Ridge Augmentation and Extrapolation Diagnostics

Why Ridge Augmentation Needs Diagnostics

Synthetic-control weights combine donor outcomes to construct a counterfactual for a treated unit. Even after choosing weights to match pre-intervention information, the treated unit and its synthetic comparison may differ on measured predictors. Augmentation adds an outcome-model correction for that mismatch. Ridge regression can stabilize this correction when predictors are numerous or correlated, but it does not make unsupported predictions safe: the treated unit may lie outside the range or combinations represented by donors.

This tutorial focuses on the fitted correction and on diagnosing its sensitivity. The correction is not a substitute for the design assumptions required for causal interpretation, including the absence of relevant spillovers and a defensible donor pool. It also does not, by itself, quantify uncertainty in the treatment-effect estimate.

Definition: Let unit \(1\) be treated and let \(j=1,\ldots,J\) index untreated donors. For a post-intervention time \(t\), let \(Y_{1t}\) be the treated unit's observed outcome and \(Y_{jt}\) the observed outcome for donor \(j\). Let \(z_1\) and \(z_j\) be vectors of pre-intervention predictors for the treated unit and donor \(j\). Given synthetic-control weights \(w_j\) with \(w_j\geq 0\) and \(\sum_{j=1}^J w_j=1\), ridge augmentation estimates the treated unit's untreated outcome by adding an outcome-model correction to the weighted donor outcome.

The Ridge-Augmented Estimate

For each post-intervention time \(t\), fit an outcome model among donors, using their post-intervention outcomes as responses and their pre-intervention predictors as features. The donors must remain untreated at time \(t\), and the predictors should be constructed without using post-intervention outcomes. For clarity, suppose the features have been centered and scaled using donor data, and write the ridge model as \(Y_{jt}=a_t+z_j^\top\beta_t+\varepsilon_{jt}\). The intercept is not penalized.

(\hat a_t,\hat\beta_t) = \arg\min_{a,\beta} \left\{ \sum_{j=1}^J (Y_{jt}-a-z_j^\top\beta)^2 +\lambda\sum_{k=1}^p\beta_k^2 \right\},\qquad \lambda\geq 0.

The ridge-augmented estimate of the untreated outcome and the corresponding treatment-effect estimate are

\widehat{Y_{1t}(0)} = \sum_{j=1}^J w_jY_{jt} + (z_1-\sum_{j=1}^Jw_jz_j)^\top\hat\beta_t, \qquad \hat\tau_t=Y_{1t}-\widehat{Y_{1t}(0)}.

The first term is the ordinary synthetic-control prediction. The second is the correction: the fitted outcome-model slope applied to the difference between the treated features and the weighted donor features. Because the weights sum to one, the model intercept cancels from this difference. A positive \(\hat\tau_t\) means the treated outcome exceeds the estimated untreated outcome, in the outcome's original units.

What the Penalty Changes

Ridge regression shrinks fitted slopes toward zero. At \(\lambda=0\), the fit is ordinary least squares when the coefficients are identifiable. As \(\lambda\) increases, the slope vector is generally pulled closer to zero; at very large \(\lambda\), the correction approaches zero and the prediction approaches the ordinary synthetic-control prediction. Ridge does not generally set individual slopes exactly to zero, and it does not automatically identify which predictors matter.

A short derivation shows how the penalty controls the correction. After centering the donor responses and features to remove the unpenalized intercept, let \(Z\) be the donor-by-feature matrix and \(y\) the centered donor outcome vector. Differentiating the ridge objective with respect to \(\beta\) and setting the derivative to zero gives

-2Z^\top(y-Z\hat\beta)+2\lambda\hat\beta=0, \qquad (Z^\top Z+\lambda I)\hat\beta=Z^\top y.

When \(\lambda>0\), \(Z^\top Z+\lambda I\) is invertible, so \(\hat\beta=(Z^\top Z+\lambda I)^{-1}Z^\top y\). This formula explains why ridge can yield a stable coefficient estimate even when features are collinear. It also makes clear that the penalty acts on the outcome-model slopes—not on the synthetic-control weights in this formulation.

Feature scaling matters because the penalty is applied to squared coefficients: a change in measurement units changes the coefficient scale and therefore changes the effective penalty. Scale features using donor information and use the same transformation for the treated unit. State the scaling and the penalty rule; penalty values from differently scaled models are not directly comparable.

