Bayesian Causal Forests for Zero-Inflated Continuous Outcomes

arXiv:2610.18889 · stat.ML, cs.LG, stat.ME · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Bayesian Causal Forests for Zero-Inflated Continuous Outcomes".

Jane: Comprehensive Research Summary: Zero-Inflated Continuous Bayesian Causal Forests with Duan’s Smearing (ZIC-BCF-Smear) This paper introduces Zero-Inflated Continuous Bayesian Causal Forests with Duan’s Smearing (ZIC-BCF-Smear), a novel,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Now, let’s look at what the paper actually says about the core mechanism of ZIC-BCF-Smear. Jane Can you explain that decoupling concept in terms of what it avoids? Tom They are avoiding data augmentation noise and posterior variance inflation that you often see when you try to model everything in one single joint model >

Lu: The main idea is fitting a Probit BCF on the full sample to model participation, and then fitting a Gaussian BCF strictly on the active subpopulation to model the continuous positive intensity >

Jane: So, if I’m driving and I want to know how this treatment affects me, is this method trying to give me two separate answers? Tom It’s essentially giving you a hurdle estimate for whether you even get a positive result, and then a separate estimate for the size of that positive effect >

Meng: But what about the estimation of the continuous part itself? That’s usually where things get messy when you have these complex distributions >

Tom: They introduce Duan’s non-parametric smearing estimator to handle that. Jane Smearing sounds like a transformation, but why do they need it instead of just using standard parametric multipliers? Tom Because those parametric scale multipliers are fragile; they break easily if the model assumptions are slightly off, so Duan’s estimator re-transforms the predictions back to the original response scale using an empirical factor >

Lalam: So it’s a distribution-free way to get rid of those fragile multipliers that might be wrong because we assumed the wrong shape for things like log-normal hurdles or Gamma hurdles >

Lu: That’s crucial because it bypasses the fragility of those parametric multipliers and makes the method much more robust under non-log-normal intensities >

The paper's summary: Tom: Let’s talk about the specific technical improvements they introduced beyond just decoupling. Jane Are there any other clever tweaks they made to make this work better in practice? Tom Yes, the authors introduced SPA, which we already touched on, but it’s a key part of the refinement >

Lu: SPA uses that active-subset propensity score as a regularizing control covariate in the prognostic forest >

Jane: So it’s about controlling for selection bias when you condition on being in that positive group, which is important for causal inference >

Meng: From an engineering standpoint, how does this decoupling help with computational cost? Tom By decoupling the process, they avoid data augmentation noise and they avoid the prohibitive complexity of BNP mixtures >

Meng: That’s good. BNP mixtures can be incredibly slow to train on large datasets when you're shipping models to production >

Jane: But what about when the data is really messy? The paper seems worried about that too, right? Tom They show they have exact Gibbs sampling for the joint posterior >

Lalam: That exact Gibbs sampler comes from updating the hurdle block and then the intensity block sequentially within each sweep, which keeps things computationally tractable while maintaining exactness for the joint posterior >

The paper's improvements: Tom: So to wrap up this paper, ZIC-BCF-Smear is a decoupled method designed for semicontinuous outcomes. Jane It seems like the main win here is that it’s stable and calibrated when compared against well-specified parametric hurdle models without committing to any specific parametric family >

Lu: They show it remains stable where competitors, like the Gamma hurdle benchmark, either diverge numerically or over-cover for nonlinear data >

Meng: It’s interesting that they can recover coverage gradients directly on the dollar-scale conditional effect in real-world data from the MEPS survey, which is a big win for practical application >

Jane: So what does this mean for someone who just listens to the show? Tom It means we have a tool that can give us a stable estimate of heterogeneous treatment effects even when the outcome isn't nice and clean like it is in many real-world medical situations >

Lalam: The paper "Bayesian Causal Forests for Zero-Inflated Continuous Outcomes" gives us this flexible framework to estimate those effects more reliably by focusing on what actually matters >

Lu: This method is valuable because it avoids those pitfalls of ignoring the zero mass or imposing fragile parametric assumptions on the outcome distribution >

Tom: It’s a lot of work, but it provides a stable and interpretable way to handle the hurdle part separately from the continuous intensity >

Conclusion: Tom: So we’ve been looking at "Bayesian Causal Forests for Zero-Inflated Continuous Outcomes," which is this method for estimating treatment effects in outcomes that have both a zero mass and a continuous positive part. Jane It seems like the core idea is this two-part structure, decoupling the hurdle from the intensity >

