Lists the sampler, model-fit, and search-algorithm settings
for design(method = "BSDA") and its validation. These are the
deeper knobs a user rarely changes; the planning targets (measure,
measure_value, target_pow) stay in design's own
arguments.
Usage
bsda_control(
gwish_sampler = "direct",
gwish_tol = 1e-08,
gwish_iter = 500L,
gwish_burnin = 500L,
edge_threshold = 0.5,
fit_iterations = 10000L,
fit_burnin = 5000L,
alpha = 0.05,
H_scout = 50L,
n_scout = 6L,
n_main = 10L,
scout_frac = 1/3,
max_iter = 10L,
n_boot = 5000L,
eps = 0.001,
tol_frac = 0.01,
init = "empty",
verbose = FALSE,
seed = NULL
)Arguments
- gwish_sampler
G-Wishart sampler,
"direct"or"block".- gwish_tol, gwish_iter, gwish_burnin
Tolerance, iterations, and burn-in for the G-Wishart sampler.
- edge_threshold
Posterior inclusion probability above which an edge is selected. Default 0.5.
- fit_iterations, fit_burnin
MCMC length and burn-in for fitting each simulated study.
- alpha
Significance level used in the fit. Default 0.05.
- H_scout,
Number of outer Monte Carlo replications for the scout
- n_scout, n_main
Grid sizes for the scout and main search passes.
- scout_frac
Fraction of
fit_iterationsused in the scout pass.- max_iter
Maximum refine-evaluate-invert iterations.
- n_boot
Bootstrap resamples for the probit-inversion CI.
- eps
Numerical tolerance guarding the probit inversion.
- tol_frac
Tolerance fraction of the relative sample size difference used to stop the search. Default 0.01, matched to typical H = 50. Lower it if H raises substantially.
- init
Initial graph for the BDMCMC sampler; it can be one of:
"prior"(the elicited prior G graph),"empty"(empty graph),"random"(random edges drawn from the elicited PIPs), or"median"(median graph, built as PIP > 0.5).- verbose
Stream progress from the C++ routine. Default
FALSE.- seed
Random seed for reproducibility. Default
NULL(no seed).
See also
Other sample size planning:
design(),
design.ggm_elicited(),
power_curve(),
power_curve.ggm_elicited(),
validate(),
validate.ggm_design()