Evaluating Time-Varying Causal Effects in Observational Data Using Synthetic Control Methods with Robust Inference in High-Dimensional Settings

Project Abstract

Evaluating time-varying causal effects in observational data presents a persistent challenge due to confounding, non-stationarity, and high dimensionality of covariates typically encountered in real-world settings. This study develops a unified framework that integrates synthetic control methods with robust statistical inference to quantify how causal effects evolve over time when treatment adoption is staggered or continuous and observational data mechanisms dominate. Our approach extends classical synthetic control by employing time-varying weights, regularized matrix factorization, and high-dimensional balancing techniques to construct credible counterfactual trajectories for treated units. We address non-stationarity through adaptive kernel-based smoothing of outcome paths and by incorporating heterogeneity through unit- and time-specific random effects, enabling accurate estimation of heterogeneous treatment effects across time and units. The core methodological contributions include (i) a robust inference procedure combining placebo tests, conformal prediction intervals, and bootstrap-based uncertainty quantification tailored for high-dimensional synthetic controls, ensuring valid coverage under model misspecification and potential spillovers; (ii) a dynamic weighting scheme that optimally learns from pre-treatment data to minimize pre-treatment imbalances while guarding against overfitting in settings with many covariates; (iii) a time-varying treatment effect estimator that decomposes observed differences into instantaneous and lagged components, capturing persistence and delayed responses commonly observed in policy changes, digital interventions, and environmental regulations; and (iv) a framework for assessing the sensitivity of findings to unobserved confounding through partial identification bounds and scenario analyses. The empirical evaluation comprises simulations designed to mimic complex observational environments with staggered adoption, high-dimensional covariates, and varying degrees of confounding strength, demonstrating superior bias reduction and interval coverage relative to benchmark methods such as conventional synthetic control, difference-in-differences, and mismatched control approaches. We then apply the proposed framework to multiple real-world settings (a) evaluating the time-varying impact of a region-wide price reform on consumer welfare using high-dimensional household and firm-level data; (b) assessing the dynamic effectiveness of an educational policy on student achievement with granular socio-economic covariates; and (c) analyzing the longitudinal health outcomes following a nationwide public health intervention, incorporating spatially correlated unobservables and network spillovers. Across applications, we highlight the ability to recover nuanced temporal patterns, including short-term shocks, medium-term adaptation, and long-run stabilization, while providing reliable uncertainty quantification under challenging data-generating processes. The results illustrate that incorporating time-varying synthetic control with robust inference markedly improves the credibility of causal claims in high-dimensional observational studies, yielding interpretable trajectories of treatment effects that inform policy design, implementation timing, and transferability of findings. The framework offers practitioners a versatile tool to discern when and how interventions produce durable, time-sensitive improvements, guiding evidence-based decision-making in economics, public policy, and the social sciences.

Project Overview

What This Project Is About

---

A plain-language overview of the topic and what the project investigates.

---

---

The Problem It Addresses

---

Many real-world studies compare groups over time using simple before-after comparisons. These can be biased when there are differences between groups or when effects change over time. This project tackles how to measure how outcomes would have changed if a treatment had not occurred, even when we have many potential factors to consider and limited randomization.

---

---

Objectives of the Project

---

1. Clarify what “time-varying causal effects” means in plain terms.
2. Explain synthetic control ideas in a simple way and why robustness matters.
3. Identify data requirements and potential sources.Common pitfalls will be highlighted.
4. Show how high-dimensional data complicates analysis and ways to address it.

---

---

What You Will Do Step by Step

---

1) Learn basic concepts of causal inference and observational data. 2) Find a suitable dataset with a clear intervention and multiple potential predictors. 3) Build a simple synthetic control example to illustrate the idea. 4) Introduce the idea of robustness and how to test results under different assumptions. 5) Analyze time-varying effects with visualization and simple checks. 6) Discuss limitations and practical considerations for real-world data.

---

---

---

---

Expected Outcome

A clear, approachable explanation of how time-varying treatments might influence outcomes, with a simple example, limitations, and guidance for future work. The project aims to give a practical framework that students can apply to real data while understanding the caveats of observational studies.