Overview
The described work focuses on Bayesian change-plane regression, a statistical method designed to identify heterogeneity in treatment effects. This approach models patient assignment to distinct groups based on whether a linear score derived from baseline covariates crosses a specific threshold. Unlike traditional hard-threshold methods that introduce nonregularity in likelihood-based inference, this framework employs a smoothed probit gate with a declared smoothing scale. The primary estimand reported is the smoothed subgroup rule.
Research Context
Treatment effect heterogeneity refers to variations in how patients respond to a given treatment. Change-plane regression offers an interpretable rule for delineating these differences, assigning patients to one of two groups according to a linear score of baseline covariates relative to a threshold. However, direct likelihood-based inference for such hard thresholds presents challenges due to its nonregular nature. This nonregularity invalidates standard large-sample approximations, complicating statistical inference related to treatment effects and subgroup identification.
Approach
The methodology introduces a Bayesian framework for change-plane regression. Instead of a sharp indicator function for the threshold crossing, the approach substitutes it with a smooth probit gate. The degree of this smoothing is determined by a declared smoothing scale. The smoothed subgroup rule is treated as the reported estimand within this framework.
Key theoretical properties characterize this approach:
- At a fixed smoothing scale, the posterior distribution satisfies a misspecified Bernstein–von Mises theorem. This theorem centers the posterior around the smoothed rule.
- Under a vanishing smoothing schedule, the boundary can be learned at a rate faster than the parametric rate.
- Within an explicit schedule window, both the regression block and the treatment-effect contrast satisfy a Bernstein–von Mises theorem. This theorem centers on the hard-threshold values and demonstrates known-subgroup oracle efficiency. This implies that effect inference incurs no first-order price for the introduced regularization.
- A decision-theoretic reporting protocol is incorporated, offering consistency guarantees. This protocol is designed to distinguish evidence of heterogeneity from the act of reporting a boundary.
Computational aspects of the method include:
- Latent-variable augmentation.
- A great-circle elliptical slice sampler.
- A normalized horseshoe prior.
Findings
The developed Bayesian framework addresses the nonregularity issue in likelihood-based inference for hard thresholds in change-plane regression. By replacing sharp indicators with smooth probit gates, the method supports valid large-sample approximations:
- At a fixed smoothing scale, the posterior distribution is centered around the smoothed rule, as described by a misspecified Bernstein–von Mises theorem.
- The boundary learning rate can exceed the parametric rate when a vanishing smoothing schedule is employed.
- Within a defined schedule window, the regression block and treatment-effect contrast exhibit known-subgroup oracle efficiency relative to the hard-threshold values. This implies that the regularization process does not impose a first-order penalty on the inference of effects.
- A decision-theoretic reporting protocol was established to ensure consistency, allowing for the differentiation between detecting heterogeneity and reporting a specific boundary.
Why This Matters
The ability to accurately model and infer treatment effect heterogeneity is relevant for improving patient stratification and personalized medicine. The proposed framework provides a statistically robust method for this analysis, addressing limitations of prior approaches related to nonregularity in inference. The preservation of known-subgroup oracle efficiency for effect inference suggests that the methodological advancements do not compromise the statistical power for estimating treatment effects.
The method was illustrated through simulations and an application to the PREMIER trial.