Lu: Exactly. They fit a Probit Bayesian Causal Forest on the full sample just to model that participation hurdle—the probability of getting exactly zero >

Meng: But fitting two separate models sounds like it could cause some noise issues when you combine them, right? Tom They address that with something called Subpopulation Propensity Adjustment, or SPA >

Jane: What exactly does SPA do for the active subpopulation? Tom It reestimates the treatment propensity specifically within that active subset and uses that result to control for selection bias in the continuous part of the forest >

Lalam: So, it’s like they are making sure that when you look at people who actually got a positive outcome, you aren't accidentally biasing your estimate because of how you filtered them in the first place >

Lu: It’s about mitigating regularization-induced confounding within that active subpopulation >

Jane: That makes sense for precision medicine where we need to know exactly who benefits from a treatment >

Tom: And they also use Duan’s non-parametric smearing estimator to handle the continuous part, which is really smart because it avoids fragile parametric multipliers >

Lalam: Smearing sounds like a transformation, but why do they need it instead of just using standard parametric multipliers? Tom Because those parametric scale multipliers are fragile; they break easily if the model assumptions are slightly off, so Duan’s estimator re-transforms the predictions back to the original response scale using an empirical factor >

Lu: So it’s a distribution-free way to get rid of those fragile multipliers that might be wrong because we assumed the wrong shape for things like log-normal hurdles or Gamma hurdles >

Jane: That’s crucial because it bypasses the fragility of those parametric multipliers and makes the method much more robust under non-log-normal intensities >

Meng: That’s good. BNP mixtures can be incredibly slow to train on large datasets when you're shipping models to production >

Tom: They show they have exact Gibbs sampling for the joint posterior, which keeps things computationally tractable while maintaining exactness for the joint posterior >

Lalam: That exact Gibbs sampler comes from updating the hurdle block and then the intensity block sequentially within each sweep >

Lu: This method is valuable because it avoids those pitfalls of ignoring the zero mass or imposing fragile parametric assumptions on the outcome distribution >

Tom: It’s a lot of work, but it provides a stable and interpretable way to handle the hurdle part separately from the continuous intensity >

Jane: So that's what "Bayesian Causal Forests for Zero-Inflated Continuous Outcomes" is about, giving us this flexible framework to estimate those effects more reliably by focusing on what actually matters >

Lu: This method is valuable because it avoids those pitfalls of ignoring the zero mass or imposing fragile parametric assumptions on the outcome distribution >

Meng: If we’re going to deploy this, we still need to watch out for the fact that there's mild under-coverage of CATE credible intervals when dealing with complex, non-linear conditional mean surfaces due to conservative regularizing priors >

Tom: That’s a necessary caveat. It means the estimates are good but we still can't guarantee one hundred percent precision on every single treatment effect >

Jane: So while it’s really useful for high-stakes causal inference, maybe the next step is focusing on fixed-horizon restricted mean or multistate formulations to sharpen those safety estimands and standardize follow-up time >

Lalam: Yeah, that makes sense. We need to get better at standardizing exposure and follow-up time for these kinds of estimates >

Lu: I think that’s the direction where we should head next, trying to move toward more practical applications of these models on the ground >

Hugo Gobato Souto, Eric Wang, Igor Shuryak

University of São Paulo · Columbia University Irving Medical Center

stat.ML, cs.LG, stat.ME

Submitted: 2026-10-08

Updated: 2026-10-08

Code: https://github.com/hugogobato/zicbcf

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: Comprehensive Research Summary: Zero-Inflated Continuous Bayesian Causal Forests with Duan’s Smearing (ZIC-BCF-Smear) This paper introduces Zero-Inflated Continuous Bayesian Causal Forests with

Key concepts

Zero-Inflated Continuous Outcomes: This type of data has two distinct parts: some patients have an exact zero value (like zero days of toxicity), while others have a continuous distribution for positive values. This complexity is common in real-world health data, making standard models inadequate.
Decoupled Two-Part Modeling: The method splits the estimation into two independent parts: first, a Probit forest models the probability of being in the zero group (the hurdle). Second, a Gaussian forest models only the intensity of positive outcomes for those who are not zero. This separation prevents noise from one part affecting the other.
Duan’s Smearing Re-transformation: This technique is used to convert predictions back to the original scale. Instead of using fragile parametric multipliers, this non-parametric method uses an empirical smearing factor to transform log-scale results back to the actual response scale, ensuring the final effect estimate is interpretable.
Subpopulation Propensity Adjustment (SPA): SPA adjusts treatment propensity specifically within the group that has positive outcomes. This step controls for confounding bias that arises when we only look at patients who actually experienced a positive result, making the analysis more accurate.

