Bias Correction in Augmented Synthetic Control

Why Add a Correction to a Synthetic Control?

A synthetic control estimates the untreated outcome of a treated unit by combining outcomes from untreated donor units. Even when the donor weights are chosen to make the units similar before intervention, the weighted donors may not match the treated unit on every pre-intervention predictor. If untreated outcomes depend on those predictors, residual imbalance can translate into bias in the weighted post-intervention outcome.

Augmented synthetic control keeps the weighted donor estimate and adds an outcome-model adjustment for that imbalance. The correction is not a replacement for a credible donor pool or a suitable comparison design. It relies on an outcome model and on conditions linking the donors to the treated unit.

Definition: Let unit 1 be treated and let units \(j=2,\ldots,J+1\) be untreated donors. At a post-intervention time \(t\), let \(Y_{it}(0)\) denote unit \(i\)'s outcome without treatment and \(Y_{it}(1)\) its outcome with treatment. The period-specific treatment effect is \(\tau_{1t}=Y_{1t}(1)-Y_{1t}(0)\). Let \(X_1\) be a \(p\)-dimensional vector of pre-intervention predictors for the treated unit, and let \(X_0\) be the \(p\times J\) matrix whose columns are the corresponding donor vectors. Given synthetic control weights \(w=(w_2,\ldots,w_{J+1})'\), define the weighted predictor discrepancy as \(D=X_1-X_0w\).

The Augmented Estimate

In the usual synthetic control construction, the weights are nonnegative and sum to one. The weighted donor outcome at time \(t\) is \(\sum_{j=2}^{J+1}w_jY_{jt}\). An outcome model supplies a coefficient vector \(\hat\beta_t\) relating the predictors to the untreated outcome. The augmented estimate of the treated unit's untreated outcome is

\hat Y_{1t}^{\,ASCM}(0) = \sum_{j=2}^{J+1}w_jY_{jt} + (X_1-X_0w)'\hat\beta_t.

The first term is the synthetic estimate. The second term adjusts it according to the predictor discrepancy. For example, if a predictor is higher for the treated unit than for the weighted donors, and higher values predict higher untreated outcomes, its contribution to the correction is positive. The estimated treatment effect is the observed treated outcome minus the augmented counterfactual:

\hat\tau_{1t}^{\,ASCM} = Y_{1t} - \hat Y_{1t}^{\,ASCM}(0).

A useful model for understanding the correction is an affine untreated-outcome mean, \(m_t(X)=a_t+X'\beta_t\). Here \(a_t\) is an intercept and \(\beta_t\) describes how the untreated outcome changes with the predictors. Because the weights sum to one, the weighted intercept equals \(a_t\), so the difference between the treated unit's model prediction and the weighted donors' model prediction is exactly \(D'\beta_t\). The correction therefore targets the model-predicted part of the imbalance.

Worked Example: A One-Predictor Correction

Suppose three hypothetical donors have pre-intervention predictor values \(2,4,6\) and post-intervention untreated outcomes \(10,14,18\), respectively. The treated unit has predictor value \(5\). The synthetic control weights are \(0.5,0.3,0.2\). Assume the untreated outcome model is \(m(x)=6+2x\), and that it applies to the treated unit as well as the donors. For this illustration, the donor outcomes lie exactly on that model.

First check the weights: they are nonnegative and sum to \(0.5+0.3+0.2=1\). The weighted donor outcome is \(0.5(10)+0.3(14)+0.2(18)=5+4.2+3.6=12.8\). The weighted predictor is \(0.5(2)+0.3(4)+0.2(6)=1+1.2+1.2=3.4\). Thus the treated-minus-weighted predictor discrepancy is \(5-3.4=1.6\).

The model slope is \(2\), so the correction is \(1.6(2)=3.2\). The augmented counterfactual is \(12.8+3.2=16.0\). This agrees with the model prediction for the treated unit, \(6+2(5)=16\), as it should when the model holds exactly. If the treated unit's observed post-intervention outcome is \(20\), the estimated treatment effect is \(20-16.0=4.0\) outcome units.

The conclusion depends on the stated assumptions: the donor units are untreated and unaffected by the intervention, the model describes the untreated outcome relationship for the treated unit, and the observed treated outcome is its outcome under treatment. The calculation alone cannot establish those assumptions.

