Synthetic Control Bias Theory

Summary

The formal justification for synthetic controls rests on the linear factor model — a generalization of difference-in-differences that allows unobserved confounders to have time-varying loadings. Under this model, Abadie, Diamond, and Hainmueller (2010) derive a bias bound: the bias of the synthetic control estimator is inversely proportional to the number of pre-treatment periods $T_0$, provided the synthetic control closely tracks the treated unit’s pre-treatment trajectory. Variable selection (via the V matrix and cross-validation) is the mechanism that enforces this fit.

Overview

The basic synthetic control estimator in Synthetic Control is intuitive — find a weighted combination of donor units that matches the treated unit before the intervention, then attribute post-treatment divergence to the treatment. But why is this a valid identification strategy? The answer is the linear factor model.

The key insight: a synthetic control that reproduces the treated unit’s pre-treatment outcomes implicitly matches on the unobserved common factors ${\mu }_j$ — exactly those confounders that would bias a regression estimator. The more pre-treatment periods available, the more constraints the matching imposes, and the better the identification.

The Linear Factor Model

Definition: Linear Factor Model (Abadie et al. 2010, Eq. 10)

The potential outcome without treatment for unit $j$ at time $t$ follows:

$$Y_{jt}^N={\delta }_t+{\mathbf{\theta }}_t{\mathbf{Z}}_j+{\mathbf{\lambda }}_t{\mathbf{\mu }}_j+{\epsilon }_{jt}$$

where:

• ${\delta }_t$ = common time trend (constant factor loading)
• ${\mathbf{Z}}_j$ = observed covariates (time-invariant or unaffected by treatment), with time-varying coefficients ${\mathbf{\theta }}_t$
• ${\mathbf{\mu }}_j$ = vector of unobserved unit-specific factors
• ${\mathbf{\lambda }}_t$ = vector of time-varying factor loadings (common factors)
• ${\epsilon }_{jt}$ = zero-mean idiosyncratic shocks

This model generalizes the standard panel data fixed-effects model. The difference-in-differences / fixed-effects restriction ${\lambda }_t=\lambda$ (constant over time) is obtained by restricting ${\mathbf{\lambda }}_t$ to be time-invariant. The linear factor model allows ${\mathbf{\lambda }}_t$ to change over time, so the “parallel trends” assumption of DiD is a special case.

Connection to DiD

DiD assumes ${\mathbf{\lambda }}_t=\mathbf{\lambda }$ (constant factor loadings), so unobserved confounders affect all units equally across time. The linear factor model relaxes this: each unit $j$ has its own loading ${\mathbf{\mu }}_j$ on common factors ${\mathbf{\lambda }}_t$ that may drift over time. Synthetic control handles this by matching on the trajectory, not just the level.

The Bias Bound

Theorem: Bias Bound for Synthetic Controls (Abadie, Diamond, and Hainmueller 2010)

Under the linear factor model, suppose a synthetic control with weights $W^{\ast }=(w_2^{\ast },\dots ,w_{J+1}^{\ast }{)}^{\prime }$ reproduces the characteristics of the treated unit:

$${\mathbf{X}}_1\approx {\mathbf{X}}_0{\mathbf{W}}^{\ast }$$

where ${\mathbf{X}}_1$ and ${\mathbf{X}}_0$ include pre-intervention outcomes and predictors. Then for $t>T_0$, the bias of ${\hat{\tau }}_{1t}$ is bounded by a function that is:

• Inversely proportional to $T_0$ (the number of pre-treatment periods)
• Increasing in $J$ (the size of the donor pool)
• Controlled by the quality of fit: ${\mathbf{X}}_1-{\mathbf{X}}_0{\mathbf{W}}^{\ast }$

Implication: A large $T_0$ alone does not guarantee low bias — the synthetic control must also achieve a close pre-treatment fit. Conversely, a close fit with small $T_0$ may still produce substantial bias if the idiosyncratic transitory shocks ${\epsilon }_{jt}$ are large.

Key practical implications of the bias bound:

1. Long pre-treatment windows are valuable. More pre-treatment periods impose more matching constraints on ${\mathbf{\mu }}_j$.
2. Imperfect fit is a warning sign. If ${\mathbf{X}}_1\ne {\mathbf{X}}_0{\mathbf{W}}^{\ast }$, Abadie, Diamond, and Hainmueller (2010) advise against using synthetic controls.
3. Large donor pools can increase bias. A large $J$ gives more flexibility to fit the pre-treatment data but may introduce interpolation biases between the treated unit and distant donor units.

Sparsity: The Geometry of Synthetic Controls

One of synthetic control’s most important properties is sparsity — typically, only a small number of donor units receive nonzero weight.

Theorem: Sparsity of Synthetic Controls (Abadie 2021, Section 3.2)

When ${\mathbf{X}}_1$ falls outside the convex hull of the columns of ${\mathbf{X}}_0$, and the columns of ${\mathbf{X}}_0$ are in general quadratic position, the synthetic control is unique and sparse — with the number of nonzero weights bounded by $k$ (the number of predictors in ${\mathbf{X}}_1$).

