Horseshoe Forests for High-Dimensional Causal Survival Analysis
Pith reviewed 2026-05-19 02:16 UTC · model grok-4.3
The pith
A Bayesian tree ensemble places the horseshoe prior on step heights to estimate heterogeneous treatment effects from high-dimensional censored survival data.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The central claim is that placing a horseshoe prior on the step heights of a Bayesian tree ensemble, combined with a reversible-jump Gibbs sampler, yields accurate estimates of heterogeneous treatment effects in high-dimensional right-censored survival data even when effects are nonlinear and sparsity varies.
What carries the argument
Horseshoe prior placed directly on the step heights of the tree ensemble, which supplies global-local shrinkage for regularization instead of relying on tree structure for sparsity.
If this is right
- The model recovers treatment effects accurately across different sparsity levels in the covariates.
- It handles nonlinear treatment effect functions without requiring the trees alone to enforce sparsity.
- The sampler accommodates the non-conjugate horseshoe prior inside the tree-ensemble framework for censored outcomes.
- The method produces usable estimates when applied to high-dimensional genomic survival data such as pancreatic cancer records.
Where Pith is reading between the lines
- Similar shrinkage on step heights could be tested in tree ensembles for other censored outcomes beyond survival.
- The approach suggests a route to combine horseshoe regularization with other causal survival models that already use tree structures.
- Further checks could examine how the method scales when the number of covariates grows well beyond the sample size.
Load-bearing premise
The reversible-jump Gibbs sampler can efficiently explore the posterior when the horseshoe prior sits on tree step heights and the survival times are right-censored.
What would settle it
If simulations in high-dimensional settings with nonlinear treatment effects show that the method recovers the true effects no better than standard approaches, or if the PDAC re-analysis yields implausible effect estimates, the central performance claim would be undermined.
read the original abstract
We develop a Bayesian tree ensemble model to estimate heterogeneous treatment effects in censored survival data with high-dimensional covariates. Instead of imposing sparsity through the tree structure, we place a horseshoe prior directly on the step heights to achieve adaptive global-local shrinkage. This strategy allows flexible regularisation and reduces noise. We develop a reversible jump Gibbs sampler to accommodate the non-conjugate horseshoe prior within the tree ensemble framework. We show through extensive simulations that the method accurately estimates treatment effects in high-dimensional covariate spaces, at various sparsity levels, and under non-linear treatment effect functions. We further illustrate the practical utility of the proposed approach by a re-analysis of pancreatic ductal adenocarcinoma (PDAC) survival data from The Cancer Genome Atlas.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes Horseshoe Forests, a Bayesian tree ensemble for heterogeneous treatment effect estimation in high-dimensional right-censored survival data. Sparsity is induced via a horseshoe prior placed directly on tree step heights rather than through the tree structure itself. A reversible-jump Gibbs sampler is developed to sample from the resulting non-conjugate posterior. The central empirical claim is that extensive simulations demonstrate accurate recovery of treatment effects across sparsity levels and non-linear effect functions; a re-analysis of TCGA PDAC survival data is also presented.
Significance. If the sampler produces reliable posterior draws, the approach would offer a useful regularization strategy for causal survival analysis in p >> n regimes by combining the flexibility of tree ensembles with global-local shrinkage. The simulation design (varying sparsity and non-linearity) and the real-data illustration provide a reasonable test bed. Credit is given for explicitly addressing the non-conjugacy problem with a tailored reversible-jump sampler rather than relying on approximate methods.
major comments (1)
- The reversible-jump Gibbs sampler is introduced to accommodate the horseshoe prior on step heights under right-censoring, yet the manuscript reports no convergence diagnostics, effective sample sizes, Gelman-Rubin statistics, or multiple-chain comparisons for the high-dimensional simulation settings. Because the central claim rests on the accuracy of posterior-based estimates obtained from these simulations, the absence of such verification is load-bearing; poor mixing would mean the reported performance cannot be attributed to correct inference.
minor comments (2)
- The simulation section should explicitly state the number of Monte Carlo replications, the precise definitions of the performance metrics (e.g., bias, MSE, interval coverage) under censoring, and whether the same censoring mechanism is used in both data generation and estimation.
- Notation for the tree step heights and the horseshoe scale parameters could be introduced earlier and used consistently when describing the prior and the sampler.
Simulated Author's Rebuttal
We thank the referee for their careful reading and constructive feedback. The single major comment raises an important point about the need to demonstrate sampler reliability. We address it directly below and commit to revisions that will strengthen the empirical support for our claims.
read point-by-point responses
-
Referee: The reversible-jump Gibbs sampler is introduced to accommodate the horseshoe prior on step heights under right-censoring, yet the manuscript reports no convergence diagnostics, effective sample sizes, Gelman-Rubin statistics, or multiple-chain comparisons for the high-dimensional simulation settings. Because the central claim rests on the accuracy of posterior-based estimates obtained from these simulations, the absence of such verification is load-bearing; poor mixing would mean the reported performance cannot be attributed to correct inference.
Authors: We agree that explicit convergence diagnostics are necessary to substantiate the reliability of the posterior estimates, especially given the non-conjugate nature of the model and the high-dimensional regime. The original manuscript emphasized the sampler derivation and overall performance metrics but did not include these diagnostics, which was an oversight. In the revised version we will add: (i) Gelman-Rubin statistics and effective sample sizes computed from four independent chains for representative high-dimensional simulation scenarios; (ii) trace plots and autocorrelation functions for selected parameters; and (iii) a brief discussion of burn-in and thinning choices. These additions will directly address the concern that poor mixing could undermine the reported results. We do not claim the diagnostics were already performed; they will be newly generated and reported. revision: yes
Circularity Check
New Bayesian tree ensemble with horseshoe prior validated on independent simulations; no load-bearing self-referential reductions.
full rationale
The paper proposes a horseshoe-prior tree ensemble for heterogeneous treatment effects in right-censored survival data and introduces a reversible-jump Gibbs sampler for the non-conjugate prior. Performance claims rest on extensive simulations across sparsity levels and non-linear effects, which constitute external benchmarks rather than fits to the same data or self-citation chains. No equations reduce reported accuracy to a quantity defined by construction from the model inputs, and the sampler development is presented as a technical contribution tested empirically rather than assumed by definition. This yields a minor score reflecting normal self-citation of prior horseshoe or BART literature without circular load-bearing.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.