Intercept-Shifted Synthetic Controls

• Intercept-shifted synthetic controls are a method that introduces unit-specific additive shifts to account for systematic pre-treatment level differences between treated and control units.
• The method de-means donor series using pre-treatment means, focusing the synthetic control construction on matching dynamic patterns for improved bias reduction.
• It has been shown to improve bias and efficiency over standard synthetic controls and DID, especially in settings with staggered adoption and heterogeneous baselines.

Intercept-shifted synthetic controls generalize the canonical Synthetic Control Method (SCM) by introducing unit-specific additive shifts (“intercepts”) to explicitly account for persistent level differences between treated and control units. This approach absorbs non-dynamic, cross-sectional mean biases—gaps in pre-treatment levels that cannot be bridged by convex combinations of donors—thereby allowing the synthetic control construction to focus on matching residual dynamics and improving both bias and efficiency in causal effect estimation. The method is now standard in contexts with staggered or imperfect adoption, heterogeneous unit baselines, or structural model misspecification in the mean levels.

1. Formal Foundations and Identification

The formalized setting is a panel: units $i=1,\dots,N$ observed over $t=1,\dots,T$. For each unit $i$, let $Y_{it}(s)$ denote the potential outcome at time $t$ if treatment had occurred at time $s$; let $T_i$ be the adoption time ($T_i=\infty$ if unexposed), and denote $D_{it} = \mathbf{1}\{t\geq T_i\}$. The observable is $Y_{it}=D_{it}Y_{it}(T_i)+(1-D_{it})Y_{it}(\infty)$. In many designs (especially with staggered adoption or unbalanced panels), the untreated potential outcomes may admit an unknown mean-level $t=1,\dots,T$0 and possibly time-varying trends/spillovers.

The distinctive identifying assumption for intercept-shifted SCM is the existence, for each treated unit $t=1,\dots,T$1, of nonnegative donor weights $t=1,\dots,T$2 summing to one, such that the pre-treatment (demeaned) outcome trajectory of the treated is replicated by the weighted average of donor trajectories up to, but not including, the mean-level:

$t=1,\dots,T$3

with $t=1,\dots,T$4 denoting the pre-treatment mean of unit $t=1,\dots,T$5. This “weighted parallel trends with intercept shift” is a weaker and more testable condition than AR($t=1,\dots,T$6) or latent factor structural assumptions and underpins guarantees of unbiasedness and consistency given appropriate weight regularity and sample growth (Guggisberg, 7 Aug 2025).

2. Construction of the Intercept-Shifted Synthetic Control

For treated unit $t=1,\dots,T$7 and donor pool $t=1,\dots,T$8, the intercept-shifted counterfactual is constructed as follows:

• Let $t=1,\dots,T$9 be nonnegative donor weights subject to
  • $i$0, $i$1, and $i$2 if $i$3.
• Estimate a unit-specific intercept $i$4 and donor intercepts $i$5.

The synthetic control estimator predicts the untreated outcome for treated unit $i$6 at post-adoption time $i$7 as

$i$8

This de-means the donor series over the pre-treatment period and re-centers at the treated's mean, absorbing level mismatches and focusing donor fit on dynamic pattern resemblance (Ben-Michael et al., 2019).

The closed-form $i$9 (given fixed weights), for $Y_{it}(s)$0 pre-treatment periods, is

$Y_{it}(s)$1

This yields an intercept-shifted synthetic matching the treated unit's pre-treatment mean (Ben-Michael et al., 2019, Ferman et al., 2019).

3. Optimization and Algorithmic Properties

The estimator solves a quadratic program over $Y_{it}(s)$2—the vector of unit intercepts and the donor-weight matrix—minimizing a convex combination of pre-treatment root mean squared errors (RMSE):

• Separate fit: $Y_{it}(s)$3 — average unit-specific RMSE.
• Pooled fit: $Y_{it}(s)$4 — RMSE for the pooled average treated path.

The objective, with normalization $Y_{it}(s)$5 and a small ridge penalty $Y_{it}(s)$6, is

$Y_{it}(s)$7

subject to the donor weight constraints.

Hyperparameters:

• $Y_{it}(s)$8 governs the trade-off between pooled and separate fit; data-driven heuristics and Pareto diagnostics are recommended.
• $Y_{it}(s)$9 is usually set to a small value for strict convexity and regularization. Estimation typically uses quadratic programming solvers (e.g., quadprog, MOSEK), with $t$0 eliminated by substitution or treated as an explicit parameter (Ben-Michael et al., 2019).

