A Relaxation Approach to Synthetic Control

arXiv:2508.01793 · econ.EM · Submitted 2025-08-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "A Relaxation Approach to Synthetic Control".

Warren: Here's the gist of the paper. The synthetic control method (SCM) relaxation approach proposes a novel machine learning algorithm to estimate weights for counterfactual prediction when a donor pool contains more control units than time periods.

Ruth: This method minimizes an information-theoretic measure of the weights subject to relaxed linear inequality constraints in addition to the simplex constraint, allowing it to approximate equal weights within each group when a group structure is present.

Warren: Pruning redundant reasoning tokens The SCM-relaxation method is designed for counterfactual prediction by minimizing an information-theoretic measure of the weights subject to a set of relaxed linear inequality constraints in addition to the simplex constraint.

Ruth: When the donor pool exhibits a group structure, SCM-relaxation approximates the equal weights within each group to diversify the prediction risk.

Paper summary: Warren: So to wrap up on this paper called "A Relaxation Approach to Synthetic Control," they propose SCM-relaxation as a way to estimate weights for counterfactual prediction when you have many control units but few time periods.

Ruth: And the core thesis is that by minimizing an information-theoretic measure of the weights under relaxed constraints, they can approximate equal weights within each group when there’s a group structure present.

Iris: The authors establish three original theoretical contributions here: first, setting up a general framework that accommodates the information-theoretic approach in formulating SCM by maintaining the simplex constraint.

Warren: Second, they overcome a key technical challenge in handling the many non-negativity constraints that arise from the simplex constraint.

Ruth: And third, they allow both the number of groups and the number of factors to diverge to infinity regardless of their relative order. Theorem one shows that their estimator converges to its target under both L1 and L2 norms with desirable rates of convergence <ref:2508.01793#pg1>.

Iris: The paper also suggests that because of this, the L2 SCM-relaxation enjoys desirable properties under fewer conditions and is thus preferred over other methods when you need a reliable result.

Warren: They’ve tested this on Brexit data, showing the estimated counterfactual without Brexit would be substantially higher than realized values.

Ruth: The implication is that for complex policy evaluation where you have lots of control units, this relaxation approach gives you a tool with good asymptotic performance to get a reliable counterfactual.

Conclusion: Ruth: So we're wrapping up on "A Relaxation Approach to Synthetic Control." Basically, they’ve cooked up this new machine learning way to figure out those weights when you have way more control units than time periods.

Warren: Right, so it’s about making the math a bit easier to handle when the donor pool is huge compared to the time periods.

Iris: The authors are building on some existing work, specifically that L2 relaxation method from Shi et al., and they’re using it to solve a constrained optimization problem that includes these relaxed inequality constraints.

Ruth: They’re focusing on how this method handles those group structures—when you have defined groups within your donor pool—by trying to approximate equal weights inside each group.

Warren: And the big takeaway is that asymptotically, this estimator actually hits oracle performance in terms of prediction accuracy when you look at it in the long run.

Iris: They’ve laid out three main theoretical contributions here: first, setting up a general framework that keeps the information-theoretic approach while keeping the simplex constraint; second, dealing with those many non-negativity constraints from the simplex constraint; and third, showing that both the number of groups and the number of factors can go to infinity without them messing up.

Ruth: So for a listener who only cares about what this means, it’s about building a more robust tool for making counterfactual predictions in situations where you have a lot of potential control units.

Warren: It's not just a fancy algorithm; they showed in their example with Brexit that the estimated counterfactual without Brexit was substantially higher than what actually happened.

Iris: The method is built on L2 relaxation, which they prefer because it seems to handle these conditions better than other approaches under fewer assumptions.

Ruth: It really suggests that for complex policy evaluations, you can get a more reliable picture of a counterfactual even when the data setup is messy.

Warren: And once we understand how these weights are calculated—how they balance across those groups—we can start thinking about how to actually use this for things like estimating average treatment effects.

