ShrinkageTrees
Overview
ShrinkageTrees provides a unified framework for survival analysis using Bayesian regression tree ensembles, with a particular focus on causal inference and high-dimensional data.
The package implements Horseshoe Trees, Causal Horseshoe Forests, and their more general counterparts — Shrinkage Trees and Causal Shrinkage Forests — alongside well-known Bayesian tree models such as BART, DART, and BCF. All models are adapted to right-censored data via accelerated failure time (AFT) formulations.
Its central methodological innovation is the Horseshoe regularisation mechanism applied directly to tree step heights, enabling adaptive global–local shrinkage in high-dimensional settings. In addition to classical BART priors, the package supports:
- Horseshoe priors
- Forest-wide Horseshoe shrinkage
- Empirical Bayes Horseshoe calibration
- Half-Cauchy priors
- Dirichlet splitting priors (DART)
These models enable flexible non-linear modelling for:
- High-dimensional prediction
- High-dimensional causal inference
- Estimation of heterogeneous (conditional average) treatment effects (CATE)
Supported outcome types:
- Continuous outcomes
- Binary outcomes
- Right-censored survival times (AFT framework)
All models are implemented with an efficient C++ backend via Rcpp, allowing scalable MCMC sampling in high-dimensional settings.
⭐ Core Contribution: Horseshoe Trees
Traditional BART and DART primarily regularise model complexity through the **tree structure (e.g., depth constraints or splitting probabilities).
ShrinkageTrees instead introduces global–local shrinkage directly on the leaf (step height) parameters via the Horseshoe prior.
A global parameter controls overall shrinkage, while local parameters allow strong signals to escape shrinkage. Small effects are aggressively shrunk toward zero, whereas large effects are preserved due to the heavy-tailed prior.
This strategy retains all covariates, reduces noise in high-dimensional settings, and improves robustness in causal models.
This methodology is introduced in:
Horseshoe Forests for High-Dimensional Causal Survival Analysis
T. Jacobs, W.N. van Wieringen, S.L. van der Pas
https://arxiv.org/abs/2507.22004