Xu 2016 - Generalized Synthetic Control Method: Causal Inference with Interactive Fixed Effects Models

Xu (2016) — Generalized Synthetic Control Method

Summary

Yiqing Xu proposes the Generalized Synthetic Control (GSC) method, which unifies difference-in-differences and the synthetic control method under a single interactive fixed effects (IFE) framework. GSC handles multiple treated units and variable treatment periods, produces frequentist uncertainty estimates (standard errors and confidence intervals) via parametric bootstrap, and embeds cross-validation for automatic model selection — addressing the key limitations of both DID (parallel trends) and SC (single treated unit, no inference).

Citation

Xu, Yiqing. 2017. “Generalized Synthetic Control Method: Causal Inference with Interactive Fixed Effects Models.” Political Analysis 25 (1): 57–76. https://doi.org/10.1017/pan.2016.2

Overview

DID requires parallel trends — that treated and control units would have followed the same trajectory in the absence of treatment. Synthetic controls require a single treated unit and provide no formal uncertainty estimates. The GSC method addresses both limitations by modeling unobserved time-varying confounders explicitly via an interactive fixed effects (IFE) model.

The key insight: by estimating the IFE model on the control group only, then projecting treated units’ pre-treatment outcomes onto the estimated factor space, the method imputes what the treated units’ outcomes would have been without treatment. DID is the special case where factor loadings are constant (${\lambda }_i=\lambda$); synthetic control is the special case with one treated unit and no parametric model.

Paper Structure

Section	Title	Key content
1	Introduction	DID parallel trends problem; IFE as solution; GSC as unification
2	Framework	Assumption 1 (IFE functional form); ATT estimand; Assumptions 2–5
2.1	Identification assumptions	Strict exogeneity (Assumption 2); weak serial dependence
3	Estimation Strategy	3-step GSC estimator; Algorithm 1 (cross-validation); Algorithm 2 (bootstrap)
3.1	Model selection	LOO cross-validation for number of factors $r$
3.2	Inference	Parametric bootstrap procedure for $\text{Var}(\widehat{ATT})$
4	Monte Carlo Evidence	Simulation: GSC vs DID, IFE, SC; Table 1 (bias, SD, RMSE, coverage)
5	Empirical Example	Election Day Registration (EDR) and voter turnout, 47 US states 1920–2012
6	Conclusion	Caveats: $T_0<10$ or $N_{co}<40$; diagnostics recommended

Key Contributions

1. Unification of DID and SC: The GSC framework encompasses DID (constant factor loadings) and SC (single treated unit) as special cases.

2. Multiple treated units, variable treatment timing: The IFE model is estimated once on the control group; factor loadings for each treated unit are estimated separately in Step 2. No need to run separate SC optimizations for each treated unit.

3. Frequentist uncertainty estimates: The parametric bootstrap (Algorithm 2) produces standard errors and confidence intervals, filling the gap left by the permutation-only inference of canonical SC.

4. Automatic model selection: Cross-validation (Algorithm 1) selects the number of factors $r$ without researcher discretion, guarding against specification search.

5. Performance vs. alternatives: Monte Carlo shows GSC has less bias than DID/IFE when treatment is heterogeneous and less bias than IFE when treatment effect is heterogeneous across units; more efficient than original SC.

Relationship to Existing Notes

Concept	Vault Note	Relationship
Basic SC estimator	Synthetic Control	GSC generalizes: multiple units, parametric bootstrap
Linear factor model	Synthetic Control Bias Theory	Same model; GSC estimates it directly on controls
Bias bound	Synthetic Control Bias Theory	GSC bias → 0 as $N_{co},T_0\to \infty$ (Remark 2)
DID/parallel trends	Differences-in-Differences	DID is the special case ${\lambda }_i=\lambda$ in Assumption 1
Multiple treated units	Synthetic Control Extensions	GSC is the primary multi-unit extension (vs. penalized SC)
Bayesian IFE/TSCS	Bayesian Difference in Differences	Bayesian analog; Pang (2014) cited as a complement

Caveats (from Conclusion)

1. Small samples: Bias increases when $T_0<10$ or $N_{co}<40$ due to imprecise estimation of factor loadings (“incidental parameters” problem)
2. Common support requirement: If treated and control units do not share common support in factor loadings, GSC may extrapolate — diagnostics essential (plot factor loadings as in Figure 3b)
3. Limitations vs. complex DGPs: Cannot accommodate dynamic treatment–outcome relationships, structural breaks, multiple treatment intensities, or random coefficients for observed covariates

Notes Generated from This Paper

• Xu 2016 - Overview — this note
• Generalized Synthetic Control Method — full formal treatment: IFE model, assumptions, 3-step estimator, cross-validation, bootstrap, empirical application

See Also

• Synthetic Control — the canonical single-unit estimator that GSC generalizes
• Synthetic Control Bias Theory — the linear factor model underpinning GSC
• Synthetic Control Extensions — survey of extensions including GSC, elastic net, matrix completion
• Differences-in-Differences — the special case when parallel trends holds
• Abadie 2021 - Overview — complementary methodological guide on canonical SC feasibility
• The Selection Problem — GSC addresses the selection problem when parallel trends (DiD) or convex hull (SC) conditions fail