Augmented Synthetic Control and Bias Correction

Why Augment a Synthetic Control?

A synthetic control estimates a treated unit’s untreated outcome by combining outcomes from untreated donor units. When the donor combination does not closely match the treated unit on predictors, however, the resulting counterfactual may retain systematic error. Augmented synthetic control (ASC) adds an outcome-model correction: it adjusts the synthetic outcome by the difference between the model’s prediction for the treated unit and its weighted predictions for the donors.

This correction can improve a counterfactual when the donor weights leave residual predictor imbalance and the outcome model captures how that imbalance relates to outcomes. It is not an automatic guarantee of lower error. The model may be estimated imprecisely, regularization may shrink predictions, and the treated unit may lie far beyond the donor units’ predictor range.

Definition: Let unit 1 be treated, and let $j=1,\ldots,J$ index untreated donor units. Let $X_i$ be a vector of predictors measured before treatment, and let $w_j$ be synthetic-control weights satisfying $w_j\geq 0$ and $\sum_{j=1}^{J}w_j=1$. Write $X_w=\sum_{j=1}^{J}w_jX_j$ for the weighted donor predictor vector. Given a fitted untreated-outcome model $\hat m_t(X)$, the augmented synthetic control estimate of unit 1’s untreated outcome at post-treatment time $t$ is $$ \hat Y_{1t}^{\,ASC}(0) = \sum_{j=1}^{J}w_jY_{jt}(0) + \left[ \hat m_t(X_1)-\sum_{j=1}^{J}w_j\hat m_t(X_j) \right]. $$ The bracketed term is the outcome-model correction. In applications, the donor outcomes in the first term are observed untreated outcomes.

The correction is a prediction gap, not simply the model’s prediction for the treated unit. It compares that prediction with the same model’s weighted prediction for the donor mixture. If the fitted model is linear, $\hat m_t(X)=X'\hat\beta_t$, define the residual predictor imbalance $d=X_1-X_w$. The correction then equals $d'\hat\beta_t$. Thus exact balance on the included predictors, $d=0$, makes the correction zero.

From Counterfactuals to Treatment Effects

Let $Y_{1t}(1)$ denote the treated unit’s observed outcome under treatment, and let $Y_{1t}(0)$ be its unobserved outcome without treatment. The estimated treatment effect at time $t$ is the observed treated outcome minus the estimated untreated counterfactual:

\hat\tau_t^{\,ASC} = Y_{1t}(1)-\hat Y_{1t}^{\,ASC}(0).

For an average effect over a set of post-treatment periods $\mathcal T_1$, one may average these period-specific differences: $\hat\tau^{\,ASC}=|\mathcal T_1|^{-1}\sum_{t\in\mathcal T_1}\hat\tau_t^{\,ASC}$. The estimate is a causal effect only under a credible design: donor units must represent the treated unit’s untreated trajectory, donors must remain untreated and unaffected by the intervention, and there must be no anticipation that changes pre-treatment outcomes. Predictor selection, weights, and model fitting should use information available without using treated post-treatment outcomes to tune the counterfactual toward a desired effect.

Worked Example: Applying the Prediction-Gap Correction

Suppose two donors have predictor values $X_1=8$ and $X_2=12$, and their synthetic weights are $0.6$ and $0.4$. The treated unit’s predictor value is $X_{\mathrm{treated}}=10$. At a post-treatment period, the donor untreated outcomes are 30 and 34, while a fitted linear outcome model has slope $\hat\beta=2$ and no intercept. The treated unit’s observed outcome is 40. Calculate the ASC counterfactual and treatment-effect estimate.

The weighted donor predictor is $X_w=0.6(8)+0.4(12)=9.6$, so the predictor imbalance is $d=10-9.6=0.4$. The weighted donor outcome is $0.6(30)+0.4(34)=31.6$. The correction is $d\hat\beta=0.4(2)=0.8$. Therefore,

\hat Y_{1t}^{\,ASC}(0)=31.6+0.8=32.4, \qquad \hat\tau_t^{\,ASC}=40-32.4=7.6.

