LoginSign Up
← Back to all sparks
R

RBesT

ANALYTICS
Velocity0.0

Tool-set to support Bayesian evidence synthesis in R

github.com/Novartis/RBesT ↗

RBesT is teaching its Bayesian decision rules to answer two-sided questions.

bayesian-statisticsclinical-trialsstanr-languageprior-elicitation
◆Current state
RBesT builds meta-analytic-predictive priors — the machinery for borrowing historical control data into a new trial — and evaluates the operating characteristics of decisions made with them. The stable line has spent several releases on effective sample size: ESS for normal mixtures via a new `family` argument, boundary corrections when no responses or no non-responses are observed, and stabilised ELIR computations. The 1.9-0 release candidate extends the normal, binomial and Poisson outcome functions to two-sided decisions.
◆Where it's heading
Two currents run through the changelog. One is ESS hardening — nearly every release since 1.7-4 fixes another edge case where the ELIR calculation aborted or returned something unstable, which is what happens when a quantity used to justify prior strength to regulators gets scrutinised. The other is Stan and brms integration debt: array syntax updates, a minimum Stan version bump, truncated prior generation for `mixstanvar`, deterministic EM. The RC's contributor list shows a second active maintainer, and the work is broader than any recent stable release.
◆Prediction
The release candidate covers all three outcome families and has already absorbed a round of review comments, so the next step is most likely the 1.9-0 CRAN release itself rather than further feature work.

◆Recent moves

  1. 5mo ago

    Two-sided decisions across normal, binomial and Poisson outcomes

    ⚡ SPARK

    The largest change in the visible history, and it is consistent across all three outcome families rather than piecemeal. Decision functions gaining a two-sided form widens which trial questions RBesT can express, not just how accurately it answers the existing ones — a different kind of release from the ESS repair work that preceded it.

    View source ↗
  2. 1y ago

    JSON read and write for mixture objects

    Adds JSON serialisation alongside smaller fixes. Modest in size, but it lets a fitted prior move between sessions, languages and pipelines as data rather than as an R object — a practical requirement when the prior has to be documented and reviewed outside the analysis that produced it.

    View source ↗
  3. 1y ago

    ess() fixed inside apply functions

    A single-issue patch for `ess()` failing when called inside the apply family. One more entry in the long ESS repair sequence, but narrow.

    View source ↗
  4. 1y ago

    ESS for normal mixtures in the exponential family

    A `family` argument lets effective sample size be computed for normal mixture densities in a one-parameter exponential family context — the case that covers a logit-transformed response scale. Also corrects `BinaryExactCI` at the boundary when no responses or no non-responses are observed, and stabilises the beta mixture information function behind ELIR.

    View source ↗
  5. 1y ago

    Truncated mixture priors for brms, plus faster Stan models

    `mixstanvar` gains automatic distribution function generation so truncated priors work with mixtures in brms, and the Stan likelihood is constructed more efficiently with smaller `gMAP` objects. The brms integration is how most users reach RBesT priors from a model they already wrote, so filling in truncation removes a real dead end.

    View source ↗
  6. 2y ago

    Stan array syntax update and CRAN system requirements

    Compliance and maintenance: system requirements now list pngquant, Stan sources move to the new array syntax with a 2.26.0 minimum, and most vignettes relocate to the package homepage to shrink the distributed tarball.

    View source ↗