Why the Correction Works

The cancellation can be shown directly. Suppose that at the post-intervention time, untreated outcomes follow an affine conditional mean plus a residual: \(Y_{it}(0)=a_t+X_i'\beta_t+e_{it}\). This representation is a modeling assumption, not an identity known from the data. The residual \(e_{it}\) is the part not accounted for by the predictors. With weights summing to one, substitute this representation into the augmented counterfactual and subtract the treated unit's untreated outcome:

\begin{aligned} \hat Y_{1t}^{\,ASCM}(0)-Y_{1t}(0) &= \sum_j w_j\{a_t+X_j'\beta_t+e_{jt}\} +D'\hat\beta_t -\{a_t+X_1'\beta_t+e_{1t}\}\\ &= D'(\hat\beta_t-\beta_t) +\sum_j w_je_{jt}-e_{1t}. \end{aligned}

In the second line, the intercept terms cancel because \(\sum_jw_j=1\). The predictor terms combine to \((X_0w-X_1)'\beta_t=-D'\beta_t\), which cancels the model's part of the outcome discrepancy when the correction uses \(\beta_t\). If the coefficient estimate is imperfect, the remaining coefficient-estimation term is \(D'(\hat\beta_t-\beta_t)\). The other remaining term is the weighted donor residual minus the treated-unit residual.

Result: Under the affine untreated-outcome model and weights that sum to one, the augmented counterfactual error is \(\hat Y_{1t}^{\,ASCM}(0)-Y_{1t}(0)=D'(\hat\beta_t-\beta_t)+\sum_jw_je_{jt}-e_{1t}\). Thus the correction removes the model-predicted component of predictor imbalance. Consistency additionally requires the coefficient-estimation error and the residual contrast to vanish, or otherwise be negligible, under the chosen asymptotic or design argument.

This identity explains both the attraction and the limitation of augmentation. If \(\hat\beta_t\) consistently estimates \(\beta_t\) and \(D\) remains bounded, then \(D'(\hat\beta_t-\beta_t)\) tends to zero. But a correct regression mean alone does not eliminate the residual contrast in every finite sample. For consistency, that contrast must also vanish in the relevant sequence of designs, for example because the weighted residuals reproduce the treated unit's residual, or because the design and averaging assumptions make the residual difference negligible.

A second route is strong predictor balance: if \(D=0\), the correction is zero regardless of \(\hat\beta_t\), and the affine model's predictor component is already balanced. This route still needs the donor-weighted residual outcome to represent the treated unit's untreated residual outcome. More generally, one must justify that the weighted synthetic estimate is valid for the untreated potential outcome, as required by the synthetic control identification assumptions.

Specify the target and comparison.
Define the treated unit's untreated post-intervention outcome as the counterfactual target. Use donor units that remain untreated and are not affected by the intervention.

Construct the weighted estimate.
Choose synthetic control weights from pre-intervention information. Confirm that they satisfy the intended constraints, including nonnegativity and a sum of one when using the standard convex combination.

Fit an outcome relationship.
Estimate the predictor-outcome coefficients using untreated information. The fitted relationship must be appropriate for the post-intervention untreated outcome being predicted, not a relationship created by the treated unit's intervention-affected outcome.

Apply and assess the correction.
Add \(D'\hat\beta_t\) to the weighted donor outcome. Interpret consistency only with support for the outcome model, coefficient estimation, and the residual comparison described by the error identity.

Worked Calculations with Multiple Predictors

Worked Example: Adjusting for Two Predictors

Consider a hypothetical analysis with two predictors and two donors. The treated predictor vector is \(X_1=(4,7)'\). Donor A has \(X_A=(2,8)'\), donor B has \(X_B=(5,4)'\), and the weights are \(0.6\) and \(0.4\). A fitted untreated-outcome model is \(m(X)=8+1.5X_1-0.5X_2\). Suppose the donors' post-intervention outcomes are \(7\) and \(13.5\), consistent here with that model, and the treated unit's observed outcome is \(14\).

The weights sum to one. The weighted predictor vector is \(0.6(2,8)'+0.4(5,4)'=(1.2,4.8)'+(2,1.6)'=(3.2,6.4)'\). Therefore \(D=(4,7)'-(3.2,6.4)'=(0.8,0.6)'\). The weighted donor outcome is \(0.6(7)+0.4(13.5)=4.2+5.4=9.6\).

The model coefficient vector is \(\hat\beta=(1.5,-0.5)'\), so the correction is \(D'\hat\beta=0.8(1.5)+0.6(-0.5)=1.2-0.3=0.9\). The augmented counterfactual is \(9.6+0.9=10.5\). Directly applying the model to the treated predictors gives \(8+1.5(4)-0.5(7)=8+6-3.5=10.5\), which verifies the arithmetic under this exact-model illustration. The estimated treatment effect is \(14-10.5=3.5\) outcome units.