The unadjusted synthetic-control effect would be $40-31.6=8.4$; the outcome-model correction lowers it by $0.8$. The result is an estimated increase of 7.6 outcome units for this period, not a claim that the correction must improve accuracy in every application.

What the Correction Can Fix

To see the roles of model error and residual imbalance, suppose the untreated outcome satisfies a linear model $Y_{it}(0)=X_i'\beta_t+u_{it}$, where $\beta_t$ is the true coefficient vector at time $t$ and $u_{it}$ is the part not explained by the predictors. Suppose the fitted model is $\hat m_t(X)=X'\hat\beta_t$. Subtract the treated unit’s true untreated outcome from its ASC estimate. Since the weighted donor outcome is $X_w'\beta_t+\sum_jw_ju_{jt}$, the difference is

\begin{aligned} \hat Y_{1t}^{\,ASC}(0)-Y_{1t}(0) &=X_w'\beta_t+\sum_{j=1}^{J}w_ju_{jt} +(X_1-X_w)'\hat\beta_t -X_1'\beta_t-u_{1t}\\ &=(X_1-X_w)'(\hat\beta_t-\beta_t) +\sum_{j=1}^{J}w_ju_{jt}-u_{1t}\\ &=d'(\hat\beta_t-\beta_t) +\sum_{j=1}^{J}w_ju_{jt}-u_{1t}. \end{aligned}

This identity separates two sources of counterfactual error: coefficient-estimation error multiplied by the predictor imbalance, and the treated-versus-weighted-donor residual difference. The sign of the first term follows from defining $d=X_1-X_w$. When $d=0$, the coefficient-estimation term vanishes; the residual contrast need not vanish. A well-balanced predictor vector therefore removes one channel of model error, but does not by itself guarantee a correct counterfactual.

Key result: Under the stated linear untreated-outcome model, ASC’s counterfactual error is $$ \hat Y_{1t}^{\,ASC}(0)-Y_{1t}(0) = d'(\hat\beta_t-\beta_t) +\sum_{j=1}^{J}w_ju_{jt}-u_{1t}. $$ If the model coefficients are estimated accurately, the correction can reduce error due to $d\neq 0$. If the weighted donor residuals differ substantially from the treated unit’s residual, or the fitted coefficients are poor, ASC may still be biased.

Worked Example: Checking the Error Decomposition

Consider one predictor and one post-treatment period. Let $X_1=10$, $X_w=8$, the true slope be $\beta_t=3$, and the fitted slope be $\hat\beta_t=3.4$. Suppose the weighted donor residual is $\sum_jw_ju_{jt}=0.5$, and the treated unit’s residual is $u_{1t}=1.2$. Calculate the ASC counterfactual error using the decomposition, then verify it directly.

Here $d=X_1-X_w=10-8=2$. The decomposition gives

d(\hat\beta_t-\beta_t)+\sum_jw_ju_{jt}-u_{1t} =2(3.4-3)+(0.5-1.2)=0.8-0.7=0.1.

For a direct check, the weighted donor untreated outcome is $8(3)+0.5=24.5$. The correction is $d\hat\beta_t=2(3.4)=6.8$, so the ASC counterfactual is $31.3$. The true treated untreated outcome is $10(3)+1.2=31.2$. Hence the ASC estimate minus the true untreated outcome is $31.3-31.2=0.1$, matching the decomposition. The positive error means this estimated counterfactual is 0.1 outcome units too high.

Regularization and Extrapolation

The outcome model is often estimated from a limited number of donor units or predictors. Regularization constrains the fitted coefficients to reduce sensitivity to the donor data. For example, ridge regression estimates coefficients by minimizing a squared-error objective plus a penalty proportional to the sum of squared coefficients. Lasso uses a penalty proportional to the sum of absolute coefficient values and can set some fitted coefficients to zero. These procedures can stabilize predictions, but they change the prediction gap used in ASC; they do not directly alter the synthetic weights.

