ShrinkageTrees 2.0.2

Bayesian bootstrap for the average treatment effect

summary(), plot(type = "ate"), and predict() for CausalShrinkageForest (and CausalHorseForest) now default to a Bayesian-bootstrap posterior for the average treatment effect: at each MCMC iteration the per-observation CATEs are reweighted with Dirichlet(1, …, 1) weights before being summed, giving a draw from the posterior of the population ATE (PATE). Credible intervals are correspondingly wider than before because they now propagate uncertainty in the covariate distribution, not only in tau(x).

This is a breaking change for printed/plotted numerics: existing scripts will report wider CIs than before. Use bayesian_bootstrap = FALSE to reproduce previous output.

ShrinkageTrees 2.0.1

ovarian dataset restructured

The ovarian dataset is now a single data frame (previously a list with $clinical and $X elements). Clinical columns (OS_time, OS_event, treatment, age, figo_stage, tumor_grade) and the 2000 gene expression columns are combined into one data frame with 2006 columns. This simplifies data access and aligns the format with the pdac dataset.

Code that previously used ovarian$clinical or ovarian$X must be updated — see ?ovarian for the new structure.

Bug fixes

ShrinkageTrees 2.0.0

TCGA ovarian cancer dataset (ovarian)

Added the ovarian dataset: a processed TCGA-OV cohort (n = 357) for high-dimensional survival prediction and causal inference.

See ?ovarian and examples/test-ovarian.R for a full worked example covering survival prediction (SurvivalBART, SurvivalDART, HorseTrees) and causal inference (SurvivalBCF, SurvivalShrinkageBCF, CausalHorseForest).

Treatment coding for causal models (treatment_coding)

All causal model functions — CausalHorseForest(), CausalShrinkageForest(), SurvivalBCF(), and SurvivalShrinkageBCF() — now accept a treatment_coding argument controlling how the treatment indicator enters the BCF decomposition y = f(x) + b * tau(x) + epsilon. Four options are available:

The predict() method for CausalShrinkageForest objects automatically carries forward the treatment coding used at training time. A propensity_test argument is available for supplying test-set propensity scores (defaults to 0.5).

Interval-censored survival outcomes

All survival-capable functions now support interval-censored data in addition to right-censored data. Supply left_time and right_time vectors (with outcome_type = "interval-censored") instead of y and status. Three censoring types are distinguished:

This convention follows survival::Surv(type = "interval2"). Censored event times are augmented within the AFT Gibbs sampler. The following functions are affected:

Multi-chain MCMC (n_chains)

All four primary model-fitting functions — ShrinkageTrees, HorseTrees, CausalHorseForest, and CausalShrinkageForest — now accept an n_chains argument (default 1). When n_chains > 1:

S3 classes and methods

MCMC convergence diagnostics (coda)

Posterior visualisation (plot)

S3 plot() methods added for ShrinkageTrees, CausalShrinkageForest, and ShrinkageTreesPrediction. Requires the suggested package ggplot2.

Vignette

Survival wrapper improvements

Bug fixes

ShrinkageTrees 1.2.0

ShrinkageTrees 1.1.0

ShrinkageTrees 1.0.3

ShrinkageTrees 1.0.2

ShrinkageTrees 1.0.1

ShrinkageTrees 1.0.0

🎉 First CRAN release of ShrinkageTrees!

This package provides Bayesian regression tree models with shrinkage priors, supporting:

It includes four core functions:

The ...Trees functions use a single learner to estimate the outcome model directly. In contrast, the Causal...Forest variants fit separate models for the treated and control regression function. This enables estimation of conditional average treatment effects (CATEs).