The boot.sample component returned by the bootstrap
functions has been restructured. For every bootstrap replicate it now
reports the row id of each selected unit in the original sample
(row.id) and the bootstrap weight used to compute the
bootstrap statistic (bootstrap.weight). The within-stratum
index previously returned as ID is replaced by
row.id, which identifies a unit in the original sample and
can therefore be used to attach any variable of the sample.
The values of the study variable are returned as
bootstrap.sample for the methods that resample units and as
study.variable for the bootstrap weight methods, where the
original values are kept and only the weights change.
In boot.weights.stsrs, and in
boot.twostage under a stratified design, the columns of
boot.sample were returned as character strings. They are
now returned with their own type, so that the bootstrap weights can be
used directly in computations.
New function boot.replicates() returns the bootstrap
replicate weights as a matrix, with one row per unit of the original
sample and one column per bootstrap replicate. The rows follow the order
of the original sample, so the bootstrap variance of any estimator is
obtained by computing the estimator with each column of weights. This
allows bootstrap variances to be estimated for, among others, a vector
of study variables, a ratio, a calibrated estimator, and estimates
within sub-groups (domains), and reproduces the bootstrap statistics of
the package exactly. It applies to the direct, pseudo-population and
bootstrap weight methods, and to the method of Rao and Wu (1988) under
simple random sampling; under a stratified design that method is
reproduced directly from boot.sample and
rescale.factor, and under the two-stage design of
boot.twostage() the study variable can be reproduced from
the rescaled values stored in boot.sample, but the
rescaling cannot be reconstructed for an arbitrary new
variable.
The names of the columns of data are no longer
fixed. The study variable, the strata, the clusters and the first-stage
inclusion probabilities are identified through the arguments
study.variable, stratum, cluster
and Pi1. The defaults for study.variable and
cluster match the column names used previously, while
stratum and Pi1 default to NULL;
a stratified design is therefore specified by passing
stratum the name of the stratum column.
The bootstrap functions return sample.size, the
number of sampled units in each stratum. direct.boot.stsrs
and boot.twostage also return rescale.factor,
the rescaling factor applied to the study variable by the method of Rao
and Wu (1988). For direct.boot.stsrs, row.id
and rescale.factor allow the bootstrap to be reproduced for
another study variable. Under the two-stage design of
boot.twostage(), the study variable can be reproduced from
the rescaled values stored in boot.sample, but the
rescaling cannot be reconstructed for an arbitrary new
variable.
The bootstrap functions now perform additional input validation
and provide informative error messages for invalid arguments. These
checks include a non-integer or non-positive number of replicates
(R, R.pop, R.samp), first-stage
inclusion probabilities Pi1 outside (0, 1] or
not constant within a cluster, a population.size that does
not match the number of strata or is smaller than the sampled size, and
missing values in the study variable, the clusters or the
strata.
direct.boot.stsrs returned incorrect results under
stratified simple random sampling. The rescaled values were taken over
the whole sample rather than within the stratum being resampled, so the
values of one stratum were combined with the survey weights of another.
Results under simple random sampling were not affected.
pseudopop.boot.stsrs stopped with an error when the
population mean was requested under stratified simple random
sampling.
boot.twostage returned an incorrect bootstrap mean
for the method of Rao and Wu (1988) under a stratified design. The
bootstrap variance was not affected.
The method of Chauvet (2007) now runs under conditional Poisson
sampling in boot.twostage.
For the method of Chauvet (2007), boot.twostage now
computes the second-stage inclusion probability of the bootstrap sample
from the pseudo-population, under both simple random sampling and
conditional Poisson sampling.
direct.boot.stsrs is more than 20 times faster. The
bootstrap indices within a stratum are now drawn in a single call rather
than one unit at a time.