Where the Data End and the Mechanism Begins: Doubly Robust Causal Digital Twins for Treatment Strategies
Abstract
Digital twins are increasingly used to ask how a patient's health would evolve under alternative treatment strategies. When a twin couples a mechanistic model with learned components and is fitted to observational records, its answer mixes information from the data with assumptions from the mechanism. For sustained strategies that begin at a clinical landmark, we split the counterfactual mean into a part the data support, where patients follow the strategy with probability bounded away from zero, and a part they do not. On the supported part, a cross-fitted sequential doubly robust estimator that uses the twin as its outcome model is consistent whenever the decision model is, so twins that fit equally well but disagree as plug-in estimators agree after correction. On the unsupported part, we characterize the values compatible with the data and with a residual of size at most epsilon around the mechanism. This set is nested in epsilon, reduces to a point when the mechanism is exact and identifiable, and recovers assumption-free bounds as epsilon grows, so the penalty weight of a physics-informed loss becomes a sensitivity parameter. Knowledge constraints on how covariates enter the mechanism shrink the set, and when support grows over time, back-testing gives a data-based lower bound on epsilon. In simulations based on an energy-balance model of weight change under glucagon-like peptide-1 receptor agonist therapy, the correction reduced the spread of estimates across four twins from 0.66 to 0.03 percentage points of weight loss with conservative coverage; on unsupported patients, the exact mechanism was rejected in every replicate, the bounds covered the truth at every residual size not rejected, and a mechanism-free extrapolation was biased by 0.5 points.
Related
- Doubly-robust causal inference for multi-valued treatment regimes with recurrent neural networks
- Doubly robust augmented weighting estimators for the analysis of externally controlled single-arm trials and unanchored indirect treatment comparisons
- Active learning confidence measures for coupling strategies in digital twins integrating simulation and data-driven submodels
- Difference-in-Differences for Heterogeneous Effects of General Treatment Strategies Accounting for Time-Varying Covariates
- Evaluation of digital binocular therapies for amblyopia after refractive adaptation in treatment-naive children aged 3–13 years: A systematic review and meta-analysis of randomised controlled trials
BibTeX
@misc{cao2026where,
title = {{Where the Data End and the Mechanism Begins: Doubly Robust Causal Digital Twins for Treatment Strategies}},
author = {Cong Cao},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.31706},
eprint = {2610.31706},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.31706},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Where the Data End and the Mechanism Begins: Doubly Robust Causal Digital Twins for Treatment Strategies
AU - Cong Cao
PY - 2026
DA - 2026/10/21
PB - arXiv
DO - 10.48550/arxiv.2610.31706
UR - https://arxiv.org/abs/2610.31706
AB - Digital twins are increasingly used to ask how a patient's health would evolve under alternative treatment strategies. When a twin couples a mechanistic model with learned components and is fitted to observational records, its answer mixes information from the data with assumptions from the mechanism. For sustained strategies that begin at a clinical landmark, we split the counterfactual mean into a part the data support, where patients follow the strategy with probability bounded away from zero, and a part they do not. On the supported part, a cross-fitted sequential doubly robust estimator that uses the twin as its outcome model is consistent whenever the decision model is, so twins that fit equally well but disagree as plug-in estimators agree after correction. On the unsupported part, we characterize the values compatible with the data and with a residual of size at most epsilon around the mechanism. This set is nested in epsilon, reduces to a point when the mechanism is exact and identifiable, and recovers assumption-free bounds as epsilon grows, so the penalty weight of a physics-informed loss becomes a sensitivity parameter. Knowledge constraints on how covariates enter the mechanism shrink the set, and when support grows over time, back-testing gives a data-based lower bound on epsilon. In simulations based on an energy-balance model of weight change under glucagon-like peptide-1 receptor agonist therapy, the correction reduced the spread of estimates across four twins from 0.66 to 0.03 percentage points of weight loss with conservative coverage; on unsupported patients, the exact mechanism was rejected in every replicate, the bounds covered the truth at every residual size not rejected, and a mechanism-free extrapolation was biased by 0.5 points.
ER -