riskRegression (version 2022.09.23)

synthesize: Cooking and synthesizing survival data

Description

Fit parametric regression models to the outcome distribution and optionally also parametric regression models for the joint distribution of the predictors structural equation models. Then the function sim.synth can be called on the resulting object to to simulate from the parametric model based on the machinery of the lava package

Usage

synthesize(object, data, ...)
# S3 method for formula
synthesize(
  object,
  data,
  recursive = FALSE,
  max.levels = 10,
  verbose = FALSE,
  ...
)
# S3 method for lvm
synthesize(
  object,
  data,
  max.levels = 10,
  logtrans = NULL,
  verbose = FALSE,
  fix.names = FALSE,
  ...
)

Value

lava object

Arguments

object

Specification of the synthesizing model structures. Either a formula or a lvm object. See examples.

data

Data to be synthesized.

...

Not used yet.

recursive

Let covariates recursively depend on each other.

max.levels

Integer used to guess which variables are categorical. When set to 10, the default, variables with less than 10 unique values in data are treated as categorical.

verbose

Logical. If TRUE then more messages and warnings are provided.

logtrans

Vector of covariate names that should be log-transformed. This is primarily for internal use.

fix.names

Fix possible problematic covariate names.

Author

Thomas A. Gerds <[e-mail]>

Details

Synthesizes survival data (also works for linear models and generalized linear models). The idea is to be able to simulate new data sets that mimic the original data. See the vignette vignette("synthesize",package = "riskRegression") for more details.

The simulation engine is: lava.

See Also

lvm

Examples

# pbc data
library(survival)
library(lava)
data(pbc)
pbc <- na.omit(pbc[,c("time","status","sex","age","bili")])
pbc$logbili <- log(pbc$bili)
v_synt <- synthesize(object=Surv(time,status)~sex+age+logbili,data=pbc)
d <- simsynth(v_synt,1000)
fit_sim <- coxph(Surv(time,status==1)~age+sex+logbili,data=d)
fit_real <- coxph(Surv(time,status==1)~age+sex+logbili,data=pbc)
# compare estimated log-hazard ratios between simulated and real data
cbind(coef(fit_sim),coef(fit_real))

u <- lvm()
distribution(u,~sex) <- binomial.lvm()
distribution(u,~age) <- normal.lvm()
distribution(u,~trt) <- binomial.lvm()
distribution(u,~logbili) <- normal.lvm()
u <-eventTime(u,time~min(time.cens=0,time.transplant=1,time.death=2), "status")
lava::regression(u,logbili~age+sex) <- 1
lava::regression(u,time.transplant~sex+age+logbili) <- 1
lava::regression(u,time.death~sex+age+logbili) <- 1
lava::regression(u,time.cens~1) <- 1
transform(u,logbili~bili) <- function(x){log(x)}
u_synt <- synthesize(object=u, data=na.omit(pbc))
set.seed(8)
d <- simsynth(u_synt,n=1000)
# note: synthesize may relabel status variable
fit_sim <- coxph(Surv(time,status==1)~age+sex+logbili,data=d)
fit_real <- coxph(Surv(time,status==1)~age+sex+log(bili),data=pbc)
# compare estimated log-hazard ratios between simulated and real data
cbind(coef(fit_sim),coef(fit_real))

#
# Cancer data
#
data(cancer)
b <- lvm()
distribution(b,~rx) <- binomial.lvm()
distribution(b,~age) <- normal.lvm()
distribution(b,~resid.ds) <- binomial.lvm()
distribution(b,~ecog.ps) <- binomial.lvm()
lava::regression(b,time.death~age+rx+resid.ds) <- 1
b<-eventTime(b,futime~min(time.cens=0,time.death=1), "fustat")
b_synt <- synthesize(object = b, data = ovarian)
D <- simsynth(b_synt,1000)
fit_real <- coxph(Surv(futime,fustat)~age+rx+resid.ds, data=ovarian)
fit_sim <- coxph(Surv(futime,fustat)~age+rx+resid.ds, data=D)
cbind(coef(fit_sim),coef(fit_real))
w_synt <- synthesize(object=Surv(futime,fustat)~age+rx+resid.ds, data=ovarian)
D <- simsynth(w_synt,1000)
fit_sim <- coxph(Surv(futime,fustat==1)~age+rx+resid.ds,data=D)
fit_real <- coxph(Surv(futime,fustat==1)~age+rx+resid.ds,data=ovarian)
# compare estimated log-hazard ratios between simulated and real data
cbind(coef(fit_sim),coef(fit_real))