Time-Aware Synthetic Control

Saeyoung Rho, Cyrus Illick, Samhitha Narasipura, Alberto Abadie, Daniel Hsu, Vishal Misra

Keywords: Synthetic Control, Causal Inference, Time Series Panel Data

TL;DR: We suggest using state-space model for synthetic-control-type of panel data to learn temporal evolution of latent factors.

Abstract: The synthetic control (SC) framework is a key tool for observational causal inference in time-series panel data, common in healthcare and clinical research. Although the data analyzed by SC are inherently time-seires, most SC approaches are invariant to permutations of the time indices in the pre-intervention data. In this work, we suggest Time-Aware Synthetic Control (TASC), which models the observations using a linear state space model and performs counterfactual inference using the Kalman filter and RTS smoothing. TASC ensures that the data maintains a low-rank signal with latent factors evolving gradually over time. As an initial demonstration, we apply the TASC approach to a case study on California’s healthcare policy (Proposition 99). Our method showed promising results in placebo tests, indicating its potential applicability in a broader range of healthcare contexts.

Submission Number: 27