Coupled unbiased MCMC for estimating Shapley effects under dependent inputs
Abstract
In global sensitivity analysis, Shapley effects provide a game-theoretic allocation of output variance among potentially dependent inputs. Their computation typically requires repeated conditional sampling, which can be challenging even when independent samples from the joint input distribution can be generated exactly. This difficulty arises naturally in hierarchical models, where conditioning on arbitrary subsets need not follow the generative order. To efficiently estimate Shapley effects under such dependence between input variables, we propose a permutation-based framework using coupled unbiased Markov chain Monte Carlo (MCMC). Our proposed method combines two complementary coupling mechanisms: (i) a joint coupling of conditional distributions along a random permutation, designed to reduce the variance of estimated marginal contributions, and (ii) a meeting-based coupling of Markov chains that removes finite-time MCMC bias through a telescoping correction. Under suitable conditions, we prove that the resulting estimators for the unnormalized Shapley effects are unbiased, and have finite variance and finite expected computational cost. Numerical experiments with various test cases illustrate the computational efficiency of the proposed method.
Related
- SAUSS: Stochastic Approximation with Unbiased Simulated Scores for Limited Dependent Variable Models
- Quadratic Point Estimate Method for Uncertainty Quantification with Dependent Non-Gaussian Inputs
- Two-Point Dependent Wild Bootstrap for Weakly Dependent Estimating Equations
- Decay Rates and Domain Dependence of a Coupled Wave-Heat System with Spatially Dependent Heat Coefficients
- From Log-Odds to Shapley Values: An Explanatory Geometry for the Weighted Naive Bayes Classifier
BibTeX
@misc{hirono2026coupled,
title = {{Coupled unbiased MCMC for estimating Shapley effects under dependent inputs}},
author = {Masataka Hirono and Takashi Goda},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.36690},
eprint = {2610.36690},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.36690},
}
RIS
TY - UNPB
TI - Coupled unbiased MCMC for estimating Shapley effects under dependent inputs
AU - Masataka Hirono
AU - Takashi Goda
PY - 2026
DA - 2026/10/27
PB - arXiv
DO - 10.48550/arxiv.2610.36690
UR - https://arxiv.org/abs/2610.36690
AB - In global sensitivity analysis, Shapley effects provide a game-theoretic allocation of output variance among potentially dependent inputs. Their computation typically requires repeated conditional sampling, which can be challenging even when independent samples from the joint input distribution can be generated exactly. This difficulty arises naturally in hierarchical models, where conditioning on arbitrary subsets need not follow the generative order. To efficiently estimate Shapley effects under such dependence between input variables, we propose a permutation-based framework using coupled unbiased Markov chain Monte Carlo (MCMC). Our proposed method combines two complementary coupling mechanisms: (i) a joint coupling of conditional distributions along a random permutation, designed to reduce the variance of estimated marginal contributions, and (ii) a meeting-based coupling of Markov chains that removes finite-time MCMC bias through a telescoping correction. Under suitable conditions, we prove that the resulting estimators for the unnormalized Shapley effects are unbiased, and have finite variance and finite expected computational cost. Numerical experiments with various test cases illustrate the computational efficiency of the proposed method.
ER -