Description
Usage
Arguments
Details
Value
Note
Author(s)
References
See Also
Examples
View source: R/bootSE.R
This function takes a model fit from an mjoint
object and
calculates standard errors and confidence intervals for the main
longitudinal and survival coefficient parameters, including the latent
association parameters, using bootstrapping (Efron and Tibshirani, 2000).
1
2
3
4
5
6
7
8
9
10
11
12  (
object,
nboot = 100,
ci = 0.95,
use.mle = ,
verbose = ,
control = (),
progress = ,
ncores = 1,
safe.boot = ,
)

object 
an object inheriting from class mjoint for a joint model
of timetoevent and multivariate longitudinal data.

nboot 
the number of bootstrap samples. Default is nboot=100 .

ci 
the confidence interval to be estimated using the
percentilemethod. Default is ci=0.95 for a 95% confidence
interval.

use.mle 
logical: should the algorithm use the maximizer from the
converged model in object as initial values for coefficients in each
bootstrap iteration. Default is use.mle=TRUE .

verbose 
logical: if TRUE , the parameter estimates and other
convergence statistics are value are printed at each iteration of the MCEM
algorithm. Default is FALSE .

control 
a list of control values with components:
nMC integer: the initial number of Monte Carlo samples to be
used for integration in the burnin phase of the MCEM. Default is
nMC= 100K.
nMCscale integer: the scale factor for the increase in Monte
Carlo size when Monte Carlo has not reduced from the previous iteration.
Default is nMCscale=5 .
nMCmax integer: the maximum number of Monte Carlo samples
that the algorithm is allowed to reach. Default is nMCmax=20000 .
burnin integer: the number of iterations for 'burnin' phase
of the optimization algorithm. It is computationally inefficient to use a
large number of Monte Carlo samples early on until one is approximately
near the maximum likelihood estimate. Default is burnin= 100K
for type='antithetic' or type='montecarlo' and
burnin= 5 for type='sobol' or type='halton' . For
standard methods, such a large burnin will generally be unnecessary and
can be reduced on an applicationspecific basis.
mcmaxIter integer: the maximum number of MCEM algorithm
iterations allowed. Default is mcmaxIter=burnin+200 .
convCrit character string: the convergence criterion to be
used. See Details.
gammaOpt character string: by default (gammaOpt='NR' ),
γ is updated using a onestep NewtonRaphson iteration, with the
Hessian matrix calculated exactly. If gammaOpt='GN' , a GaussNewton
algorithmtype iteration is implemented, where the Hessian matrix is
approximated based on calculations similar to those used for calculating
the empirical information matrix? If it is used, then the steplength is
adjusted by a nominal scaling parameter of 0.5 in order to reduce the
chance of overshooting the maximizer.
tol0 numeric: tolerance value for convergence in the
parameters; see Details. Default is tol0=1e03 .
tol1 numeric: tolerance value for convergence in the
parameters; see Details. Default is tol1=1e03 .
tol2 numeric: tolerance value for convergence in the
parameters; see Details. Default is tol2=5e03 for
type='antithetic' or type='montecarlo' and tol2=1e03
for type='sobol' or type='halton' .
tol.em numeric: tolerance value for convergence in the
multivariate linear mixed model (MVLMM). When K > 1, the optimal
initial parameters are those from the MVLMM, which is estimated using a
separate EM algorithm. Since both the E and Msteps are available in
closedform, this algorithm convergences relatively rapidly with a high
precision. Default is min(1e04 , tol2 ).
rav numeric: threshold when using convCrit='sas' that
applies absolute change (when <rav ) or relative change (when
≥rav ) criterion; see Details. Default is
0.1 , which is an order of magnitude higher than the SAS
implementation.
type character: type of Monte Carlo integration method to
use. Options are
type='montecarlo' Vanilla Monte Carlo sampling.
type='antithetic' Variance reduction method using antithetic
simulation. This is the default option.
type='sobol' QuasiMonte Carlo with a low
deterministic Sobol sequence with Owentype scrambling.
type='halton' QuasiMonte Carlo with a low deterministic
Halton sequence.

progress 
logical: should a progress bar be shown on the console to
indicate the percentage of bootstrap iterations completed? Default is
progress=TRUE .

ncores 
integer: if more than one core is available, then the bootSE
function can run in parallel via the foreach
function. By default, ncores=1 , which defaults to serial mode. Note
that if ncores >1, then progress is set to FALSE by
default, as it is not possible to display progress bars for parallel
processes at the current time.

safe.boot 
logical: should each bootstrap replication be wrapped in a
tryCatch statement to catch errors (e.g. during the
optimisation progress)? When model fitting throws errors, a new bootstrap
sample is drawn for the current iteration and the model is refit; this
process continuex until a model fits successfully. Default is FALSE .

... 
options passed to the control argument.

Standard errors and confidence intervals are obtained by repeated
fitting of the requisite joint model to bootstrap samples of the original
longitudinal and timetoevent data. Note that bootstrap is done by
sampling subjects, not individual records.
An object of class bootSE
.
This function is computationally intensive. A dynamic progress bar is
displayed showing the percentage of bootstrap models fitted. On computer
systems with more than one core available, computational time can be
reduced by passing the argument ncores
(with integer value >1) to
bootSE
, which implements parallel processing via the
foreach
package. Note: if parallel
processing is implemented, then the progress bar is not displayed.
Due to random sampling, an mjoint
model fitted to some bootstrap
samples may not converge within the specified control parameter settings.
The bootSE
code discards any models that failed to converge when
calculating the standard errors and confidence intervals. If a large
proportion of models have failed to converge, it is likely that it will
need to be refitted with changes to the control
arguments.
Graeme L. Hickey (graemeleehickey@gmail.com)
Efron B, Tibshirani R. An Introduction to the Bootstrap. 2000; Boca
Raton, FL: Chapman & Hall/CRC.
mjoint
for approximate standard errors.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22  ## Not run:
# Fit a joint model with bivariate longitudinal outcomes
()
hvd < !($log.grad) & !($log.lvmi), ]
fit < (
formLongFixed = ("grad" = log.grad ~ + sex + hs,
"lvmi" = log.lvmi ~ + sex),
formLongRandom = ("grad" = ~ 1  num,
"lvmi" = ~  num),
formSurv = Surv(fuyrs, status) ~ age,
= (hvd, hvd),
inits = ("gamma" = (0.11, 1.51, 0.80)),
timeVar = "time",
verbose = )
fit.boot < (fit, 50, use.mle = , control = (
burnin = 25, convCrit = "either",
tol0 = 6e03, tol2 = 6e03, mcmaxIter = 60))
## End(Not run)

joineRML documentation built on Jan. 5, 2021, 5:07 p.m.