Synthetic Difference-in-Differences

Synthetic difference-in-differences (abbreviated SDID) is a quasi-experimental panel-data causal inference estimator in biostatistics, epidemiology, and econometrics developed by Dmitry Arkhangelsky, Susan Athey, David A. Hirshberg, Guido W. Imbens, and Stefan Wager in 2021 (American Economic Review) to evaluate the average treatment effect on the treated (ATT) of population-level interventions—such as state tobacco taxes, Medicaid expansions, sugar-sweetened beverage excise taxes, environmental emissions standards, or hospital vaccination mandates—across longitudinal panel datasets. SDID unifies the two primary quasi-experimental frameworks of modern public health policy evaluation: traditional two-way fixed effects difference-in-differences (DID) and the Abadie–Diamond–Hainmueller synthetic control method (SC). Specifically, SDID estimates the causal treatment effect τ by solving a doubly weighted two-way fixed effects regression that simultaneously incorporates unit weights ω̂_i (which reweight untreated control jurisdictions so their pre-intervention outcome trend parallels the treated group) and time weights λ̂_t (which reweight pre-intervention calendar periods to prioritize baseline periods most predictive of post-intervention control outcomes), while retaining additive unit fixed effects α_i and time fixed effects β_t. By combining data-driven local reweighting with additive fixed-effects differencing, SDID achieves reduced bias, lower estimation variance, and formal asymptotic normality under interactive latent factor data-generating processes where either standard DID or standard SC alone would fail.

Limitations of Unweighted Difference-in-Differences and Pure Synthetic Control

Consider a balanced panel dataset of N aggregate units (such as states, counties, or health districts, indexed by i = 1, …, N) observed across T calendar periods (t = 1, …, T). Partition the N units into N_co untreated control units (i = 1, …, N_co) that never adopt the intervention and N_tr = N − N_co treated units (i = N_co + 1, …, N) that adopt the intervention simultaneously beginning at period T_pre + 1 (creating T_pre pre-intervention periods and T_post = T − T_pre post-intervention periods). Let Y_it denote the observed health or economic outcome and W_it ∈ {0, 1} denote the binary policy exposure indicator (W_it = 1 if and only if i > N_co and t > T_pre).

Both classical difference-in-differences and the synthetic control method can be written as weighted linear least-squares regressions, yet each imposes restrictive structural assumptions:

1. Standard Two-Way Fixed Effects DID: Estimates the treatment effect τ̂^did by solving the unweighted regression:

Because DID assigns uniform weight 1/N_co to every control unit and uniform weight 1/T_pre to every pre-treatment period, it is valid only when the unweighted average of all control units satisfies the parallel trends assumption across the pre- and post-treatment window. In public health applications where treated states have distinct demographic trajectories or secular mortality trends, unweighted parallel trends are routinely violated.

1. Pure Synthetic Control (SC): Estimates τ̂^sc by constructing simplex unit weights ω̂_i^sc ≥ 0 (∑_i=1^N_co ω̂_i^sc = 1) that match the treated units' pre-intervention outcome levels, and solving a unit-weighted regression without unit fixed effects α_i and without time weights λ_t:

Because pure SC omits unit fixed effects α_i, it forces the convex combination of donor units to match the unshifted level of the treated unit rather than just its slope (failing whenever the treated jurisdiction lies outside the convex hull of donor levels), and it places equal weight on distant historical pre-treatment periods (t = 1) and periods directly preceding policy enactment (t = T_pre).

The Doubly Weighted SDID Objective Function

Arkhangelsky et al. (2021) resolved both limitations by defining the Synthetic Difference-in-Differences estimator τ̂^sdid as the solution to the doubly weighted two-way fixed effects least-squares problem:

Here:

• Unit Weights ω̂_i^sdid: Non-negative weights on the N_co control units satisfying ∑_i=1^N_co ω̂_i^sdid = 1 (with treated units assigned uniform weight ω̂_i^sdid = 1/N_tr for i > N_co), chosen so that the ω̂-weighted pre-treatment trajectory of the control group is parallel to the pre-treatment trajectory of the treated group up to a constant vertical intercept shift.
• Time Weights λ̂_t^sdid: Non-negative weights on the T_pre pre-intervention periods satisfying ∑_t=1^T_pre λ̂_t^sdid = 1 (with post-intervention periods assigned uniform weight λ̂_t^sdid = 1/T_post for t > T_pre), chosen so that the λ̂-weighted average of pre-treatment outcomes for each control unit predicts its average post-treatment outcome up to a constant additive shift.

Because the SDID objective includes both unit fixed effects α_i and time fixed effects β_t, the estimator has a transparent closed-form representation as a double difference of weighted averages:

where Y_{tr, post} is the simple average of treated units in the post-treatment period, Y_{tr, t} is the average of treated units at pre-treatment time t, and Y_{i, post} is the average of control unit i across the post-treatment period.

