Spectral Truncation in Synthetic Control

Technical Summary: Spectral Truncation in Synthetic Control

Problem Statement
Synthetic Control (SC) methods estimate a treated unit's counterfactual trajectory by constructing a weighted average of untreated donor units such that the weighted donor pre-treatment path matches the treated unit's raw pre-treatment path. This paper investigates an alternative approach: Spectral Synthetic Control, which performs matching not in the raw time coordinates, but in the coordinates of the donor panel's leading temporal singular vectors. The motivating hypothesis is that if the outcome process follows a low-rank interactive fixed-effects model, matching in the directions carrying systematic variation might more directly target underlying factor loadings. However, the paper posits that this approach introduces specific risks: weight underdetermination (where the balancing problem has infinitely many solutions) and basis-estimation noise (where unremoved fixed effects contaminate the spectral basis).

Methodology
The authors propose and analyze a family of estimators defined by a spectral truncation parameter $K$ and a hybrid mixing weight $\eta$:

1. Spectral SC: The donor pre-treatment matrix is decomposed via Singular Value Decomposition (SVD). The treated unit and donors are projected onto the top $K$ singular vectors ($V_K$). Weights are chosen to minimize the squared distance between the treated unit's projected path and the weighted donor projected paths, subject to simplex constraints and a ridge penalty.
2. Hybrid Estimator: To avoid the abruptness of truncation, a hybrid estimator assigns a strictly positive but shrunk weight $\eta \in [0, 1]$ to directions outside the retained $K$-dimensional subspace.
   • $\eta = 0$: Recovers truncated Spectral SC.
   • $\eta = 1$: Recovers raw-path SC exactly (equivalent to matching in the full $T_0$-dimensional space).
   • $\eta \in (0, 1)$: Interpolates between the two, constraining all directions but with unequal weighting.
3. Hyperparameter Selection: The ridge penalty ($\lambda_\omega$) and the mixing weight ($\eta$) are tuned via donor-only leave-one-out placebo validation. The rank $K$ is fixed by design for each experimental regime and is not tuned.

Key Theoretical Contributions
The paper establishes three formal results governing the behavior of these estimators:

• Full-Rank Equivalence: At full rank ($K = T_0$), the spectral objective is mathematically identical to the raw-path SC objective. The projection preserves the Euclidean balancing discrepancy exactly.
• Weight Underdetermination: The paper proves that exact spectral balance (without ridge penalty) is an underdetermined system whenever the number of donors $N_0$ exceeds $K + 1$. The solution set forms an affine subspace of dimension $N_0 - K - 1$. Consequently, the specific weights selected by the estimator depend heavily on the regularization term rather than the data, as the data alone cannot pin down a unique solution in the directions orthogonal to the retained subspace.
• Bias Decomposition and Bound: Using a Best Linear Predictor (BLP) construction, the authors derive a finite-sample bound relating the treatment-effect bias to the spectral balancing quality. The bias is shown to depend on:
  1. The achieved spectral residual (controlled by the estimator).
  2. The operator norm of the BLP coefficient matrix ($\|A_K\|_{op}$).
  3. The BLP residual ($r_i$), which captures basis-estimation noise (errors in estimating the true factor subspace) and omitted factors. Crucially, if fixed effects are not removed prior to SVD, the leading singular vectors may be dominated by cross-unit level differences rather than the factor structure, inflating the BLP residual and rendering the spectral bound uninformative.

Empirical Results
The authors evaluate the estimators across eleven data-generating regimes specifically designed to favor spectral truncation (e.g., low-rank structures, sparse donors, weak factors), using 400 replications per regime.

• Raw-Input Performance: When applied to raw pre-treatment paths (without removing fixed effects), Truncated Spectral SC consistently underperforms tuned raw-path SC. The Root Mean Squared Error (RMSE) of Spectral SC is significantly higher (4 to 11 standard errors) across all regimes.
• Hybrid Estimator Behavior: The hybrid estimator, free to choose any $\eta$, overwhelmingly selects $\eta = 1$ (raw-path matching) in the majority of replications across all regimes. It is statistically indistinguishable from tuned raw-path SC in most cases.
• Preprocessing Sensitivity (Robustness Check): A critical finding emerges when the data is preprocessed. When unit and time fixed effects are removed (two-way demeaning) before the spectral decomposition:
  • The performance gap between Spectral SC and raw-path SC nearly vanishes.
  • The placebo-tuning procedure reverses its preference, selecting full truncation ($\eta = 0$) in a majority of replications.
  • This confirms the theoretical bound: removing fixed effects reduces the BLP residual, allowing the spectral basis to better approximate the true factor subspace.

Significance and Claims
The paper does not claim that spectral methods should replace raw-path SC. Instead, it frames the work as a diagnostic result identifying the specific mechanisms that cause spectral truncation to fail under standard conditions:

1. Underdetermination: Low-dimensional balancing leaves too much freedom for regularization to dictate the weights.
2. Basis Contamination: In raw data, unremoved fixed effects distort the estimated spectral basis, inflating the error term in the bias bound.

The authors conclude that the negative results observed in the main simulation study are specific to raw-path input. The paper demonstrates that when the condition of the theoretical bound is met (by removing fixed effects), spectral matching can become viable. The work serves to clarify when and why spectral truncation helps or hurts, rather than proposing a new universal estimator.

Open Questions
The paper explicitly notes several limitations and open areas for future research:

• The robustness of the preprocessing reversal across all eleven regimes (only the baseline was checked).
• Principled selection of the rank $K$ (currently fixed by design).
• Inference procedures that account for the uncertainty in the estimated basis and hyperparameters.
• Extensions to staggered adoption and multiple treated units.