Hierarchical Piecewise Regression with Smoothed Change-Points

Fits Bayesian hierarchical piecewise regression models with multiple logistic-smoothed change-points. Non-linear parameters (change-point locations and transition sharpness) and linear parameters can each be conditioned on covariates and factors via flexible design matrices. A random-intercept structure is supported for any parameter. Spike-and-slab regularization is supported for selecting the number of breakpoints. Posterior inference uses a Metropolis-within-Gibbs sampler implemented in 'Rust' for speed. Methods are based on the smooth transition piecewise regression model of Bacon and Watts (1971) and variable selection spike-and-slab priors of Kuo and Mallick (1998) < https://www.jstor.org/stable/25053023>. Methods are described in Bindoff (2026) .

smoothbp

Fits smoothed hierarchical piecewise regression with multiple change-points using an optimised Metropolis-within-Gibbs sampler implemented in Rust.

Features

• Multi-Breakpoint Models: Support for an arbitrary number of change-points.
• Spike-and-Slab Regularization: Automatic selection of the number of breakpoints using smoothbp_ss().
• Flexible Predictors: Change-point locations (omega), slope changes (delta), and transition sharpness (rho) can all be conditioned on covariates.
• Hierarchical Structure: Random intercepts for all parameters.
• High Performance: Core MCMC engine written in Rust.

Installation

# Requires Rtools45 on Windows for Rust compilation
pak::pkg_install("ABindoff/smoothbp")