Dynamic Policy Evaluation and Learning with Spatio-temporal Data
Abstract
Although sequential decision-making is ubiquitous across domains, policy evaluation and learning with spatio-temporal data remain challenging due to spatial spillover and temporal carryover effects. We develop methods for evaluating and learning individualized dynamic policies under spatio-temporal interference. Under a semiparametric additive outcome model that allows for complex spillover and carryover effects, we consider a family of stabilized estimators for evaluating the performance of a given individualized policy. From this family, we select a data-adaptive optimal estimator that minimizes the asymptotic variance. We then derive the asymptotic distribution of the proposed policy evaluation estimator, and establish the finite-sample regret bounds of our policy learning estimator. We further propose a statistical test to select the complexity of the semiparametric additive model by determining the appropriate order of interactions. Through simulations, we assess the finite-sample performance of our estimators and the validity of the proposed test. Our motivating application examines the optimal allocation of economic aid in Iraq from February 2007 to July 2008. Drawing on declassified conflict data, we study the weekly assignment of aid projects across districts and find that reallocating aid away from regions with a persistently high level of violence can substantially reduce insurgent attacks.
Related
- A Non-Stationary Spatio-Temporal Covariance Model with Dynamic Advection Effects for Rainfall Data
- STCFormer: Adaptive Spatio-Temporal Modeling with Dynamic Cluster Transformer for Station-based Weather Forecasting
- Resolving the Mobility-Equity-Emissions Trilemma: A Causal Machine Learning Framework for Dynamic Tolling Policy
- Reinforcement Learning Under Runtime Confounding: Policy Learning and Off-Policy Evaluation
- Scalable Dynamic Pricing of Substitutable Products through Structure-Guided Policy Learning
BibTeX
@misc{zhou2026dynamic,
title = {{Dynamic Policy Evaluation and Learning with Spatio-temporal Data}},
author = {Lingxiao Zhou and Kosuke Imai and Jason Lyall and Georgia Papadogeorgou},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.15718},
eprint = {2609.15718},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.15718},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - UNPB
TI - Dynamic Policy Evaluation and Learning with Spatio-temporal Data
AU - Lingxiao Zhou
AU - Kosuke Imai
AU - Jason Lyall
AU - Georgia Papadogeorgou
PY - 2026
DA - 2026/09/14
PB - arXiv
DO - 10.48550/arxiv.2609.15718
UR - https://arxiv.org/abs/2609.15718
AB - Although sequential decision-making is ubiquitous across domains, policy evaluation and learning with spatio-temporal data remain challenging due to spatial spillover and temporal carryover effects. We develop methods for evaluating and learning individualized dynamic policies under spatio-temporal interference. Under a semiparametric additive outcome model that allows for complex spillover and carryover effects, we consider a family of stabilized estimators for evaluating the performance of a given individualized policy. From this family, we select a data-adaptive optimal estimator that minimizes the asymptotic variance. We then derive the asymptotic distribution of the proposed policy evaluation estimator, and establish the finite-sample regret bounds of our policy learning estimator. We further propose a statistical test to select the complexity of the semiparametric additive model by determining the appropriate order of interactions. Through simulations, we assess the finite-sample performance of our estimators and the validity of the proposed test. Our motivating application examines the optimal allocation of economic aid in Iraq from February 2007 to July 2008. Drawing on declassified conflict data, we study the weekly assignment of aid projects across districts and find that reallocating aid away from regions with a persistently high level of violence can substantially reduce insurgent attacks.
ER -