Synthetic Blips: Generalizing Synthetic Controls for Dynamic Treatment Effects

Anish Agarwal, Sukjin Han, Dwaipayan Saha, Vasilis Syrgkanis, Haeyeon YoonarXiv 2022 · arXiv

TL;DR

This paper introduces Synthetic Blips, a novel generalization of the synthetic control framework designed for sequential, adaptive, and multi-treatment settings. By breaking down complex treatment histories into recursive "blip effects," the method bypasses the exponential donor sample complexity that plagues naive extensions of synthetic controls, allowing for valid causal inference even in the presence of unobserved time-varying confounders and feedback loops.

What Problem Does This Paper Address?

Synthetic control (SC) methods are a staple in econometrics for estimating counterfactuals when dealing with unobserved confounding in panel data. However, traditional SC assumes a simple "one-shot" treatment or a staggered adoption design where treatment is permanent and irreversible.

In the real world, interventions are often dynamic and adaptive. For example, a patient might receive multiple different therapies sequentially, or a firm might receive financial loans and insurance at various times based on its fluctuating performance. Naively extending SC to these settings fails for two reasons:

  1. Donor Sample Complexity: To find a match for a specific sequence of treatments over time, you would need donor units that received that exact same sequence. As the number of time steps ($T$) and treatment options ($A$) grow, the required number of donors explodes exponentially ($A^T$).
  2. Strict Exogeneity Requirements: Existing panel data methods often assume that treatment assignments don't depend on past outcomes (ruling out feedback loops), which is unrealistic for data-adaptive policies like recommendation algorithms or targeted government aid.

Methods

The authors propose Synthetic Blips, which uses a dynamic programming-like backward induction process to recursively estimate counterfactuals.

The core idea is to decompose the potential outcome under any arbitrary sequence of interventions into a baseline outcome (under control) plus a sum of blip effects. A blip effect is the incremental causal impact of applying a treatment at a specific time step $t$ while holding all other periods fixed.

Under an additive low-rank latent factor model assumption—where latent factors capture unobserved time-varying confounding signals—the authors show that the blip effect for a target unit can be expressed as a linear combination of the blip effects of a carefully chosen subgroup of donor units. Instead of matching the entire history, donors only need to match the control history up to time $t-1$ and receive the specific treatment at time $t$. The weights for these synthetic controls and blips are estimated using Principal Component Regression (PCR), which naturally de-noises the observed data through low-rank approximation.

The paper explores two main variants of this model:

  • Linear Time-Varying (LTV): Where treatment effects can vary flexibly across calendar time.
  • Linear Time-Invariant (LTI): Where treatment effects depend only on the elapsed time (lag) since the treatment was administered.

Datasets

The authors evaluate their framework using a unique, newly assembled Korean firm-level panel dataset (2006–2015). This dataset links three distinct sources:

  • The Survey of Business Activities (SBA) from Statistics Korea (detailed firm characteristics).
  • Export insurance data from the Korea Trade Insurance Corporation (K-SURE).
  • Export loan data from the Export-Import Bank of Korea (EXIM).

The final balanced estimation sample consists of 661 firms (162 supported, 499 never-supported) observed over a 10-year period, focusing on how sequential financial support impacts a firm's export value.

Key Insights and Ideas

  • The Power of Blip Decomposition: Treating sequential treatments as a series of additive "blips" reduces the donor sample complexity dramatically. Instead of scaling exponentially, the sample complexity scales linearly with time or treatments.
  • Embracing Feedback Loops: Unlike traditional difference-in-differences or standard synthetic controls, Synthetic Blips explicitly accommodates adaptive policies where past shocks feed back into future treatment decisions.
Comparison of Donor Sample Complexity and Adaptivity Across Models
  • Retrospective vs. Prospective Policy Learning: The framework distinguishes between optimizing schedules for firms already in the dataset (retrospective) and training interpretable decision trees to guide future allocations for new firms (prospective).

Key Results

The empirical application to Korean exporting firms yielded several striking insights:

  • Loans vs. Insurance: Sustained export loans generated a massive positive cumulative effect, increasing export values by an average of 120.4 million USD per firm by directly relaxing working-capital constraints. Conversely, sustained public insurance showed a cumulative negative effect (-36.3 million USD), likely due to moral hazard or selection bias (public insurance covering riskier profiles that the private market rejects).
  • Timing Matters: For loans, concentrated schedules (front-loading or back-loading) performed significantly better than spreading them out evenly, as consecutive support sustains working capital across production cycles.
  • Massive Scope for Policy Optimization: Retrospective policy learning revealed that an optimal cost-constrained allocation rule could increase cumulative export outcomes by 114% while actually using fewer total resources (335 supports vs. the observed 346).
  • The Value of Coordination: If K-SURE and EXIM jointly optimized their decisions rather than acting independently, average potential outcomes would increase by 39% at a lower total cost.

Limitations and Future Work

  • Additivity Assumption: The model relies heavily on the assumption that treatment effects are additive across time steps. While the authors discuss a "hybrid" model that allows for interactions within a short time window, fully non-linear or highly interactive dynamic treatment effects remain a challenge.
  • Fixed Lag Dependence: To prevent estimation errors from compounding geometrically during backward induction as $T \to \infty$, the authors must assume a fixed lag dependence (i.e., treatments only affect outcomes for a finite number of future periods $q$).
  • Winner's Curse: The reported gains in retrospective policy learning are subject to optimization bias over noisy estimates, meaning real-world gains might be slightly more modest than the in-sample estimates suggest.

Summary updated

Abstract

We propose a generalization of the synthetic control methods to the setting with dynamic treatment effects, in which each unit receives multiple treatments sequentially, according to an adaptive policy that depends on a latent, endogenously time-varying confounding state. Under a low-rank latent factor model assumption, which admits linear time-varying and time-invariant dynamic triangular systems as special cases, we develop an identification strategy for any unit-specific mean outcome under any sequence of interventions. Our method, which we term synthetic blips, is a backward induction process in which the blip effect of a treatment at each period for a target unit is recursively expressed as a linear combination of the blip effects of other units that received the designated treatment, avoiding the combinatorial donor requirements of naive synthetic control extensions. We provide easy-to-implement estimation algorithms that yield consistent estimators. Using unique Korean firm-level panel data, we estimate individualized dynamic treatment effects and optimal allocation rules in the context of financial support for exporting firms.
BibTeX
@misc{agarwal2022synthetic,
  author = {Anish Agarwal and Sukjin Han and Dwaipayan Saha and Vasilis Syrgkanis and Haeyeon Yoon},
  title = {{Synthetic Blips: Generalizing Synthetic Controls for Dynamic Treatment Effects}},
  year = {2022},
  eprint = {2210.11003},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2210.11003},
}
Figure