Extensions include auxiliary covariates in the balancing objective, cohort-level pooling for shared adoption dates, and bootstrap or jackknife inference (Ben-Michael et al., 2019).

4. Theoretical Properties: Bias, Variance, and Consistency

Intercept shifts eliminate bias from fixed additive level differences, as the pre-treatment means are matched by construction. Under AR($t$1) or factor models, error bounds for the average treatment effect on the treated (ATT) explicitly decompose into dynamic (pooled and individual) imbalance and a remainder term due to noise and approximation error:

$t$2

Replacing outcomes with de-meaned series $t$3 removes systematic level error, so ATT error depends only on the dynamic components (Ben-Michael et al., 2019).

Recent results (Guggisberg, 7 Aug 2025) show that under the weighted parallel trends plus intercept shift assumption, regular weights, and growing pre-treatment windows, the intercept-shifted SCM estimator is consistent for the time- and unit-average ATT, even allowing for heavier-tailed shocks than sub-Gaussian. Demeaning reduces bias and variance over standard SCM and typically outperforms difference-in-differences (DID) unless time-varying confounding is present (Ferman et al., 2019).

5. Bayesian and Penalized-Likelihood Approaches

A Bayesian approach interprets the intercept shift as a free parameter, casting the optimization as maximum a posteriori (MAP) estimation under pseudo-Gaussian likelihood subject to parallelly shiftable convex hull constraints. The MAP solution is:

$t$4

subject to $t$5 for $t$6 and $t$7, with $t$8 free ($t$9 is the intercept).

The KKT conditions admit a closed-form for $s$0 (the average residual with given weights), and the Gibbs sampling scheme enables credible interval estimation and covariate selection (Goh et al., 2020).

In empirical work (Basque Country application), the Bayesian intercept-shifted SCM yields coverage-accurate posterior intervals and selects donor pools and intercepts that closely match pre- and post-treatment dynamics; the intercept directly absorbs mean-level disparities (Goh et al., 2020).

6. Practical Diagnostics and Implementation Guidance

Empirical diagnostics for intercept-shifted SCM focus on validating design-based assumptions:

• Pre-treatment gap plots: Low RMSE in placebo periods evidences weighted parallel trends with intercept.
• Effective donor count: $s$1, with $s$2 recommended to limit variance inflation.
• Sensitivity to donor set and window length: Monotonic RMSE reduction with increased pre-treatment periods or donors signals convex hull adequacy.
• Specification test: Permutation/block-based tests of the de-meaned estimator against DID can detect violation of the parallel trends assumption (Ferman et al., 2019).
• Reporting: Present results for standard SCM, intercept-shifted SCM, and DID; proximity across estimators (with large $s$3-values) supports parallel trends and bias reduction, while divergence suggests misspecification.

Algorithmic steps for the single treated unit case include:

1. Compute pre-treatment means for treated and donor units.
2. Solve the convex QP for donor weights on the demeaned outcomes.
3. Estimate the effect by subtracting the re-centered synthetic prediction in the post-treatment period.
4. Optionally, conduct specification testing via permutation.

7. Extensions and Comparison with Hybrids

Intercept-shifted SCM generalizes both classical SCM (no intercept) and DID (uniform weights with intercept) (Ferman et al., 2019). It is closely related to recent “SC+DiD hybrid” methods, such as the “partially pooled SCM with intercepts” and the “synthetic difference-in-differences” estimator, but differs by explicitly absorbing mean-level differences while allowing flexible weighting structures (Ben-Michael et al., 2019, Guggisberg, 7 Aug 2025).

Compared to factor- or AR-model-based SCM, intercept-shifted methods weaken structural assumptions by requiring only observable parallel trends in dynamics after demeaning, and impose only design-based regularity on weight distributions and error tails.

Empirical simulations and applied work demonstrate improved bias and variance for ATT over standard SCM and DID, both in correctly specified and certain misspecified/incomplete pre-treatment fit settings (Ben-Michael et al., 2019, Goh et al., 2020, Ferman et al., 2019).

---

In summary, intercept-shifted synthetic controls provide a robust, flexible framework for causal inference in panels where treated and donor units may differ systematically in mean levels. By incorporating unit-specific intercepts, these estimators relax convex hull constraints, absorb persistent bias, and anchor identification to explicit and empirically testable design assumptions, thereby expanding the methodology's applicability and empirical reliability (Ben-Michael et al., 2019, Goh et al., 2020, Ferman et al., 2019, Guggisberg, 7 Aug 2025).