bsgw {BSGW}R Documentation

Bayesian Survival using Generalized Weibull Regression

Description

Bayesian survival model - with stratification and shrinkage - using Weibull regression on both scale and shape parameters, resulting in time-dependent (i.e. dynamic) hazard ratios.

Usage

bsgw(formula, data, formulas=formula, weights, subset, na.action=na.fail, init="survreg"
  , ordweib=FALSE, scale=0, control=bsgw.control(), print.level=2)
bsgw.control(scalex=TRUE, iter=1000, burnin=round(iter/2), sd.thresh=1e-4
  , lambda=0.0, lambdas=lambda, nskip=round(iter/10), alpha.min=0.1, alpha.max=10.0
  , beta.max=log(20), betas.max=5.0, memlim.gb=8)
## S3 method for class 'bsgw'
print(x, ...)

Arguments

formula

Survival formula expressing the time/status variables as well as covariates used in regression on scale parameter. Currently, only right and left censoring is supported. Must include intercept term.

data

Data frame containing the covariates and response variable.

formulas

Formula expressing the covariates used in regression on shape parameter. No left-hand side is necessary since the response variable information is extracted from formula. Default value is formula. Must include intercept term.

weights

Optional vector of case weights. *Not supported yet*

subset

Subset of the observations to be used in the fit. *Not supported yet*

na.action

Missing-data filter function. *Not supported yet (only na.fail behavior works)*

init

Initialization behavior. Currently, three options are supported: 1) If init="survreg", an ordinary Weibull regression is performed and coefficients are used to initialize the bsgw MCMC run. 2) If init is a survreg object, e.g. from a previous Weibull regression fit, the object can be directly passed as parameter. 3) If init is any other value, or if survreg produces error or warning, we simply set all coefficients to zero.

ordweib

If TRUE, a Bayesian ordinary Weibull model is estimated, in which any covariates in formulas are stripped away, and the inverse-logit transformation in the shape-parameter regression is replaced with a simple exponential transformation. If shrinkage parameters are kept at 0, the result is a Bayesian equivalent of an ordinary Weibull regression.

scale

If scale>0, the value of the shape parameter is fixed, i.e. not estimated from data.

control

See bsgw.control for a description of the parameters inside the control list.

print.level

Controlling verbosity level.

scalex

If TRUE, each covariate vector is centered and scaled before model estimation. The scaling parameters are saved in return object, and used in subsequent calls to predict function. Users are strongly advised against turning this feature off, since the quality of Gibbs sampling MCMC is greatly enhanced by covariate centering and scaling.

iter

Number of MCMC samples to draw.

burnin

Number of initial MCMC samples to discard before calculating summary statistics.

sd.thresh

Threshold for standard deviation of a covariate (after possible centering/scaling). If below the threshold, the corresponding coefficient is removed from sampling, i.e. its value is clamped to zero.

lambda

Bayesian Lasso shrinkage parameter for scale-parameter coefficients.

lambdas

Bayesian Lasso shrinkage parameter for shape-parameter coefficients.

nskip

Controlling how often to print progress report during MCMC run. For example, if nskip=10, progress will be reported after 10,20,30,... samples.

alpha.min

Lower bound on the shape parameter.

alpha.max

Upper bound on the shape parameter.

beta.max

Upper bound on absolute value of coefficients of scale parameter (with the exception of the intercept).

betas.max

Upper bound on absolute value of coefficients of shape parameter (with the exception of the intercept).

memlim.gb

User-specified limit on total memory (in GB) available during prediction. Hazard, cumulative hazard, and survival prediction objects are all three-dimensional arrays which can quickly grow very large, depending on data length, number of MCMC samples collected, and number of time points along which prediction is made.

x

Object of class 'bsgw', usually the result of a call to the bsgw.

...

Arguments to be passed to/from other methods.

Value

The function bsgw.control returns a list with elements identical to the input parameters. The function bsgw returns an object of class bsgw, with the following components:

call

The matched call.

formula

Same as input.

formulas

Same as input.

weights

Same as input. *Not supported yet*

subset

Same as input. *Not supported yet*

na.action

Same as input. *Not supported yet* (current behavior is na.fail)

init

Initial values for scale and shape coefficients used in MCMC sampling, either by performing an ordinary Weibull regression or by extracting estimated coefficients from a previously-performed such regression.

ordweib

Same as input.

survreg.scale.ref

Value of scale parameter, estimated using ordinary Weibull regression by calling the survreg function in the survival package.

ordreg

The "survreg" object returned from calling the same function for initialization of coefficients.

scale

Same as input.

control

Same as input.

