7.1 Motivation and Setup

Credibility and efficiency often pull in opposite directions. Design-based methods earn trust by avoiding strong functional-form assumptions, but they pay a price. When the treated unit lies outside the donor pool’s convex hull, as discussed in Section 6.1, weighting alone cannot close the gap. The fit deteriorates, bias creeps in, and pure synthetic control, for all its transparency, sometimes fails. In the factor-model perspective from Chapter 6, this failure corresponds to a mismatch in the treated unit’s latent factor loadings relative to any convex combination of donor loadings, so that the remaining imputation error can be first-order. This chapter confronts that failure and offers a resolution.

Hybrid methods blend weighting with outcome modelling to gain flexibility while staying anchored to the design logic of synthetic control. They can improve pre-treatment fit and reduce variance, but they also introduce modelling and stability assumptions. The question is whether the compromise works, and when it does not.

Consider what happens when a retailer pilots a loyalty programme in five flagship stores. The pilot runs for two years. The retailer has perhaps twenty control stores, but none resembles the flagships. These stores anchor high-traffic urban centres, draw different customer segments, and generate revenue patterns that no convex combination of suburban or regional stores can replicate. Standard synthetic control produces weights, but the pre-treatment fit is poor. The gap between the synthetic and actual trajectories signals trouble. Any post-treatment estimate carries that residual bias forward.

Hybrid methods attack this problem from three directions. Augmented synthetic control pairs the weighting estimator with a regression adjustment that corrects for the residual imbalance. If the weights produce a synthetic control that undershoots the treated unit’s pre-treatment revenue by five per cent, the regression model estimates that gap and uses it to re-centre the synthetic control’s counterfactual trajectory before computing the treatment effect. Regularised synthetic control takes a different path: it shrinks the weights toward simplicity, trading a small increase in bias for a larger reduction in variance. Synthetic difference-in-differences introduces time weights alongside unit weights, aligning both the cross-sectional and temporal dimensions before computing contrasts.

Each approach relaxes a constraint that pure synthetic control imposes. They respond to two distinct failure modes. The first is interpolation failure: when the treated unit sits outside the donor convex hull, no choice of non-negative weights can match its untreated path. The second is overfitting: when the donor pool is large or the pre-period is noisy, the optimisation can chase idiosyncratic fluctuations and produce a synthetic path that does not generalise.

You gain flexibility, but flexibility has costs. Adding a regression adjustment introduces model dependence. Regularisation requires tuning choices that practitioners must justify. Time weights rely on a reweighted parallel-trends-style comparison that may or may not be credible. Hybrid methods are not strictly superior to their predecessors. They occupy a different point on the bias–variance frontier, and the right choice depends on the data structure you face.

The factor model foundation developed in Chapter 6 remains central. As defined there, untreated potential outcomes decompose into unit-specific loadings and time-varying factors. When the treated unit’s loadings lie outside the convex hull of donor loadings, pure synthetic control cannot recover the counterfactual without bias. Hybrid methods modify the constraints that govern this imputation. Augmented synthetic control uses a regression adjustment to repair residual mismatch after weighting. Regularised variants penalise extreme weight vectors and can stabilise estimates when plain synthetic control overfits. Synthetic difference-in-differences reweights both units and pre-treatment periods, changing which comparisons carry identifying weight.

Return to the loyalty programme. The five flagship stores differ from controls not just in observable characteristics such as square footage, foot traffic, and product mix, but in unobservable ways that load onto latent factors. Urban flagship stores may respond more strongly to macroeconomic shocks, and their customer base may exhibit different seasonal patterns. Pure synthetic control tries to match on pre-treatment outcomes, hoping the match implicitly captures these factor loadings. When it fails, the treated unit sits outside the convex hull, and no combination of donors can replicate its trajectory.