The distinction matters because regularization can trade variance for bias. Shrinking coefficients may reduce the impact of noisy estimates, particularly when predictors are highly correlated, but can also understate the outcome change associated with a predictor imbalance. The correction is especially sensitive to this trade-off when the treated unit’s predictor vector is far from the donor mixture.

Worked Example: Ridge Shrinkage Under Extrapolation

Suppose an outcome model without an intercept is fitted to two donor observations: predictor values $-1$ and $1$, with untreated outcomes $-2$ and $2$. Ridge regression minimizes $\sum_i(Y_i-\beta X_i)^2+\lambda\beta^2$, with $\lambda=2$. A treated unit has predictor value $5$, and the synthetic donor mixture has predictor value $1$. For illustration, suppose the true untreated relationship is $Y(0)=2X$ with no residual error. Compare the ridge-augmented counterfactual with the one using the unpenalized fitted slope.

For these data, $\sum_iX_iY_i=(-1)(-2)+(1)(2)=4$ and $\sum_iX_i^2=2$. Differentiating the ridge objective gives $\hat\beta_{\mathrm{ridge}}=\sum_iX_iY_i/(\sum_iX_i^2+\lambda)=4/(2+2)=1$. The predictor gap is $d=5-1=4$, so the ridge correction is $4(1)=4$. The donor at $X=1$ has outcome 2, giving an ASC counterfactual of $2+4=6$.

Without the penalty, the fitted slope is $4/2=2$, the correction is $4(2)=8$, and the counterfactual is $2+8=10$. Under the stipulated true relationship, the treated unit’s untreated outcome is $2(5)=10$. In this example, ridge shrinkage understates the correction by 4 outcome units. The treated predictor value 5 lies well outside the donor predictor range from $-1$ to $1$; the example illustrates why regularization does not make a long extrapolation reliable.

Practical Conditions and Diagnostics

ASC is most persuasive when the donor combination is reasonably comparable to the treated unit, the outcome model is fitted using appropriate untreated information, and the correction is not driven by extreme extrapolation. The method does not require the synthetic weights to achieve perfect predictor balance: the correction is designed to address the remaining gap. But that feature is useful only to the extent that the fitted relationship maps the gap to a credible outcome adjustment.

Several checks help make the adjustment transparent:

Report the predictor imbalance $d=X_1-X_w$, not only the final counterfactual. A large gap makes the result more dependent on the outcome model.
Describe how the outcome model was fitted and regularized, including the data used to select penalty strength.
Compare the fitted predictions for the treated unit and weighted donor mixture. A large prediction gap warrants scrutiny, especially if the treated predictors lie outside donor support.
Examine pre-treatment fit and placebo diagnostics using the relevant procedures. Good pre-treatment fit is informative but does not prove that post-treatment untreated outcomes would have followed the same pattern.

Common pitfall: It is incorrect to argue that adding a prediction gap must remove bias because the treated unit’s prediction is more individualized. ASC adjusts only for the model-predicted part of the treated-donor difference. If coefficients are inaccurate or the treated unit requires extrapolation, the correction can add error. The relevant question is whether the adjusted prediction is supported by the donor data and a credible outcome relationship.

The central advantage of ASC is that it combines two sources of information: a weighted donor outcome and a model-based adjustment for remaining predictor imbalance. Its central vulnerability is the same combination: the adjustment depends on both the size of the imbalance and the reliability of predictions across that gap. Reporting the unadjusted fit, the correction, and the degree of extrapolation makes that dependence visible.

Check Your Understanding

Answer each question using the definitions and calculations above.

Write the ASC counterfactual for a linear outcome model and define the predictor imbalance $d$.
If the treated and weighted donor predictor vectors match exactly, which component of the counterfactual error decomposition disappears? Which component may remain?
Suppose $d=3$, $\hat\beta_t-\beta_t=-0.2$, and the weighted donor residual minus the treated residual is $0.5$. Calculate the ASC counterfactual error.
Why can regularization reduce sensitivity to noisy donor data while still making the ASC correction less accurate?
What does it mean for the treated unit’s prediction to involve extrapolation, and why is a large predictor gap a warning sign?