brr: builds the replicate-weight matrix balanced repeated
replication needs – a variance-estimation technique from survey
statistics that works by refitting an estimator on nr
half-sample replicates of the data (nr, the number of
replications required, is determined from strata and
secu and reported back) and taking the spread across them as
the variance, rather than a model-based formula. This returns the
design itself – TDA's orthogonal (Hadamard-based) construction,
not practical to reproduce independently in R – so it can be applied
to whatever estimator is being replicated; actually refitting a model
nr times and computing the BRR variance from the results is
ordinary R code once the design is in hand, not shown here since it
depends entirely on what is being estimated.
Usage
tda_brr(strata, secu, dir = tempfile("tda"), ...)Arguments
- strata
number of strata.
- secu
number of sampling units per stratum;
2in the overwhelming majority of real BRR designs, and the only value tested here.- dir
working directory.
- ...
passed to
tda_run.
Value
A list: replications (nr, how many are needed),
and design, the strata by replications matrix
itself (TDA's description: design[i, j] is which
sampling unit of stratum i to use in replication j).
Details
Not every (strata, secu) combination succeeds: secu must
be at least 2 (one stratum, one unit gives nothing to replicate), and
at least one combination that satisfies every other documented
constraint (brr = 4,2) can still fail with “no success
in calculating orthogonal weights”: 2, 3, and 8 strata at
secu = 2 all succeed, 4 does not. This is a limitation in
TDA's Galois-field construction for
specific sizes, not something this wrapper can lift.
See also
Other descriptive statistics:
tda_cov(),
tda_dstat(),
tda_freq(),
tda_independence(),
tda_ineq(),
tda_loglin(),
tda_quant(),
tda_rcorr(),
tda_segr(),
tda_subm()