Worked Example: Calculating the Ridge Correction and Its Sensitivity

Suppose four hypothetical donors have one pre-intervention feature \(x=(0,1,2,3)\) and post-intervention outcomes \(Y=(2,3,5,6)\), respectively. Assume this feature is already on the chosen analysis scale. The treated unit has \(x_1=4\), and the synthetic-control weights place all weight on the fourth donor. The observed treated outcome is \(Y_{1t}=9\). Fit a ridge line with an unpenalized intercept and \(\lambda=2\).

The donor means are \(\bar x=1.5\) and \(\bar Y=4\). The centered cross-product is \(7\), and the centered sum of squares for \(x\) is \(5\). Thus the ridge slope and intercept are

\hat\beta = \frac{\sum_j(x_j-\bar x)(Y_{jt}-\bar Y)} {\sum_j(x_j-\bar x)^2+\lambda} = \frac{7}{5+2}=1, \qquad \hat a=4-(1)(1.5)=2.5.

The synthetic-control prediction is \(6\). Its feature value is \(3\), so the feature discrepancy is \(4-3=1\) and the ridge correction is \(1\times1=1\). The augmented untreated-outcome estimate is \(6+1=7\), giving \(\hat\tau_t=9-7=2\) outcome units.

For comparison, with \(\lambda=0\), the slope is \(7/5=1.4\); the correction is \(1.4\), the untreated-outcome estimate is \(7.4\), and \(\hat\tau_t=1.6\). With \(\lambda=5\), the slope is \(7/10=0.7\); the correction is \(0.7\), the untreated-outcome estimate is \(6.7\), and \(\hat\tau_t=2.3\). The point estimate changes by \(0.7\) units across these fits. The treated feature value \(4\) is also beyond the donor range \([0,3]\), so this sensitivity is a warning to examine extrapolation, not a reason to select whichever penalty gives a preferred effect.

Diagnosing Support and Extrapolation

The correction depends on both the feature discrepancy \(d=z_1-\sum_jw_jz_j\) and the fitted slope. A large correction can arise from a large discrepancy, a steep slope, or both. Neither a small correction nor a small treatment-effect change across penalties proves that the treated unit is supported by the donors. In multiple dimensions, compensating feature differences can even make the correction small while the treated feature vector lies far outside the donor combinations.

Use diagnostics that examine the feature geometry as well as the correction itself. A useful first check is whether the treated feature vector lies within the convex hull of donor feature vectors: that is, whether it can be written as a weighted average of donor vectors with nonnegative weights summing to one. Being within this hull does not guarantee a good fit or causal identification, but being outside it signals that predictions at the treated features require extrapolation in at least one direction. Also inspect predictor-by-predictor differences between the treated unit and the weighted donor combination, the size and direction of the correction, and how these quantities change under reasonable penalty choices.

Worked Example: A Mismatch Without Extrapolation Beyond the Range

Use the four donor observations from the preceding example, with ridge slope \(\hat\beta=1\) at \(\lambda=2\). Now suppose the treated feature is \(x_1=1.5\), the weights are \(0.5\) on the donor with \(x=0\) and \(0.5\) on the donor with \(x=2\), and the observed treated outcome is \(Y_{1t}=6\).

The weighted donor feature is \(0.5(0)+0.5(2)=1\), so the feature discrepancy is \(1.5-1=0.5\). The weighted donor outcome is \(0.5(2)+0.5(5)=3.5\). The correction is \(0.5(1)=0.5\), so the augmented untreated-outcome estimate is \(3.5+0.5=4.0\), and the estimated effect is \(6-4.0=2.0\) outcome units.

Here \(1.5\) lies inside the donor feature range from \(0\) to \(3\), so the treated feature is not beyond the one-dimensional donor range. Yet the weighted synthetic feature is \(1\), not \(1.5\), and augmentation corrects that mismatch. A nonzero correction is therefore not, by itself, evidence of extrapolation beyond donor support.

Worked Example: A Small Correction Despite Multidimensional Extrapolation

Consider three hypothetical donors with two features \(z_1,z_2\) and post outcomes as follows: \((0,0)\) with outcome \(10\), \((2,0)\) with outcome \(12\), and \((0,2)\) with outcome \(8\). Use equal donor weights for fitting the outcome model, an unpenalized intercept, and \(\lambda=1\). The treated feature vector is \((2,2)\). For the synthetic prediction, place weight \(0.5\) on each of the donors at \((2,0)\) and \((0,2)\); the observed treated outcome is \(13\).