The Chinese University of Hong Kong

econ.EM

Submitted: 2025-08-03

Updated: 2026-11-04

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 88/100

The gist: The synthetic control method (SCM) relaxation approach proposes a novel machine learning algorithm to estimate weights for counterfactual prediction when a donor pool contains more control units than…

Key concepts

Synthetic Control Method (SCM)
SCM is a technique used for counterfactual prediction, creating a synthetic control unit by weighting existing control units to mimic the treated unit's pre-treatment behavior. The goal is to estimate what would have happened to the treated unit in the absence of treatment.
SCM-relaxation
This novel machine learning algorithm relaxes standard SCM constraints by minimizing an information-theoretic measure of weights. When group structures exist, it specifically approximates equal weights within each group to better diversify prediction risk and improve estimation accuracy.
Information-theoretic measure
This is a mathematical objective function used in the optimization problem. It quantifies how good or bad a set of estimated weights is at representing the true underlying structure of the data, guiding the algorithm toward optimal weights.
Oracle performance
Oracle performance means that in large samples, this estimator achieves the best possible predictive accuracy achievable. It indicates that as more data is available, the method's predictions become as accurate as theoretically possible.

Terminology

Summary

The synthetic control method (SCM) relaxation approach proposes a novel machine learning algorithm to estimate weights for counterfactual prediction when a donor pool contains more control units than time periods. p. 0 This method minimizes an information-theoretic measure of the weights subject to relaxed linear inequality constraints in addition to the simplex constraint, allowing it to approximate equal weights within each group when a group structure is present. p. 0

Pruning redundant reasoning tokens

The SCM-relaxation method is designed for counterfactual prediction by minimizing an information-theoretic measure of the weights subject to a set of relaxed linear inequality constraints in addition to the simplex constraint. When the donor pool exhibits a group structure, SCM-relaxation approximates the equal weights within each group to diversify the prediction risk. p. 0 Asymptotically, the proposed estimator achieves oracle performance in terms of out-of-sample prediction accuracy. p. 0 The method is built on the L2-relaxation method introduced by Shi et al. p. 1 (2025) and formulates the weight estimation as a constrained optimization problem where the objective function is an information-theoretic divergence measure of the weight vector. p. 1

How the training signal is built

The model sets up data with J + 1 cross-sectional units and T0 + T1 time periods, where unit j = 0 is the sole individual that received a treatment during t ∈ T1. The standard SCM estimator minimizes the L2-distance between the treated unit and synthetic control in the pre-treatment periods, i. p. 5e., wbSC:= arg min w∈∆J∥y0 − Yw∥22. p. 5 To address challenges arising from many potential control units, SCM-relaxation introduces a relaxed form of the first-order condition (FOC) derived from the original SCM formulation. p. 2 This relaxation is defined by solving a convex optimization problem min w∈∆J,γ∈R∥w∥22 s. p. 6t.∥Σbw − Υb + γ1J∥∞ ≤ η, where η = ηT0,J is a user-specified tuning parameter. p. 6

Where the gains come from

The paper establishes three original theoretical contributions regarding the asymptotics of SCM-relaxation estimators. Firstly, it sets up a general framework that accommodates the information-theoretic approach in formulating SCM by maintaining the simplex constraint. p. 2 Secondly, it overcomes a key technical challenge in handling many non-negativity constraints rising from the simplex constraint. p. 2 Thirdly, it allows both the number of groups (K) and the number of factors (r) to diverge to infinity regardless of their relative order. p. 3 Theorem 1 shows that the estimator wb converges to w∗ under both L1- and L2-norm with desirable rates of convergence. p. 2

Theoretical Foundations and Performance

The asymptotic analysis relies on several key assumptions, including Assumption 1 (Factors and loadings) which bounds the spectrum of the core covariance matrix Σbco based on whether K ≤ r or K > r. Lemma S.1 shows that when K ≤ r, the eigenvalues of the “core” covariance matrix Σbco are bounded away from zero so that Σbco is non-degenerate. p. 9 If K > r, then the Lagrange multiplier µ1 associated with the balancing constraint satisfies∥µ1∥2 = Op(r/J). p. 6 Theorem 3 guarantees that wb(g) behaves as well as the oracle weight w∗(g) in large sample, as in the following corollary. p. 2

Empirical Validation and Application

Monte Carlo simulations demonstrate the finite sample performance of SCM-relaxation in comparison with other off-the-shelf machine learning methods. In an empirical application assessing the economic impact of Brexit on the United Kingdom’s real GDP, the estimated weights of the control units from our method are more interpretable than those from SCM. p. 2 The paper shows that all methods except fsPDA outperform the canonical SCM for all combinations of K and r. p. 2 The L2-SCM-relaxation method is consistently the best performer and is recommended in practice. p. 0 The estimated counterfactual without Brexit would be substantially higher than the realized values. p. 2