Geometric interpretation: The synthetic control is the projection of ${\mathbf{X}}_1$ onto the convex hull of ${\mathbf{X}}_0$. Since ${\mathbf{X}}_1$ is typically outside this hull (curse of dimensionality), the projection touches the hull at a face with at most $k$ vertices.

Sparsity is a feature, not a bug:

• Interpretability: The synthetic counterfactual is a named weighted average of specific donor units (e.g., 42% Austria + 22% United States + …)
• Transparency: The contribution of each donor unit is explicit, allowing subject-matter evaluation of the counterfactual’s plausibility
• Contrast with regression: Regression weights are dense (all units contribute), unrestricted (can be negative), and allow extrapolation — obscuring potential biases

The V Matrix and Variable Selection

The synthetic control optimization (Eq. 7 in Abadie 2021) requires choosing a $k\times k$ positive definite matrix $\mathbf{V}=\text{diag}(v_1,\dots ,v_k)$ that weights the relative importance of each predictor:

$$\underset{\mathbf{W}}{\min }\Vert {\mathbf{X}}_1-{\mathbf{X}}_0\mathbf{W}\Vert ={\left(\sum _{h=1}^k{v}_h{\left(X_{h1}-\sum _{j=2}^{J+1}{w}_j{X}_{hj}\right)}^2\right)}^{1/2}$$

Definition: Cross-Validation for V Matrix Selection

Split the pre-intervention periods $t=1,\dots ,T_0$ into a training period $t=1,\dots ,t_0$ and a validation period $t=t_0+1,\dots ,T_0$ (with $t_0=T_0/2$ as a default). Then:

1. For each candidate $\mathbf{V}$, compute weights $\tilde{\mathbf{W}}(\mathbf{V})$ using training period data only
2. Evaluate the MSPE of the resulting synthetic control on the validation period:

$$\sum _{t=t_0+1}^{T_0}{\left(Y_{1t}-w_2(\mathbf{V})Y_{2t}-\cdots -w_{J+1}(\mathbf{V})Y_{J+1,t}\right)}^2$$

3. Select ${\mathbf{V}}^{\ast }$ that minimizes the validation-period MSPE
4. Use ${\mathbf{W}}^{\ast }=\mathbf{W}({\mathbf{V}}^{\ast })$ for estimation

Simple alternatives to cross-validation:

• $v_h=1/\text{Var}(X_{h1},\dots ,X_{hJ+1})$: rescale all predictors to unit variance (equivalent to standardizing)
• Equal weights $v_h=1/k$: appropriate when predictors are on similar scales

Pre-Intervention Outcomes as Predictors

Including pre-intervention outcome values of $Y_{jt}$ in ${\mathbf{X}}_1$ and ${\mathbf{X}}_0$ is often the single most important variable selection decision. Pre-treatment outcomes are powerful predictors of post-treatment outcomes (via the factor model), and they are automatically absorbed into ${\mathbf{\mu }}_j$ in the factor model framework. However, including many individual time-period outcomes (rather than summary measures) can lead to overfitting during the training period. Use aggregate summaries (means, subperiod averages) as default.

Contrast: Synthetic Control vs. Regression

Property	Synthetic Control	Regression
Weight constraints	$w_j\ge 0$, $\sum w_j=1$	Unconstrained
Extrapolation	Precluded by convex combination	Allowed (negative weights)
Sparsity	Yes (bounded by $k$)	No (dense weights)
Fit transparency	Explicit: ${\mathbf{X}}_1-{\mathbf{X}}_0{\mathbf{W}}^{\ast }$	Hidden: regression forces ${\mathbf{X}}_0{\mathbf{W}}^{reg}={\mathbf{X}}_1$ exactly
Bias when units dissimilar	Visible in large ${\mathbf{X}}_1-{\mathbf{X}}_0{\mathbf{W}}^{\ast }$	Hidden by extrapolation
Pre-analysis plan	Weights registerable before outcomes observed	Cannot preregister

The regression estimator forces a perfect fit of the covariates (${\bar{\mathbf{X}}}_0{\mathbf{W}}^{reg}={\bar{\mathbf{X}}}_1$) even when the untreated units are completely dissimilar to the treated unit, allowing extrapolation. Regression weights in table 3 of Abadie (2021) include negative values for four OECD countries in the German reunification example — synthetic control weights in table 2 are all nonnegative.

See Also

• Synthetic Control — basic estimator, California Prop 99 example, constrained optimization
• Synthetic Control Inference and Diagnostics — how bias theory informs the RMSPE test statistic
• Synthetic Control Requirements — the convex hull condition and contextual requirements derived from this theory
• Differences-in-Differences — the linear factor model nests DiD as a special case (${\lambda }_t=$ constant)
• Abadie 2021 - Overview — full paper overview
• Bayesian Difference in Differences — Bayesian DiD is the parallel-trends special case (${\lambda }_t=$ constant) of this linear factor model