Building on the Event-Study Designs in Chapter 5, Chapter 6 introduces the Synthetic Control Method (SCM). While event studies and Difference-in-Differences (DiD) rely on the powerful but restrictive assumption of parallel trends, SCM provides a flexible alternative by constructing a custom counterfactual for the treated unit.

SCM is particularly valuable in marketing when:

  1. Parallel Trends are Implausible: The treated unit (e.g., a flagship market or a specific country) follows a unique trajectory that no single control unit matches.
  2. Donor Pool is Heterogeneous: Many potential control units exist, but none is a “perfect” match on its own.
  3. Transparency is Critical: Stakeholders need to see exactly which markets are being used as comparisons and with what weights.

The Logic of Synthetic Control

The core idea is to find a weighted average of untreated “donor” units that closely matches the treated unit’s pre-treatment outcome path.

If $Y_{1t}$ is the outcome for the treated unit and $\{Y_{jt}\}_{j \in J}$ are the outcomes for the donor pool $J$, the synthetic control is:

$$ \hat{Y}_{1t}(0) = \sum_{j \in J} w_j^* Y_{jt} $$

where the weights $w_j^*$ are chosen to minimize the difference between the treated unit and its synthetic counterpart during the pre-treatment period.

The Factor Loading Perspective

Identification in SCM relies on a latent factor structure. If outcomes are driven by common factors $f_t$ and unit-specific loadings $\lambda_j$:

$$ Y_{jt}(0) = \alpha_j + \lambda_t + f_t' \lambda_j + \epsilon_{jt} $$

The synthetic control is approximately unbiased if we can find weights that match the treated unit’s factor loadings:

$$ \lambda_1 = \sum_{j \in J} w_j^* \lambda_j $$

This implies that the treated unit must lie within the convex hull of the donor pool’s factor loadings. If the treated unit is an extreme outlier (e.g., the largest market), no combination of smaller donors can perfectly replicate its dynamics.

The Convexity Constraint

A defining feature of SCM is that weights are non-negative ($w_j \ge 0$) and sum to one ($\sum w_j = 1$). This convexity constraint handles several critical issues:

  • No Extrapolation: It ensures the counterfactual is an interpolation within the observed donor data, rather than an extrapolation beyond it.
  • Realistic Counterfactuals: It prevents the model from assigning negative weights, which could produce “impossible” counterfactuals unsupported by any real-world unit.
  • Regularization: The constraint acts as an implicit regularizer, discouraging extreme weights and reducing variance (at the cost of some potential bias).

Limitations and Overfitting

While SCM is powerful, it is not a “magic bullet.” Good pre-treatment fit (measured by RMSPE) is a necessary diagnostic, but it does not guarantee low post-treatment bias.

  • Overfitting to Noise: A model that fits every idiosyncratic wiggle in the pre-treatment data may be fitting noise rather than underlying signal, leading to poor generalization after the intervention.
  • Specification Search: The choice of predictors (outcomes, covariates) and their importance weights ($V$) can significantly affect results. Transparency in the final weights can hide the “search” process used to find them.

Conclusion

Synthetic Control shifts the focus from finding a single comparable unit to constructing an optimal composite unit. By explicitly modeling the pre-treatment match and adhering to the convexity constraint, SCM provides a transparent and theoretically grounded framework for causal inference in complex panels where parallel trends cannot be assumed.