Summary

Comprehensive Research Summary: Zero-Inflated Continuous Bayesian Causal Forests with Duan’s Smearing (ZIC-BCF-Smear)

This paper introduces Zero-Inflated Continuous Bayesian Causal Forests with Duan’s Smearing (ZIC-BCF-Smear), a novel, decoupled two-part nonparametric estimation framework designed to accurately estimate heterogeneous treatment effects in outcomes that exhibit semicontinuous distributions—specifically those possessing both a discrete point mass at zero and a skewed continuous distribution for positive values. The authors address the significant shortcomings of existing methods, which typically either ignore the zero mass, impose fragile parametric assumptions, or suffer from prohibitive computational costs and data augmentation noise.

Core Methodology: Decoupling Hurdle and Intensity Processes

The ZIC-BCF-Smear method employs a sophisticated two-part Bayesian forest structure that fundamentally decouples the modeling of the participation hurdle from the modeling of the positive outcome intensity:

Hurdle Modeling (Participation): A Probit Bayesian Causal Forest (BCF) is fitted on the full sample to explicitly model and capture the zero-inflation hurdle—the probability that a patient contributes an exact zero to the outcome.
Intensity Modeling (Positive Outcome): A strictly Gaussian BCF is then fitted only on the active subpopulation (those with positive outcomes) to model the continuous, positive intensity of the effect.

This decoupling strategy is crucial as it avoids data augmentation noise and posterior variance inflation often associated with joint modeling approaches.

Key Innovations and Technical Components

The paper introduces three key technical innovations that enhance robustness, flexibility, and interpretability:

Subpopulation Propensity Adjustment (SPA): This component is introduced to mitigate regularization-induced confounding within the active subpopulation. SPA reestimates the treatment propensity specifically within this active subset and supplies it to the prognostic component of the continuous-stage forest, thereby controlling for selection bias induced by conditioning on a post-treatment active status.
Non-Parametric Duan’s Smearing Re-transformation: To bypass the fragility of parametric scale multipliers (which are highly sensitive to model misspecification), Duan’s non-parametric smearing estimator is incorporated. This estimator re-transforms the predictions from the log scale back to the original response scale using an empirical, distribution-free smearing factor, effectively replacing fragile parametric multipliers with a more robust transformation.
Exact Gibbs Sampling: The joint posterior is fitted using a single Gibbs sampler where the hurdle block and the intensity block are updated sequentially within each sweep. This results in an exact Gibbs sampler for the joint posterior due to the factorization of the two-part likelihood, ensuring computational tractability while maintaining exactness.

Validation and Performance Benchmarks

The method's efficacy is rigorously validated through extensive simulation benchmarks across three distinct Data Generating Processes (DGPs) and link regimes (including log-normal hurdles, Gamma hurdles, and Tweedie distributions).

Superior Stability: ZIC-BCF-Smear is shown to match or outperform well-specified parametric hurdle models without committing to any specific parametric family. Critically, it remains stable and calibrated where competitors—such as the Gamma hurdle benchmark—either diverge numerically or pathologically over-cover for nonlinear data.
Robustness to Misspecification: The authors demonstrate that ZIC-BCF-Smear is categorically more stable under link non-linearity than parametric Gamma positive-intensity likelihoods, which are known to break down in such regimes.
Scale Recovery: In applications like dental expenditures from the Medical Expenditure Panel Survey (MEPS), ZIC-BCF-Smear recovers documented coverage gradients directly on the dollar-scale conditional effect—a feat that a parametric Gamma hurdle benchmark fails to achieve, which only reproduces these gradients on its participation margin.

Applications and Causal Inference Findings

The method is applied to real-world data from the SPECTRUM randomized phase III head-and-neck cancer trial, estimating the causal effect of assignment to panitumumab plus chemotherapy on the number of days patients spend with severe toxicity.

Clinical Interpretation: The results credibly show that assignment to panitumumab plus chemotherapy increases this severe-toxicity burden by about ten days per patient under the trial’s own differential treatment duration.
Decomposition Insight: A key finding is that the two-part decomposition localizes this effect mainly in the positive-intensity part of the model rather than solely in whether the reconstructed severe-AE burden is positive at all.

