limits.bca(boot.obj, probs=c(0.025, 0.05, 0.95, 0.975), details=F, z0=NULL, acceleration=NULL, group.size=NULL, frame.eval = <<see below>>, subset.statistic=1:p, narrow=F)
z0
,
acceleration
,
group.size
, and the adjusted probability levels for each BCa confidence
limit
along with the BCa limits. If
details=F
, only the BCa confidence limits are
returned.
L
component of
boot.obj
,
if it is present, otherwise using
.
In the latter case, the statistic must allow samples of reduced size;
see
.
It may also be calculated by
.
acceleration
. It is time consuming to do a full delete-1 jackknife
if
n
is large, so the
default is to use a group size of
floor(n/20)
, which yields roughly 20
groups of points, and hence 20 jackknife replicates. For the full jackknife,
use
group.size=1
.
boot.obj
can be found.
You need to specify this if objects can't be found by their
original names, or have changed; see
.
narrow=FALSE
(the default), then intervals are wider and coverage probabilities
generally more accurate than with
narrow=TRUE
.
details=F
,
returns a matrix of BCa confidence limits with one row per
parameter of interest.
Otherwise, returns a list with components:
If
acceleration
or
z0
are large and requested probabilities are
near 0 or 1, BCa limits are inaccurate, and may be ill-defined.
If ill-defined the largest or smallest bootstrap replicate is returned.
A warning about accuracy is provided if the adjusted probability levels
and number of bootstrap replications are such that the limits are
determined by a small number (currently 5) of the smallest or largest bootstrap
replicates.
The values of
z0
and
acceleration
estimated by
and the
acceleration
value estimated
by this function based on calling
or using the
L
component of
boot.obj
assume that the bootstrap sample size is the number of observations
in the original data -- i.e. that you are not using the
size
argument to the
sampler
passed to
.
Multiply those values by
sqrt(n/size)
to correct for the
size
argument.
This function assumes that
boot.obj
was created by ordinary
bootstrapping, or something similar (the
sampler
argument to
bootstrap could be
samp.bootstrap
,
samp.boot.bal
, or
samp.bootknife
, with
size=n
).
If these assumptions are violated the intervals may be incorrect.
Bootstrap BC (bias-corrected) intervals can be obtained by setting
acceleration=0
.
z0
and
acceleration
should have length equal
to the length of the
observed
component of
boot.obj
.
Davison, A.C. and Hinkley, D.V. (1997), Bootstrap Methods and Their Application, Cambridge University Press.
Efron, B. and Tibshirani, R.J. (1993), An Introduction to the Bootstrap, San Francisco: Chapman & Hall.
temp <- bootstrap(stack.loss, var) limits.bca(temp) limits.bca(temp, details=T) infl <- influence(stack.loss, var) limits.bca(temp, z0 = infl$estimate$z0, acceleration = infl$estimate$acceleration)