Options for the RoBMA package

Inspect or change package-level defaults used by RoBMA.

Usage

RoBMA.options(...)

RoBMA.get_option(name)

Arguments

...

named option(s) to change. Names must be exact public option names, nonempty, and unique.

name

a single non-missing character string matching one public option exactly; for available options, see details below.

Value

RoBMA.options() invisibly returns a named list with all current options after applying any changes. RoBMA.get_option() returns the current value of the requested option.

Details

The available options are:

max_cores

number of cores to use for parallel computing (default is one fewer than detected logical cores, with a minimum/fallback of 1)

native_threads

maximum number of native threads used to evaluate independent posterior rows of the compiled selection-likelihood batches in post-processing such as likelihood-aware densities, z-plots, and predictions. The default NA inherits the model's fitting setup: models fitted with parallel = TRUE evaluate their post-processing rows with up to max_cores native threads, unless this option sets an explicit thread count. Parallel work is arranged so that every reduction sums in a fixed order, so results do not depend on the thread count; single-state fitting and bridge-sampling evaluations stay serial. A model fitted with parallel = FALSE does its post-processing serially unless this option sets an explicit thread count.

check_scaling

whether to check scaling of predictors (default TRUE)

silent

whether to suppress JAGS output when fitting or extending models (default FALSE); the model-averaging constructors RoBMA(), BMA(), BMA.norm(), BMA.glmm(), RoBMA.mv(), and BMA.mv() default to silent = TRUE and ignore the option

jags.worker_output

file path for parallel JAGS worker stdout and stderr when fitting or extending. The parent directory must exist; workers append to the same file and messages may interleave. The default empty string disables capture. Set this option inside any background job that fits the model. It does not change sampling or numerical integration settings.

autocompute.loo

whether to automatically compute LOO (default FALSE); a GLMM whose nuisance-prior likelihood the quadrature cannot certify warns and the fit is returned without it

autocompute.waic

whether to automatically compute WAIC (default FALSE); a GLMM whose nuisance-prior likelihood the quadrature cannot certify warns and the fit is returned without it

autocompute.marglik

whether to automatically compute marginal likelihood (default FALSE)

cluster_likelihood.n_gamma

number of Gauss-Hermite nodes formerly used for weighted normal cluster-unit log-likelihoods (default 15). Normal cluster-unit log-likelihoods, weighted or not, are now integrated in closed form, so the option is still accepted but no longer affects them; selected-normal likelihoods use their own quadrature schedules and integration controls.

selection.cache_max_bytes

one shared storage budget for exact and coarse selection-normalizer caches across all chains in a fit. The default "auto" uses the smaller of 4 GiB and one quarter of currently available system RAM, resolved before fitting. Supply a nonnegative whole number of bytes to set an explicit total; 0 disables caching. Storage grows as needed. Chains receive equal shares, pooled within each worker process; sequential chains share the total.

selection.cache_retain

whether to save selection-normalizer cache entries with the fitted object after fitting or extending (default FALSE). Retained entries survive saveRDS() and are restored when extending a compatible fit. See remove_selection_cache for memory costs and removal.

selection.sampler

"default" retains ordinary JAGS sampling. "coarse_corrected" uses coarse slice proposals with a full-target Metropolis correction for eligible scalar updates involving multivariate, cluster or factor step selection blocks (see selection_sampler_info for the blocks that take part). The posterior target is unchanged. Settings are captured when each model is compiled.

selection.coarse_grid

named nonnegative grid steps c(mean = .05, variance = .05, log_weight = .01) of the dense multivariate step blocks; cluster and factor blocks evaluate their coarse rule at the actual parameters and ignore the grid. Mean spacing is its step times the largest original sampling SE in the block; covariance-diagonal spacing is its step times that SE squared. Each diagonal is rounded upward on a grid whose origin is its own squared original sampling SE, retaining all off-diagonal entries. Positive weights are rounded on the natural-log scale; zero weights remain zero. A zero step retains that coordinate exactly but does not disable the coarse rule budget or fallback. These settings affect proposals, not the corrected target.

selection.coarse_max_rules

coarse quadrature rule budget (default 3, minimum 3). Dense multivariate blocks try at most this many rules from the beginning of the fitted quadrature schedule for a coarse anchor; unsupported or unsuccessful coarse integration uses the prescribed unit normalizer, without QMC fallback. Cluster and factor blocks use the one rule of this number in their own schedule (the last rule of a shorter schedule) at the actual parameters, and the exact density where that rule is unavailable. The full-target evaluator retains its ordinary numerical controls and diagnostics.

default_UISD.effect

default scaling of the unit information standard deviation for the effect size parameter (default 0.5)

default_UISD.heterogeneity

default scaling of the unit information standard deviation for the heterogeneity parameter (default 0.25)

default_UISD.mods

default scaling of the unit information standard deviation for the moderators (default 0.25)

default_UISD.scale

default scaling of the unit information standard deviation for the scale parameter (default 0.5)

default_lograte.sd

default standard deviation of the estimate-specific midpoint log-rate prior in Poisson GLMMs (default 1)

default_informed_priors.mods

default scaling of informed priors for moderators (default 0.5)

default_informed_priors.scale

default scaling of informed priors for the scale parameter (default 0.5)

default_bias_weightfunction.alpha

default alpha for the weightfunction (default 1)

default_bias_PET.scale

default scale for the PET (default 1)

default_bias_PEESE.scale

default scale for the PEESE (default 5)

Boolean options require scalar TRUE or FALSE values. max_cores must be integer-like and at least 1; cluster_likelihood.n_gamma must be integer-like and at least 3. native_threads accepts NA or a positive integer. Scale and alpha defaults must be finite positive numbers.