L2-Regularized Simplex Program for Unit Weights ω

A central innovation of SDID over classical synthetic control is how the unit weights ω̂^sdid ∈ Δ^N_co (where Δ^N_co = {ω ∈ ℝ₊^N_co : ∑_i=1^N_co ω_i = 1} is the unit simplex) are estimated from the pre-treatment data:

Two mathematical components distinguish this optimization from Abadie–Diamond–Hainmueller SC:

1. Free Intercept ω₀ ∈ ℝ (Slope Alignment Rather Than Level Alignment): Because the downstream SDID regression includes unit fixed effects α_i that subtract out each unit's baseline level, the unit-weight optimization includes an unconstrained scalar intercept ω₀. Consequently, the donor weights ω_i do not need to reproduce the unshifted pre-treatment level of the treated group; they only need to reproduce its time-varying fluctuations (pre-trend) up to an additive constant ω₀. Even if the treated jurisdiction has the highest or lowest baseline mortality rate in the country (lying outside the convex hull of donor levels), SDID constructs a valid parallel-trend counterfactual without negative-weight extrapolation.
2. Noise-Scaled L₂ Ridge Penalty (ζ² T_pre ||ω||₂²): Pure unpenalized simplex matching often assigns weight to only a few donor units, which inflates variance when period-specific idiosyncratic noise ε_it is present. SDID adds a ridge penalty scaled to the empirical variance of first-differenced noise among the control units:

and ΔY_it = Y_i,t+1 − Y_it. This L₂ regularization disperses weights across a broader set of similar donor units to bound the variance norm ||ω̂^sdid||₂ while preserving simplex non-negativity.

Simplex Program for Pre-Treatment Time Weights λ

Symmetrically, the pre-treatment time weights λ̂^sdid ∈ Δ^T_pre (where ∑_t=1^T_pre λ_t = 1 and λ_t ≥ 0) are chosen by matching the pre-treatment periods to the post-treatment average across all N_co untreated control units:

where a small regularization parameter ζ_λ = 10⁻⁶ σ̂ prevents degeneracy when multiple pre-treatment periods are collinear.

In intuitive epidemiological terms, if disease rates or health expenditures follow a persistent autoregressive or random-walk process, distant historical periods (t ≪ T_pre) carry substantial drift relative to the post-policy era, whereas the periods directly preceding policy implementation (t ≈ T_pre) strongly predict post-policy counterfactual outcomes. The time-weight optimization concentrates λ̂_t^sdid on the most predictive pre-intervention periods, removing systematic time-series drift that would otherwise bias an unweighted pre–post comparison.

Low-Rank Matrix Decomposition of Potential Outcomes

To establish consistency and asymptotic distribution theory, Arkhangelsky et al. (2021) analyzed SDID under a general interactive fixed effects (latent factor) model for the untreated potential outcomes Y_it(0):

Here L = [L_it] is a structured low-rank signal matrix combining additive unit effects α_i, time effects β_t, and an R-dimensional interactive factor component γ_i^T υ_t (capturing unobserved unit-specific sensitivities γ_i to time-varying macro-epidemiological shocks υ_t), while ε_it represents mean-zero idiosyncratic shocks across units and periods.

Decomposing the N × T signal matrix L into block form separating control vs. treated units (co vs. tr) and pre-treatment vs. post-treatment periods (pre vs. post), the bias of the SDID estimator conditional on the weights (ω̂, λ̂) is governed by the bilinear residual:

Factoring the interactive structure L_it − α_i − β_t = γ_i^T υ_t reveals that this residual is the inner product of the unit-factor balance error and the time-factor balance error:

This factorization yields a double-robustness property analogous to targeted maximum likelihood estimation and augmented inverse-probability weighting in observational cohort studies:

1. Whenever the unit weights ω̂ balance the unobserved unit factor loadings (γ_tr ≈ ∑ ω̂_i γ_i), and
2. The time weights λ̂ balance the unobserved time factors (υ_post ≈ ∑ λ̂_t υ_t),

the resulting estimator bias is the product of the two small approximation errors, converging to zero at a faster rate than either a unit-weighted synthetic control estimator or a time-weighted regression estimator alone.

Clustered Bootstrap, Jackknife, and Placebo Inference

Arkhangelsky et al. (2021) proved that when the panel dimensions (N_co, T_pre, N_tr, T_post) grow under regularity conditions on the rank of L and error serial correlation, the standardized SDID estimator is asymptotically normal:

Because the unit and time weights (ω̂^sdid, λ̂^sdid) depend on the data through constrained quadratic programs, three complementary variance estimators V̂_τ are used in empirical health-policy research depending on the number of treated units N_tr:

