Research map / Probabilistic and scientific modeling / Treatment effects

Double Machine Learning for Causal Inference under Shared-State Interference

ICML 2025 · Probabilistic and scientific modeling › Treatment effects · approach: Distributional treatment effects · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Extends double machine learning inference to settings where interference is channeled through a Markov-evolving shared state, proving asymptotic normality and a consistent variance estimator, and instantiating it for average direct effects and switchback GATE estimation.

Problem

Causal inference of treatment effects when units interact through shared states (prices, recommendations, inventory, social signals) that evolve over time, so outcomes are dependent across units and standard iid DML does not apply.

Approach

Models interference as fully mediated by a Markov-evolving shared state H_t, assuming potential outcomes are independent conditional on the shared state, and proves a DML theorem (Neyman-orthogonal estimator, auxiliary-sample nuisance training) with asymptotic normality under geometric ergodicity with detailed balance or m-dependence, plus a Markov-chain-consistent variance estimator (batching or

Key results

• Proves asymptotic normality of the DML estimator under shared-state interference with a consistent variance estimator under either geometric ergodicity with detailed balance or m-dependence.
• Instantiates the theorem for average direct effect (ADE) in observational settings and GATE variance reduction in switchback experiments.
• Simulations show the estimators concentrate around true effects and the variance estimators give approximately nominal confidence interval coverage.

Surprising findings

• Treating the shared state as an iid covariate in standard DML can yield inconsistent treatment effect or variance estimators.

Limitations

• Requires the shared state to be the sole channel of interference and a Markov structure on the shared state.
• Requires an auxiliary sample for nuisance training unless the models are Donsker; sample splitting is not guaranteed valid under dependence.
• Requires known bound m on dependence length for the m-dependent case.

Experimental setup

Metrics: coverage probability of confidence intervals finite-sample concentration of estimators

Baselines: naive iid DML treating shared state as covariate

Builds on: double machine learning exposure mappings AIPW

Closest papers by meaning

• Higher-Order Causal Message Passing for Experimentation with Complex Interference (NeurIPS 2024 · similarity 0.84)
• Active Treatment Effect Estimation via Limited Samples (ICML 2025 · similarity 0.84)
• Off-policy estimation with adaptively collected data: the power of online learning (NeurIPS 2024 · similarity 0.83)
• Causal Inference out of Control: Estimating Performativity without Treatment Randomization (ICML 2024 · similarity 0.83)
• Impossibility of Distribution-Free Predictive Inference for Individual Treatment Effects (NeurIPS 2026 · similarity 0.83)
• Independent-Set Design of Experiments for Estimating Treatment and Spillover Effects under Network Interference (ICLR 2024 · similarity 0.83)