Limitations and Future Directions

Despite its significant advancements, the authors acknowledge several limitations:

Under-coverage: There is mild under-coverage of CATE credible intervals when dealing with complex, non-linear conditional mean surfaces due to conservative regularizing priors.
Assumptions: The model relies on an assumption of homoskedastic log-scale errors for the Stage 2 intensity component and uses plug-in estimates for propensity scores fixed before the sampler begins.
Estimand Focus: The estimator targets a contrast in the cumulative outcome as recorded, lacking a mechanism to standardize exposure or follow-up time.

The authors suggest future work should focus on fixed-horizon restricted mean or multistate formulations to sharpen safety estimands by standardizing follow-up time and accommodating competing events.

Conclusion for High-Stakes Research

ZIC-BCF-Smear represents a significant methodological leap in causal inference for semicontinuous outcomes. By intelligently decoupling the zero hurdle from the continuous intensity process and employing distribution-free re-transformation techniques, it provides a flexible, stable, and clinically interpretable estimator that avoids the pitfalls of both overly restrictive parametric models and computationally intractable nonparametric alternatives. Its demonstrated ability to recover known heterogeneity on the correct response scale makes it a highly valuable tool for high-stakes causal effect estimation.

Improvements for AI systems

Bold header: Decoupled Hurdle-Intensity Modeling for Semicontinuous Outcomes. This allows AI systems to estimate Conditional Average Treatment Effects (CATEs) on complex outcomes like severe toxicity burden by fitting a Probit BCF on the full sample to model participation and a Gaussian BCF strictly on the active subpopulation.
Bold header: Robustness Against Scale Misspecification. The method utilizes Duan’s Non-Parametric Smearing Estimator to map predictions back to the response scale, which replaces the fragile parametric scale multiplier exp(σ 2/2) with a distribution-free multiplier, resolving scale misspecification bias under non-log-normal intensities.
Bold header: Mitigating Selection Bias in Active Subpopulations. The inclusion of Subpopulation Propensity Adjustment (SPA) mitigates regularization-induced confounding by using the active-subset propensity score as a regularizing control covariate in the prognostic forest, which is crucial for accurate causal inference when conditioning on post-treatment status.
Bold header: Handling High Computational Complexity Efficiently. By decoupling the process, the method avoids data augmentation noise and avoids the prohibitive complexity of BNP mixtures, making it feasible to estimate heterogeneous treatment effects where traditional methods suffer from extreme computational costs and data augmentation noise.
Bold header: Valid Causal Inference on Real-World Data. The estimator is applied to individual-participant data from the SPECTRUM trial, allowing for a causal effect estimation of the number of days patients spend with severe toxicity, which is then interpreted in the context of the trial’s own differential treatment duration.

Abstract

Estimating (heterogeneous) treatment effects is still challenging when the outcome is semicontinuous (i.e., exhibiting both a discrete point mass at zero and a skewed continuous distribution for positive values), despite the strong presence of this data type in real world settings. Existing methods either ignore the zero mass, impose fragile parametric assumptions, or suffer from extreme computational costs and data augmentation noise. This paper proposes Zero-Inflated Continuous Bayesian Causal Forests with Duan's Smearing (ZIC-BCF-Smear), a decoupled two-part nonparametric estimator that fits a Probit BCF on the full sample to model participation and a Gaussian BCF strictly on the active subpopulation to model positive intensity. Subpopulation Propensity Adjustment (SPA) mitigates regularization-induced confounding within the active subpopulation, and Duan's non-parametric smearing estimator re-transforms predictions to the response scale, bypassing the fragile scale multipliers of parametric models. A comprehensive simulation benchmark shows that ZIC-BCF-Smear matches well-specified parametric models without their assumptions and remains stable and calibrated where competitors either diverge or pathologically over-cover for nonlinear data. The method is then applied to individual-participant data from the SPECTRUM randomized phase III head-and-neck cancer trial, estimating the causal effect of assignment to panitumumab plus chemotherapy on the number of days patients spend with severe toxicity, an outcome for which a quarter of patients contribute an exact zero. Panitumumab assignment raises that burden by about ten days per patient under the trial's own differential treatment duration, and the decoupled structure localizes the effect mainly in the positive-intensity part rather than in whether the reconstructed severe-AE burden is positive.

Sources

A Bayesian Nonparametric Model for Zero-Inflated Outcomes: Prediction, Clustering, and Causal Estimation

Listen

Transcript

Key concepts

Summary

Improvements for AI systems

Abstract

Sources

Related papers