TL;DR
This paper reveals a crucial theoretical link between Synthetic Controls (SC) and balancing scores, showing that SC weights can act as balancing scores under specific conditions. This insight bridges two major causal inference frameworks, suggesting that SC methods can effectively emulate randomized controlled trials, leading to more robust causal effect estimation.
What Problem Does This Paper Address?
In causal inference, a primary goal is to estimate the effect of a treatment or intervention. Randomized Controlled Trials (RCTs) are the gold standard because they ensure treatment assignment is independent of potential outcomes, allowing for unbiased estimation. However, RCTs are often infeasible or unethical in real-world scenarios.
Two popular alternative approaches are Synthetic Controls (SC) and Balancing Score Methods (like propensity scores). SC is widely used for time-series data, especially when only a few units are treated (e.g., estimating the impact of a policy on a single country). It constructs a "synthetic" control unit by taking a weighted average of untreated units to mimic the treated unit's pre-treatment behavior. Balancing score methods, on the other hand, are more common in cross-sectional data with many units and aim to create comparable treatment and control groups by conditioning on a score that summarizes confounding variables.
While both methods use pre-treatment information to adjust for confounding, their theoretical connections have been unclear. This paper aims to bridge this gap by demonstrating when SC weights can achieve the same desirable property as balancing scores: making potential outcomes independent of treatment assignment.
Methods
The authors establish their main result by leveraging a factor model for time-series data and introducing specific assumptions. The core of their argument relies on showing that under these assumptions, the weights derived from the Synthetic Control method (often referred to as "oracle weights" because they are assumed to be known or perfectly estimated) satisfy the definition of a balancing score.
Specifically, they define a factor model where outcomes are driven by common time trends, unit-specific factors, and treatment effects. They then introduce three key assumptions:
- Feasibility: There exist weights such that the pre-treatment outcomes of the treated unit can be perfectly predicted by a weighted average of control units' pre-treatment outcomes.
- Uniformly Bounded: The SC weights are between 0 and 1.
- Cumulatively Bounded: The SC weights sum to 1.
Under these assumptions, they prove that the SC weights effectively make the potential outcomes independent of the treatment assignment, thus acting as balancing scores.
Datasets
Not discussed in the paper. The paper is theoretical and focuses on establishing a general result rather than empirical evaluation on specific datasets.
Key Insights and Ideas
- SC weights are balancing scores: The most significant insight is the formal demonstration that Synthetic Control weights, under certain conditions, fulfill the role of balancing scores. This is a crucial theoretical link between two distinct causal inference frameworks.
- Emulating RCTs: By acting as balancing scores, SC weights allow the treated unit to be compared to a "synthetic" control that is effectively randomized with respect to potential outcomes. This means SC methods can be viewed as a way to emulate an RCT, even with observational data.
- Unifying Causal Inference Tools: This finding helps to unify the understanding of different causal inference techniques, suggesting that advancements in one area (e.g., doubly robust methods in balancing scores) might be transferable to the other (Synthetic Controls).
Key Results
The paper's main result is Theorem 1: Given the causal dependencies and assumptions A.1-A.3 (feasibility, uniform boundedness, and cumulative boundedness of SC weights), the potential outcomes are independent of the treatment assignment conditional on the SC weights. In simpler terms, {Y₁(z)}t | Z | β, where β represents the SC weights. This implies that SC weights are balancing scores.
Limitations and Future Work
- Assumptions: The result relies on specific assumptions, particularly the existence of "oracle" SC weights that perfectly predict pre-treatment outcomes. While the paper notes that these weights can be estimated, the exact conditions under which these estimates are sufficiently accurate for the balancing score property to hold in practice are not fully explored.
- Heteroskedasticity: The paper mentions that exactly identifying oracle weights might not be feasible, especially with heteroskedastic noise, and that regularization is typically used to estimate weights. The implications of using estimated weights versus oracle weights on the balancing score property are a point for further investigation.
- Future Work: The authors suggest exploring doubly robust SC methods that combine propensity and prognostic scores, similar to existing cross-sectional methods. They also aim to investigate if their result holds under weaker assumptions and in scenarios where SC has been applied for estimation.
Summary updated