X

Model matrix used for regression on scale parameter, after potential centering and scaling. The corresponding vector of coefficients is called beta.

Xs

Model matrix used for regression on shape parameter, after potential centering and scaling. The corresponding vector of coefficients is called betas.

y

Survival response variable (time and status) used in the model.

contrasts

The contrasts used for scale-parameter coefficients.

contrastss

The contrasts used for shape-parameter coefficients.

xlevels

A record of the levels of the factors used in fitting for scale parameter regression.

xlevelss

A record of the levels of the factors used in fitting for shape parameter regression.

terms

The terms object used for scale parameter regression.

termss

The terms object used for shape parameter regression.

colnamesX

Names of columns for X, also names of scale coefficients.

colnamesXs

Names of columns for Xs, also names of shape coefficients.

apply.scale.X

Index of columns of X where scaling has been applied.

apply.scale.Xs

Index of columns of Xs where scaling has been applied.

centerVec.X

Vector of centering parameters for columns of X indicated by apply.scale.X.

scaleVec.X

Vector of scaling parameters for columns of X indicated by apply.scale.X.

centerVec.Xs

Vector of centering parameters for columns of Xs indicated by apply.scale.Xs.

scaleVec.Xs

Vector of scaling parameters for columns of Xs indicated by apply.scale.Xs.

idx

Vector of indexes into X for which sampling occured. All columns of X whose standard deviation falls below sd.thresh are excluded from sampling and their corresponding coefficients are clamped to 0.

idxs

Vector of indexes into Xs for which sampling occured. All columns of Xs whose standard deviation falls below sd.thresh are excluded from sampling and their corresponding coefficients are clamped to 0.

median

List of median values, with elements including beta (coefficients of scale regression), betas (coefficients of shape regression), survreg.scale (value of surgreg-style scale parameter for all training set observations).

smp

List of coefficient samples, with the following elements: 1) beta (scale parameter coefficients), 2) betas (shape parameter coefficients), 3) lp (vector of linear predictor for scale parameter, within-sample), 4) loglike (log-likelihood of model), 5) logpost (log-posterior of mode, i.e. log-likelihood plus the shrinkage term). The last two entities are used during within-sample prediction of response, i.e. during a subsequent call to predict. Each parameter has control$iter samples.

km.fit

Kaplan-Meyer fit to training data. Used in plot.bsgw method.

tmax

Maximum time value in training set. Used in predict.bsgw for automatic selection of the tvec parameter.

Author(s)

Alireza S. Mahani, Mansour T.A. Sharabiani

References

Mazucheli J., Louzada-Neto F. and Achnar J.A. (2002). Lifetime models with nonconstant shape parameters. Confiabilidade. III Jornada Regional de Estatistica e II Semana da Estatistica, Maringa.

Neal R.M. (2003). Slice Sampling. Annals of Statistics, 31, 705-767.

Park T. and Casella G. (2008). The Bayesian Lasso. Journal of the American Statistical Association, 103, 681-686.

See Also

For calculating median and lower/upper bounds on coefficients, use summary.bsgw.

For prediction, use predict.bsgw.

Examples

## model estimation using 800 samples, printing progress every 100 samples
library("survival")
data(ovarian)
est <- bsgw(Surv(futime, fustat) ~ ecog.ps + rx, ovarian
            , control=bsgw.control(iter=400, nskip=100))

## comparing shape of Weibull curves between ordinary Weibull and bsgw
## since in bsgw shape is dependent on covariates, only a population average is meaningful
## Note that survreg-style scale is inverse of bsgw shape parameter, see survreg help page
west <- survreg(Surv(futime, fustat) ~ ecog.ps + rx, ovarian)
cat("constant survreg-style scale parameter:", west$scale, "\n")
cat("population average of survreg-style scale parameter from bsgw model:"
  , mean(est$median$survreg.scale), "\n")

[Package BSGW version 0.9.4 Index]