This check confirms the model calculation, not causal identification by itself. A causal interpretation still requires that the donors provide valid untreated comparisons, that there is no anticipation or interference that invalidates the comparison, and that the outcome model is suitable for the treated unit's untreated outcome.

Worked Example: Exact Predictor Balance

Suppose one predictor has value \(4\) for the treated unit and values \(2\) and \(6\) for two hypothetical donors. Use weights \(0.5\) and \(0.5\). Assume the untreated outcome follows \(m(x)=5+3x\), with donor post-intervention outcomes \(11\) and \(23\). The treated unit's observed outcome is \(22\).

The weights sum to one, and the weighted predictor is \(0.5(2)+0.5(6)=1+3=4\), exactly the treated value. Thus \(D=4-4=0\), so the correction is \(0\hat\beta_t=0\), whatever coefficient estimate is used. The weighted donor outcome is \(0.5(11)+0.5(23)=5.5+11.5=17\). Under the stated model, the treated unit's untreated outcome is \(5+3(4)=17\). The estimated treatment effect is \(22-17=5\) outcome units.

Here exact predictor balance means that the model-based adjustment cannot change the estimate. The causal conclusion still rests on the untreated-outcome and donor-comparability assumptions; balance on this one predictor does not guarantee balance on omitted features that affect outcomes.

Conditions and Practical Interpretation

A consistency claim is conditional on a design and model, not a consequence of writing down the augmentation formula. At a minimum, interpretation requires the standard synthetic control conditions for the setting: the intervention and its timing are well defined; donor outcomes represent untreated outcomes; donors are not exposed to treatment spillovers; and there is no anticipation that changes outcomes before treatment. Predictor measurement and model estimation must also use information that is appropriate for the untreated counterfactual.

For the correction specifically, useful sufficient conditions follow from the error decomposition. The untreated conditional mean should be adequately represented by the chosen outcome model over the treated unit's predictor values and the relevant donors. The coefficient estimator should converge to the relevant coefficient, with \(D'(\hat\beta_t-\beta_t)\) tending to zero. Finally, the weighted donor residual contrast \(\sum_jw_je_{jt}-e_{1t}\) must tend to zero or be controlled by the design's identification argument. Exact balance \(D=0\) removes dependence on the fitted coefficient for the adjustment itself, but does not by itself establish that residual condition.

The outcome model may be estimated from untreated donor information, such as donor outcomes and pre-intervention predictors. The specific fitting strategy must match the target: it should estimate the untreated outcome relationship at the post-intervention time or otherwise justify transporting a relationship learned from earlier periods. The treated unit's post-intervention outcome must not be used as though it were an untreated training outcome.

A large correction deserves attention. It means the weighted donors differ materially from the treated unit on predictors that the fitted model associates with the outcome. Such a correction can reduce bias when the model is credible, but it can also make the estimate depend heavily on a model prediction beyond the information supported by the donors. Augmentation does not turn poor support or a misspecified model into reliable extrapolation.

Common Mistake

A frequent error is to treat the correction as an automatic guarantee of unbiasedness: “The regression adjusts for the mismatch, so the remaining estimate is causal.” The formula does not justify that conclusion. The correction only addresses the modeled component of predictor imbalance. If the slope is wrong, if the treated predictors are unsupported by donor data, or if weighted residual outcomes do not represent the treated unit's residual, the augmented counterfactual can still be wrong. The appropriate correction is to state the outcome-model and donor-comparability assumptions, calculate the adjustment carefully, and interpret the result as causal only when the identification conditions are defensible.

Key takeaway: Augmented synthetic control equals a weighted donor outcome plus the fitted outcome change associated with the treated-minus-weighted predictor discrepancy. Under an affine untreated-outcome model, that adjustment cancels the model-predicted imbalance; consistency also requires appropriate coefficient estimation and control of the remaining residual contrast.

Check Your Understanding

Use the definitions and error decomposition above to answer these questions.

In the augmented estimate, what does \(D\) measure, and why does the correction vanish when \(D=0\)?
With weights \(0.7\) and \(0.3\), donor predictor values \(2\) and \(6\), treated predictor value \(4\), and model slope \(1.5\), calculate the correction.
Why does the intercept cancel from the affine-model derivation when the weights sum to one?
Under the affine model, name the two remaining components of the augmented counterfactual error when \(\hat\beta_t\) is estimated rather than known.
Why is a correct outcome-model mean not, by itself, sufficient to guarantee consistency in every finite sample?
What does a large correction indicate, and why does it not automatically establish that the augmented estimate is reliable?