model.list
.
censorReg(formula = formula(data), data = sys.parent(), weights = rep(1,n), truncation, subset, na.action, distribution = "weibull", threshold = 0, initial, fixed = list(), control = NULL, model = F, x = F, y = F, method = NULL, ...)
"censor"
as computed by the
function. See the documentation for
,
,
and
for details. Two special
variable types are handled on the right-hand-side. These are:
strata()
- indicates that the sample is to be split into sub-samples,
and a separate model fit in each subsample. See
for
additional details.
offset()
- the offset for each observation is subtracted from the
(transformed if one of the logged distribution is used) response
and truncation times prior to fitting the model.
formula
,
subset
and
weights
argument.
censor
object that is used to indicate a truncation distribution.
The usual arguments to
are used in creating the truncation object,
but notice that right truncation means that the failure (or censoring must
occur between zero and the truncation point, left truncation means that the
failure (or censoring) must occur between the truncation point and infinity,
and interval truncation means that the failure (or censoring) must occur within
the truncation interval. In particular, this implies that an observation that
is right censored cannot also come from a right- or interval-truncation
distribution, and similarly for other combinations of censoring and truncation
types.
"extreme"
(minimum extreme value),
"weibull"
,
"gaussian"
(or
"normal"
),
"lognormal"
,
"logistic"
,
"loglogistic"
,
"exponential"
,
"logexponential"
,
"rayleigh"
and
"lograyleigh"
. All ten distributions have a scale parameter.
The rayleigh distribution is the minimum extreme value distribution with the
scale parameter fixed at 0.5. The exponential distribution is the minimum
extreme value distribution with the scale parameter fixed at 1.
T
is
input,
threshold
is computed as the smallest failure time minus 10% of its
absolute value. An alternate estimate of
threshold
is obtained by
specifying
threshold = "Linearized-qq"
which computes an optimization to
minimize the curvature of a quantile-quantile plot of the quantiles of the
response versus the quantiles obtained from a Kaplan-Meier estimate of
survival.
maxiter
- the maximum number of iterations. This defaults to 500.
rel.tolerance
- the convergence criterion. When the scaled change in the
all parameter estimates is less than
rel.tolerance
, convergence is
assumed. This defaults to 0.0001.
escale
- a vector of scale factors, one for each parameter. By default,
these all equal 10000. If a scalar is input, it is applied to all
elements.
TRUE
then the model frame is included as component model
in the object returned by the function.
TRUE
then the model matrix is included as
component
x
in the object returned by the function.
TRUE
then the response is included as component
y
in the object returned by the function.
method = "model.list"
). The
model list may later be fit by a call to
.
"censorReg"
is returned. See
for details. If there are stratification variables, then a list of objects of
class
"censorReg"
is produced; this is a class
"censorRegList"
object. If
method="model.list"
, an object of class
"model.list"
is returned. See
for details.
A Fletcher-Powell algorithm is used in model fitting. The basic estimation code comes from W. Q. Meeker and is described in Meeker and Duke (1981) which details a program named CENSOR. Recently, W. Q. Meeker has refined the estimation algorithms, making them more general and stable.
Meeker, W. Q. and Duke, S. D. (1981), "CENSOR - A User-Oriented Computer Program for Life Data Analysis," The American Statistician, 35(2), 112.
# Fit an exponential model censorReg(censor(futime,fustat) ~ ecog.ps + rx, data=ovarian, distribution="exponential", fixed=list(scale=1)) censorReg(censor(days, event)~voltage, data=capacitor2, weights=weights, fixed=list(voltage=-0.30))