Skip to contents

Names kept reachable for one release past their removal or rename, each pointing at its successor. They are removed in dbarts 1.1-0.

bart2 is the former name of the modern front door, which is now bart; a bart2 call is forwarded, with a warning the first time in a session. Its one default that differs from bart's is combineChains = FALSE, as in 0.9-34's bart2, so a multi-chain fit keeps its chains as a margin of its draws. The fit's stored call is the bart2 call as written.

rbart_vi and its methods are deprecated, and documented on rbart_vi.

Details

The rest of the registry is names still reachable on entry points that keep their own page, rather than functions of their own: each is accepted for one release, mapped onto its successor with a once-per-session warning of class dbartsDeprecatedWarning, which inherits dbartsWarning (or refused by name, where noted), and removed in dbarts 1.1-0.

rngSeed

On bart and dbartsControl: the engine seed, now seed.

sigma

On dbarts and xbart: the creation-time residual estimate, now sigest; like sigest, it defaults to NULL.

node.prior

On dbarts: the leaf-value prior, now leaf.prior.

degreesOfFreedom

On dbartsPriors's chi(): the degrees of freedom of the hyperprior on k, now df, as invchi() and base R's dchisq spell it. Positional calls, chi(1.5, 2), are unchanged.

tree.prior

On dbarts and dbartsSpec: the tree prior, now power, base and split.probs, as bart takes them (tree.prior = cgm(power, base, split.probs) is power, base and split.probs, and tree.prior = dart(a, b) is split.probs = dart(a, b)). A tree prior object saved by 0.9-34 is read the same way. The value is used, with a once-per-session warning. bart and xbart never took it and refuse it, naming the plain spelling.

resid.prior, sigdf, sigquant

The residual scale's prior is a setting of the families that draw one, so it rides the family object (family = gaussian(sigma = chisq(df, quant)), family = gaussian(sigma = fixed(value))); likewise student(df, sigma = ), aft(sigma = ) and hurdle.lognormal(sigma = ). sigdf and sigquant are retired on bart; resid.prior is retired the same way on bart, dbarts and xbart, and dbartsSpec refuses it, so the prior has one home on every door. A call that writes it both ways - a retired spelling and a family whose call named sigma - is refused where the two disagree. sigest, the residual-scale estimate supplied at creation, is unaffected.

proposal.probs

On bart and dbarts: the tree-move mixture, now a dbartsControl slot and formal (control = dbartsControl(proposal.probs = )).

n.trees, n.cuts, useQuantiles on dbartsControl

What a model and its data hold: n.trees is now an argument of dbarts, dbartsSpec and forest, and n.cuts and useQuantiles of dbarts and dbartsData. A value written on a control is read after a once-per-session warning; a control's value that differs from what a call takes draws a warning on every call, and a value a fit left on its control states nothing.

sigest beside a fixed residual scale

A fit that states sigest and a residual scale held fixed (family = gaussian(sigma = fixed(value))) runs, sigest having no effect, after a once-per-session warning; it is an error from dbarts 1.1-0, so drop sigest.

The n.trees, useQuantiles and binary slots of a dbartsControl object

Still readable, as control@n.trees, control@useQuantiles and control@binary, for code that reads them from the control; the values now live on the model and the data: the forests' tree counts in the model's forests, useQuantiles in the data object's useQuantiles slot, and the response family in the model's family slot. A later release removes them.

A three-element n.burn on xbart

Refused outright, naming n.burn = c(fresh, warm): 0.9-x's third element was a per-replication burn-in, and a chain is never carried between replications now.

dbartsSampler's $startThreads/$stopThreads

No-op methods (dbartsSampler): the hierarchical thread manager they drove is gone, threads now owned per run, but a 0.9-x Gibbs loop that brackets its sweeps with them still runs.

dbartsSampler's $sampleNodeParametersFromPrior

Forwarded to $sampleLeafParametersFromPrior (dbartsSampler) after a once-per-session warning.

A thread count on dbartsSampler's $run

Ignored after a once-per-session warning, whether passed as n.threads, as 0.9-x's formal numThreads, or positionally after updateState: a run uses the sampler's own count, which $setControl changes, and its draws do not depend on it.

A logical updateCutPoints on dbartsSampler's $setPredictor

Read after a once-per-session warning, TRUE as "position" and FALSE as "none": the argument is now one of "none", "position" and "value" (dbartsSampler), and "position" is what TRUE did.

NA where NULL now means not given

NULL, not NA, is the one spelling of an absent value. Where 0.9-34's own default was NA, an explicit NA is still read as NULL after a once-per-session warning: sigest on bart; seed on bart, bartBT and xbart; n.samples on dbartsControl; updateState on the dbartsSampler methods and updatePredictorPerObservationJointly; and numBurnIn and numSamples on $run. An argument new in 1.0-0 refuses an NA, naming NULL: sigest and seed on dbarts, sigest on xbart, seed on dbartsControl, and both on dbartsSpec, which is new, and seed on dbartsValidateComposition.

A BayesTree-spelled bart call

Forwarded to bartBT after a once-per-session warning, when the call carries a name from bartBT's own formals that bart does not take. A call that also names an argument bartBT does not take (data, n.samples, any modern spelling) fits under neither and is refused before it is forwarded, naming the BayesTree names given with their modern spellings and the others.

A fourth or later positional bart argument

Refused, naming both doors: 0.9-x's bart read a fourth positional argument as sigest, where bart's own fourth formal is subset.

The front-door startup message

Attaching the package prints one packageStartupMessage naming both doors, once per session (suppressPackageStartupMessages silences it); no successor, since nothing replaces it.

The front-door defaults message

A bart call that names no bartBT argument and whose first argument is not a formula - a 0.9-x call such as bart(x.train, y.train, x.test) - binds the same way under both doors, so nothing forwards it and it runs under bart's own defaults. The first such call in a session prints a message of class dbartsFrontDoorMessage naming the change and bartBT; suppressMessages silences it, and a call made from package code prints nothing.

A BayesTree-spelled pdbart or pd2bart call

Each name bartBT takes and bart does not is translated to bart's spelling, after a once-per-session warning per name and function that also fires for a call from package code; x.test and sampleronly are refused. A setting given under both spellings is refused.

The pdbart defaults message

The first pdbart or pd2bart call in a session that fits data and names no BayesTree argument prints a message of class dbartsFrontDoorMessage saying that the fit uses bart's defaults and that pdbart(bartBT(x, y, keeptrees = TRUE)) gives the 0.9-34 model; a call made from package code prints nothing.

Every entry above, bart2 and rbart_vi, is one row of dbarts:::dbartsTombstones, the registry this page documents by hand.

Value

bart2 returns what bart returns, with the chains separate unless combineChains = TRUE is given.

See also