The hardware and bandwidth for this mirror is donated by dogado GmbH, the Webhosting and Full Service-Cloud Provider. Check out our Wordpress Tutorial.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]dogado.de.

FusionForests 1.1.1

• The sampler for the centred error mixture is now exact. It samples the uncentred (parameter-expanded) mixture, whose full conditionals are all conjugate, and centres every stored draw afterwards (Yang et al., 2010): the RCT mixture mean is added to m0_sh and the difference of the two source means to d; the atoms are shifted to mean zero. Previously the forests conditioned on the centred atoms while the atoms and weights were updated as in the uncentred model. tau and c are unaffected; m0_sh and d now carry the source-specific error means. Draws are not bit-identical to 1.1.0.
• The baseline forest now fits the current augmented outcome; it used to see the outcome from one sweep earlier.
• dp_mu_g is returned for every mixture error model, and the mixture outputs are documented in full.
• With k_deviation = 0 and a source-specific error model, d equals the constant difference of the source-specific error means.

FusionForests 1.1.0

• New data sets breast_fusion and breast_fusion_truth.
• The k_* arguments now follow the paper’s scaling σ_h = k / (2√J), with defaults k = 1 and k_treat = 0.5. Double old k_* values to reproduce earlier fits.
• The default error_dist is now "source_hdp_scale".
• New error_dist = "source_gaussian": a normal error with a separate scale per source.
• fusion_estimand() now respects treatment_coding and per-source error scales.
• The survival difference for target_source = "rwd" is now causal: it uses tau only and no longer adds the confounding function c.
• Removed SimpleBART(), SimpleBCF(), the decomposition argument and the "RMST" estimand.
• FusionForest() now errors on missing covariate values.
• Documentation in the paper’s notation, code cleanup and more tests.

FusionForests 1.0.1

• Internal cleanup of the C++ layer for CRAN portability: removed unused standard-library includes, routed all console output explicitly through Rcpp::Rcout, and restricted the shared library’s symbol visibility to the registered entry points.

FusionForests 1.0.0

First release of FusionForests.

New model: FusionForest()

FusionForest() is a Bayesian tree ensemble for combining data from a randomised controlled trial (RCT) and real-world data (RWD).

The outcome (continuous, or log survival time) is decomposed over separate tree forests: a control forest, a treatment forest, a deconfounding forest and a deviation forest that captures how the treatment effect in the RWD deviates from the RCT. The observational data are not assumed to be unconfounded.

Continuous, right-censored and interval-censored outcomes are supported via an accelerated failure time formulation, with Gaussian or Dirichlet-process mixture error distributions (error_dist).

Posterior summaries

• fusion_estimand() — posterior draws of causal survival estimands: survival difference and acceleration factor.
• fusion_projection() — interpretable linear projections of the posterior treatment effect surface.
• print() and summary() methods for FusionForest fits.

Additional models

• SimpleBART() — single-forest BART.
• SimpleBCF() — Bayesian causal forest with prognostic and treatment forests.

Re-exports

The single-study causal and survival models from the ShrinkageTrees package (ShrinkageTrees(), HorseTrees(), CausalShrinkageForest(), CausalHorseForest(), SurvivalBART(), SurvivalDART(), SurvivalBCF(), SurvivalShrinkageBCF()) are re-exported, so library(FusionForests) provides every model from the accompanying paper in one namespace.

These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.
Health stats visible at Monitor.