[Paper Review] Policy Analysis using Synthetic Controls in Continuous-Time

Alexis Bellot, Mihaela van der Schaar|arXiv (Cornell University)|Feb 2, 2021

Advanced Causal Inference Techniques39 references4 citations

TL;DR

This paper introduces Neural Continuous Synthetic Controls, a continuous-time framework for counterfactual estimation using controlled differential equations to model latent counterfactual paths. By treating control unit trajectories as continuous processes and learning a dynamic vector field to combine them, the method enables accurate, flexible counterfactual prediction on irregularly sampled, multivariate time series—overcoming key limitations of discrete-time synthetic controls in dynamic and misaligned data settings.

ABSTRACT

Counterfactual estimation using synthetic controls is one of the most successful recent methodological developments in causal inference. Despite its popularity, the current description only considers time series aligned across units and synthetic controls expressed as linear combinations of observed control units. We propose a continuous-time alternative that models the latent counterfactual path explicitly using the formalism of controlled differential equations. This model is directly applicable to the general setting of irregularly-aligned multivariate time series and may be optimized in rich function spaces -- thereby improving on some limitations of existing approaches.

Motivation & Objective

• To address the limitations of discrete-time synthetic controls in handling irregularly aligned and multivariate time series.
• To model the latent counterfactual trajectory of a treated unit as a continuous dynamical system rather than a discrete linear combination of observed data points.
• To improve counterfactual estimation by capturing time-varying, nonlinear dependencies between control and treated units through a learnable vector field.
• To enable robust inference in settings where traditional synthetic controls fail due to temporal misalignment or nonstationary dynamics.
• To provide a flexible, function-space-optimized framework that generalizes beyond static weights and fixed grids.

Proposed method

• The method models the counterfactual path of the treated unit as the solution to a controlled differential equation driven by the latent paths of control units.
• It learns a latent vector field f that combines infinitesimal changes in control trajectories to reconstruct the counterfactual evolution of the treated unit.
• The vector field f is parameterized using a neural network, enabling nonlinearity and adaptivity in the combination of control paths.
• The model is trained end-to-end using observed pre-treatment data, minimizing the error between the predicted and actual outcomes before treatment.
• It supports irregular sampling by modeling continuous paths rather than discrete observations, allowing interpolation-free processing of misaligned time series.
• The framework can incorporate time-varying covariates via a data-dependent lasso regularization scheme on the weight matrix W, improving matching fidelity.

Experimental results

Research questions

• RQ1 Can synthetic control estimation be generalized to continuous-time processes to better model real-world, irregularly sampled time series data?
• RQ2 How can time-varying, nonlinear dependencies between control and treated units be captured in counterfactual modeling beyond static linear weights?
• RQ3 Can a continuous-time formulation improve counterfactual prediction accuracy in settings with temporal misalignment or sparse observations?
• RQ4 To what extent can the model maintain validity when control units are subject to spillover effects or external shocks post-intervention?
• RQ5 How can uncertainty in counterfactual estimates be quantified within a continuous-time synthetic control framework?

Key findings

• The proposed Neural Continuous Synthetic Controls model achieves improved counterfactual estimation on irregularly aligned multivariate time series compared to discrete-time synthetic controls.
• The method successfully captures dynamic, time-varying relationships between control and treated units, which static linear combinations fail to represent.
• In the California cigarette tax experiment, the model produced a counterfactual trajectory that closely matched the observed pre-treatment trend and predicted a plausible post-treatment decline.
• The framework supports flexible integration of time-variant and time-invariant covariates through a relevance-weighted regularization scheme on the weight matrix.
• The model generalizes beyond fixed grids and avoids interpolation artifacts, enabling robust inference on real-world data with irregular sampling.
• While uncertainty quantification is not explicitly modeled in the current version, the framework is compatible with stochastic differential equation extensions for future uncertainty estimation.