Augmented synthetic control (ASCM) offers a partial fix. You fit the best synthetic control you can, then estimate a regression of outcomes on covariates or lagged outcomes within the control group. You apply that regression to adjust the synthetic control’s predictions for the treated unit. Under conditions we make precise later (for example, correctly specified factor structure or correctly specified regression adjustment within the control group), this estimator has a doubly robust flavour: it can remain close to unbiased if either the weighting scheme or the regression adjustment is well specified. But “doubly robust” does not mean “doubly correct”. In small samples, and marketing panels are almost always small in the relevant dimension, misspecification on either side can still generate finite-sample bias. The insurance policy has limits.

Regularisation addresses a different pathology. When the donor pool is large, synthetic control can overfit. The optimisation finds weights that match pre-treatment outcomes precisely, but those weights may be chasing noise. A donor that happens to correlate with the treated unit’s idiosyncratic pre-treatment fluctuations receives weight it does not deserve. Out of sample, the synthetic control performs poorly. In variance terms, highly concentrated weight vectors (small effective number of donors) inflate the idiosyncratic component $\sigma^2 (1 + \sum_j (w_j^*)^2)$ from Section 6.4, making overfitted synthetic controls particularly unstable. Ridge-type penalties shrink weights toward uniformity or toward zero. Elastic-net variants combine sparsity with shrinkage. The bias–variance trade-off shifts: you accept some pre-treatment imbalance to avoid fitting noise.

Synthetic difference-in-differences (SDID) makes a structural innovation. Identification in SDID relies on a reweighted common-trend condition: after applying the estimated unit and time weights, untreated potential outcomes for treated and control units follow parallel trends in the reweighted space. SDID does not make parallel trends true by construction. It shifts the identifying burden onto whether the reweighted comparison is credible in your application. Standard synthetic control constructs a weighted average of donor units. SDID also constructs a weighted average of pre-treatment time periods. The estimator computes a double contrast: treated minus synthetic control, post-period minus a weighted pre-period. This structure nests both difference-in-differences, with equal unit weights and equal time weights, and synthetic control, with optimised unit weights and equal time weights, as special cases. By optimising both sets of weights, SDID can handle settings where neither pure method would succeed, provided its reweighted common-trend assumptions are credible.

Now consider a brand launching campaigns in twenty markets over three years. Markets adopt treatment at different times. Some start early, others late, some never. This is staggered adoption, a setting introduced in Chapter 4, and it complicates everything. You cannot simply pool post-treatment observations, because early adopters contribute to the donor pool for late adopters. The safe way to organise a staggered analysis is to define cohort–time effects $\tau(g, t)$ relative to not-yet-treated or never-treated donors, and then aggregate them into event-time summaries $\theta_k$ with transparent, non-negative weights. Hybrid methods aim to improve the quality of each $\tau(g, t)$ estimate by constructing better counterfactuals within each group–time cell. They do not, by themselves, eliminate the aggregation pitfalls discussed in Chapter 4.

The design-first philosophy remains central, but it needs reinterpretation in hybrid settings. Pure synthetic control embodies the idea that identification lives in the assignment process and the comparability of donors, not in a fitted regression. Hybrid methods introduce modelling, which creates dependence on additional stability and specification assumptions. The resolution is not to abandon design thinking but to use it as discipline. You choose models that respect the structure of the problem, validate them through diagnostics, and report sensitivity to specification. The goal is transparency about what you are assuming, not purity about what you are not.

What follows develops these ideas systematically. We begin with augmented synthetic control in Section 7.2, which addresses bias from poor pre-treatment fit by combining synthetic control weighting with outcome regression. We then examine regularised and balancing approaches in Section 7.3, which stabilise weights and improve robustness to overfitting. Synthetic difference-in-differences receives extended treatment in Sections 7.4–7.10, showing how unit and time weights reshape the comparison that underpins a parallel-trends-style argument. Finally, we briefly survey matrix completion and other machine learning hybrids that relax linearity assumptions further.

The core message is pragmatic. Hybrid methods are not a replacement for design. They are an extension, a way to make design-disciplined inference work in settings where the data would otherwise defeat it. Use them when pure methods fail. Understand what you gain and what you give up. Report your choices transparently. That is how credible evidence gets built.

7.2 Augmented Synthetic Control (ASCM)

References

  • Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 7.1.