maintainer: Hassan Pazira [e-mail]

References

Pazira H., Massa E., Weijers J.A.M., Coolen A.C.C. and Jonker M.A. (2024). Bayesian Federated Inference for Survival Models, arXiv. <https://arxiv.org/abs/2404.17464>

See Also

MAP.estimation

Examples

# Setting a seed for reproducibility
set.seed(1123)

#-------------------------
# Simulating Survival data
#-------------------------
N    <- c(7, 10, 13) # the sample sizes of 3 datasets
beta <- 1:4
p    <- length(beta)
L    <- 3

# Define a function to generate multivariate normal samples
mvrnorm_new <- function(n, mu, Sigma) {
    pp <- length(mu)
    e <- matrix(rnorm(n * pp), nrow = n)
    return(crossprod(t(e), chol(Sigma)) + matrix(mu, n, pp, byrow = TRUE))
}
Z <- list()
for (z in seq_len(L)) {
    Z[[z]] <- mvrnorm_new(n = N[z], mu = rep(0, p),
                          Sigma = diag(rep(1, p),p))
    colnames(Z[[z]]) <- paste0("Z_",seq_len(ncol(Z[[z]])))
}

# One simulated dataset from exponential distribution with no censoring:
surv_data <- surv.simulate(Z = Z[[1]], beta = beta, a = exp(-.9),
                           cen_rate = 0, gen_data_from = "exp")
surv_data
surv_data$D[[1]][,1:2] # The simulated survival data

# Calculate only 'u2' with a predefined censoring rate of 0.4:
u2_new <- surv.simulate(Z = Z[1:2], beta = beta, a = exp(-.9),
                        b = exp(1.8), u1 = 0.1, only_u2 = TRUE,
                        cen_rate = 0.4, gen_data_from = "weibul")$u2
u2_new

# Two simulated datasets with a known 'u2':
# Using 'u2_new' to help control over censoring rate (was chosen 0.4)
surv.simulate(Z = Z[1:2], beta = beta, a = exp(-.9), b = exp(1.8),
              u1 = 0.05, u2 = u2_new, gen_data_from = "weibul")

# Three simulated datasets from 'weibul' with an unknown 'u2':
surv.simulate(Z = Z, beta = beta, a = exp(-1), b = exp(1),
               u1 = 0.01, cen_rate = 0.3, gen_data_from = "weibul")

# Two simulated datasets from 'gomp' with unknown 'u2' and censoring rate of 0.3:
surv.simulate(Z = Z[2:3], beta = beta, a = exp(1), b = exp(2), u1 = 0.1,
              cen_rate = 0.3, gen_data_from = "gomp", Trace = TRUE)