1. Clustered Unit Bootstrap (N_tr Large): Resamples unit trajectories i ∈ {1, …, N} with replacement B times, recomputing (ω̂^(b), λ̂^(b), τ̂^(b)) on each bootstrap panel to construct normal-theory confidence intervals τ̂^sdid ± z_1−α/2 V̂_boot^1/2.
2. Fixed-Weights Leave-One-Unit-Out Jackknife (N_tr ≥ 2): Holds the optimal weights (ω̂^sdid, λ̂^sdid) fixed and computes the leave-one-unit-out estimates τ̂⁽⁻ⁱ⁾ across i = 1, …, N. Because it avoids re-solving the quadratic programs N times, the jackknife is fast and asymptotically conservative.
3. Placebo Variance Estimator (N_tr = 1 Single Treated Jurisdiction): When only a single state or country adopts the policy (N_tr = 1, as in California's Proposition 99), the jackknife is undefined because dropping the sole treated unit leaves no treated observations. In this regime, SDID randomly samples N_tr = 1 pseudo-treated unit from the N_co untreated donor units without replacement, applies the full SDID estimator to the remaining N_co − 1 controls to obtain placebo treatment effects τ̂_placebo^(b), and uses the empirical variance of τ̂_placebo^(b) to construct valid Gaussian confidence intervals.

Staggered Adoption and Covariate Adjustment Extensions

In many epidemiological surveillance settings, jurisdictions adopt a health policy at staggered calendar dates G_i ∈ {t₁, t₂, …, t_K} rather than simultaneously. Following the staggered-DID framework of Callaway and Sant'Anna (2021), staggered SDID partitions the treated units into adoption cohorts g defined by their initial treatment date, computes a cohort-specific SDID estimate τ̂_g^sdid using the never-treated (or not-yet-treated) jurisdictions as the donor pool over the pre-period t < g, and aggregates the cohort estimates weighted by treated unit-periods. Time-varying covariates X_it can likewise be incorporated by residualizing Y_it against X_it β jointly within the doubly weighted objective or prior to weight estimation.

When should researchers prefer SDID over standard Synthetic Control or DID?

Researchers should prefer SDID whenever evaluating aggregate panel interventions where (1) unweighted parallel trends across all untreated units are questionable (making standard DID biased), or (2) the treated jurisdiction has an outlier baseline outcome level outside the convex hull of donor states, or multiple treated units (N_tr ≥ 1) and noisy pre-treatment periods are present (where standard SC suffers from interpolation bias or elevated variance). Because SDID reduces to DID when weights are uniform and behaves like level-shifted SC when pre-trend matching dominates, it serves as a doubly robust default across both regimes.

Why does SDID regularize the unit weights ω with an L2 penalty but scarcely regularize the time weights λ?

In typical policy panels, unit-specific idiosyncratic shocks ε_it have substantial cross-sectional variance, so allowing ω̂ to concentrate on a single donor state would transmit that donor's post-treatment noise directly into the treatment effect estimate (inflating variance proportional to ||ω̂||₂²). Conversely, the pre-treatment time weights λ̂ are averaged across all N_co control units, and post-treatment shocks are frequently strongly correlated with the periods directly preceding treatment (t ≈ T_pre). Keeping the penalty on λ̂ negligible (ζ_λ = 10⁻⁶ σ̂) allows λ̂ to concentrate on the final pre-treatment periods just before the intervention, which minimizes bias from secular drift.

How can I visualize the validity of an SDID design in a public health study?

A standard SDID diagnostic plot displays two panels: (1) the parallel-trends trajectory plot, comparing the observed outcome path of the treated group against the ω̂^sdid-weighted synthetic control trajectory (shifted vertically by the estimated intercept difference) across all pre- and post-treatment periods, with shaded triangles at the bottom of the pre-period indicating the magnitude of the time weights λ̂_t^sdid; and (2) the unit-by-unit difference contribution dot plot, plotting each donor state's individual DID contrast against its assigned weight ω̂_i^sdid.

1. Arkhangelsky, D., Athey, S., Hirshberg, D. A., Imbens, G. W., & Wager, S. (2021). Synthetic difference-in-differences. American Economic Review, 111(12), 4088–4118. https://doi.org/10.1257/aer.20190159
2. Abadie, A., Diamond, A., & Hainmueller, J. (2010). Synthetic control methods for comparative case studies: Estimating the effect of California's tobacco control program. Journal of the American Statistical Association, 105(490), 493–505. https://doi.org/10.1198/jasa.2009.ap08746
3. Athey, S., Bayati, M., Doudchenko, N., Imbens, G., & Khosravi, K. (2021). Matrix completion methods for causal panel data models. Journal of the American Statistical Association, 116(536), 1716–1730. https://doi.org/10.1080/01621459.2021.1891924
4. Ben-Michael, E., Feller, A., & Rothstein, J. (2021). The augmented synthetic control method. Journal of the American Statistical Association, 116(536), 1789–1803. https://doi.org/10.1080/01621459.2021.1929245
5. Callaway, B., & Sant'Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200–230. https://doi.org/10.1016/j.jeconom.2020.12.001