Feature Engineering and Standardization

The theoretical results can be extended to allow for heterogeneous scales by standardizing the data, where yjt = θj (λ′jft + ujt). The standardized SCM-relaxation estimates the weights by solving min w∈∆J,γ∈R Xj∈[J] g(wj) s. p. 1t. Σbsw − Υb s + γ1J∞ ≤ η, where the solution is denoted as (wb s(g), γb s(g)). p. 7 This standardized version achieves near-oracle prediction risk under the same conditions imposed in Theorem 3. p. 0 The Monte Carlo simulation evidence shows that the relaxation methods are robust against mild deviation from group structure, manifesting their practical usefulness. p. 4 The L2-SCM-relaxation CIs are shorter in finite sample, especially when T0 is small. p. 2 The EL is worse than L2 and entropy (and even worse than SCM in some cases). p. 3 The performance of the estimated weights is displayed in Table 2. p. 0 In Figure 4, L2-SCM-relaxation nearly perfectly traces the group pattern with the smallest estimation errors. p. 8 The final conclusion is that L2-SCM-relaxation stands out as the best performer and is recommended in practice.

Inference on ATT

For inference on the average treatment effect on the treated (ATT), one can construct the post-treatment counterfactual of the treated as ybN 0t = Xj∈[J]wbjyj. The ATT estimator is thus τb = 1/T1 Xt∈T1(yI 0t − ybN 0t). p. 7 A debiased treatment effect (bc) estimator is constructed by partitioning the pretreatment periods into B consecutive disjoint blocks H1 ∪ H2 ∪ · · · ∪ HB ⊆ T0. p. 5 This construction is able to induce approximate independence between wb(k) and the data in Hk ∪ T1. p. 3 The resulting t-statistic TB → d tB−1, where tB−1 is a random variable following the t-distribution with degrees of freedom B − 1.

Conclusion

We propose a machine learning algorithm, termed SCM-relaxation, to estimate the weights for counterfactual prediction of a treated unit at the presence of many potential control units. Given a factor model with the latent group structures, L2-SCM-relaxation produces weights that converge asymptotically to the oracle weights. p. 0 If one chooses the EL or the entropy objective function to conduct the relaxation, under extra regularization conditions they possess similar properties as the L2 counterpart if the true weight is non-zero. p. 1 Our theoretical study suggests that L2-SCM-relaxation enjoys desirable properties under fewer conditions and is thus preferred. p. 6 We apply them to evaluate the impact of Brexit on the UK’s real GDP, and find that the estimated counterfactual without Brexit would be substantially higher than the realized values. p. 2 The first sentence is exactly what it says about its subject and finding. p. 4

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop an SCM-relaxation algorithm for counterfactual prediction that minimizes an information-theoretic measure of weights subject to relaxed linear inequality constraints in addition to the simplex constraint, allowing for dense weighting schemes when the donor pool exhibits a group structure. This system can generate more interpretable weights than standard Synthetic Control Method (SCM) by approximating equal weights within each latent group.

  2. Implement an asymptotic theory for SCM-relaxation estimators under latent group structures, showing that the estimated weight vector converges sufficiently fast to a desirable target, even when the number of control units is much larger than the number of pre-treatment time periods. This enables robust counterfactual predictions in high-dimensional settings where J is much larger than T0.

  3. Create a unified asymptotic framework for any convex divergence function with non-negative support and restricted strong convexity, allowing the method to encompass popular estimators like empirical likelihood and entropy balancing, while ensuring convergence rates are maintained.

  4. Design a feature engineering pipeline that standardizes input data by calculating sample variances for each control unit prior to feeding it into the SCM-relaxation algorithm. This extension allows theoretical results to be applied effectively when outcomes have heterogeneous scales across different individuals.

  5. Build an inferential procedure for the average treatment effect on the treated (ATT) using cross-fitting and debiasing techniques, such as constructing a bias-corrected estimator (bc) for ATT by averaging estimates from disjoint pre-treatment blocks. This allows for valid t-tests even when regularization constraints are imposed.

  6. Integrate Monte Carlo simulation capabilities to rigorously test the finite sample performance of SCM and its relaxation methods against standard machine learning techniques like Lasso, Ridge, and Group Lasso across various group structures (K r). This provides a quantitative basis for selecting the optimal relaxation method based on desired prediction error ratios.

Sources

Related papers