The donor feature means are \((2/3,2/3)\), and the donor outcome mean is \(10\). After centering, the feature cross-product matrix is

Z^\top Z= \begin{pmatrix} 8/3&-4/3\\ -4/3&8/3 \end{pmatrix}, \qquad Z^\top y= \begin{pmatrix} 4\\ -4 \end{pmatrix}.

The ridge equations give \((Z^\top Z+I)\hat\beta=Z^\top y\). By symmetry the solution has \(\hat\beta=(b,-b)\). Substituting into either equation gives \((4+1)b=4\), so \(b=0.8\) and \(\hat\beta=(0.8,-0.8)\). The fitted intercept is \(10\), since the feature means are equal and the two slopes cancel.

The weighted donor feature vector is \((1,1)\), so the discrepancy is \((2,2)-(1,1)=(1,1)\). The correction is \((1,1)^\top(0.8,-0.8)=0\). The weighted donor outcome is \(0.5(12)+0.5(8)=10\), so the augmented untreated-outcome estimate is \(10\), and the estimated effect is \(13-10=3\) outcome units.

Despite the zero correction, the treated vector \((2,2)\) is outside the donors' convex hull: every convex combination of the three donor vectors has coordinates whose sum is at most \(2\), but the treated vector's coordinates sum to \(4\). The distance to the hull is \(2/\sqrt{2}\), or about \(1.414\), in these feature units. The opposing slopes cancel in the correction; they do not remove the extrapolation warning. Inspecting only the correction's magnitude would miss it.

A Practical Sensitivity and Support Check

A disciplined diagnostic separates choices that determine the synthetic comparison from choices that determine the augmentation. Keep the donor pool, predictors, pre-intervention window, and synthetic weights fixed while examining a reasonable range of ridge penalties. This isolates how much the outcome-model correction depends on shrinkage. If these design choices are changed too, report those as separate sensitivity analyses rather than attributing all changes to the penalty.

Specify the outcome model.
List the pre-intervention features, their scaling, the response period, and which coefficients are penalized. Confirm that each donor remains untreated in the response period.

Fit and compare penalties.
Report the ridge correction and augmented estimate over a prespecified, defensible penalty range. If using donor-based cross-validation to select a penalty, do not choose it to produce a preferred treated-unit effect.

Inspect discrepancies and support.
Report the feature differences \(z_1-\sum_jw_jz_j\), check whether \(z_1\) is in the donors' convex hull, and identify dimensions or combinations that require extrapolation.

Interpret the estimate cautiously.
Describe whether the correction is stable and whether the treated features are supported. Treat substantial penalty sensitivity or unsupported features as limitations on the credibility of the counterfactual.

Leave-one-donor-out prediction error can help assess how well the outcome model predicts outcomes for donors omitted from its fit. It is a diagnostic of prediction among the available donors, not proof that the model predicts well at an unsupported treated feature vector. With few donors, such validation can itself be noisy. Penalty selection and fit diagnostics should therefore be accompanied by the feature-support assessment, not used in place of it.

Common Mistake: Treating Shrinkage as a Cure for Extrapolation

A tempting argument is: “Ridge shrinks the coefficients, so the augmented prediction cannot extrapolate far.” This is incorrect. Shrinkage limits coefficient magnitudes relative to the chosen feature scaling, but it does not move the treated feature vector into the donor support. A predictor far outside donor values can still yield a model-dependent prediction, and multiple unsupported feature differences can cancel in the total correction. Conversely, a noticeable correction can simply repair a moderate mismatch within donor support.

Report the correction and the geometry separately. A useful account states the penalty and scaling, the correction in outcome units, how the estimate varies across defensible penalties, and whether the treated feature vector lies within donor support. These diagnostics do not establish the identification assumptions or replace uncertainty analysis. They clarify how much the augmented estimate relies on the outcome model and on extending that model beyond observed donor combinations.

Key takeaway: Ridge augmentation adds a penalized outcome-model adjustment to the synthetic-control prediction. Penalty sensitivity reveals dependence on shrinkage; feature discrepancies and donor-support checks reveal where the adjustment is being applied. Neither a small correction nor stable estimates across penalties rule out extrapolation.

Check Your Understanding

Use the definitions and examples above to test your reasoning.

Why does the intercept cancel from the ridge correction when synthetic-control weights sum to one?
In the first worked example, what happens to the augmentation as the penalty becomes very large, and why?
Can the treated feature vector be within the donor range while the correction is nonzero? Explain using the second worked example.
Why does a zero correction fail to show that the treated unit is within donor support?
Name two distinct diagnostics that should accompany a report of penalty sensitivity.