Synthetic Control Method: Constructing a Weighted Counterfactual for a Single Treated Unit

The synthetic control method builds a counterfactual for a single treated unit — one state, one country, one hospital system, one firm — by combining several untreated units into a weighted composite that tracks the treated unit’s pre-intervention trend closely. Where difference-in-differences needs a comparison group whose trend you’re willing to assume would have stayed parallel, synthetic control replaces that assumption with a data-driven weighting scheme, chosen so the composite actually matches the treated unit’s own pre-treatment path. It was introduced by Alberto Abadie and coauthors, first in a 2003 study of the economic cost of the Basque Country conflict, then formalized and popularized through two widely cited applications: California’s Proposition 99 tobacco-control program (Abadie, Diamond and Hainmueller, comparing California’s per-capita cigarette sales to a synthetic California built from other states) and German reunification (comparing West Germany’s per-capita GDP to a synthetic West Germany built from other OECD economies). Both papers are the standard citations for this method and worth reading directly before applying it to a new setting.

The problem it solves: N = 1, no obvious comparison unit

Difference-in-differences works well when a plausible comparison group exists — a set of units you can defend as likely to have moved in parallel with the treated unit absent treatment, often averaged together to reduce noise. Synthetic control is built for the harder case: exactly one unit was treated, and there is no single untreated unit that resembles it closely enough to serve as a comparison on its own. California in 1988 is not well matched by any single other state; West Germany in 1990 is not well matched by any single other OECD economy. Averaging every available untreated unit with equal weight doesn’t fix this — most of those units are poor matches, and an unweighted average dilutes the ones that are good matches with the ones that aren’t.

The synthetic control method’s contribution is to let the data choose which untreated units to combine, and in what proportion, so the resulting composite reproduces the treated unit’s pre-treatment characteristics as closely as possible. The weights are not chosen to make the post-treatment result look a particular way — they are fixed using only pre-treatment data, before the outcome the study is trying to estimate is even in view.

Constructing the synthetic control: donor pool and weights

The donor pool is the set of untreated units eligible to be combined into the synthetic control. Three criteria govern who belongs in it:

• No spillover or anticipation effects. A donor unit that was itself affected by the treatment — directly, through policy diffusion, or through anticipatory behavior before the treatment date — contaminates the counterfactual. If neighboring states adjusted their own tobacco policy in response to California’s, or partially anticipated German reunification’s economic effects, they are compromised donors even if never formally “treated.”
• Comparable data-generating process. Donor units should plausibly share the same broad economic, institutional, or demographic structure as the treated unit. Including units so structurally different that no convex combination of them could ever resemble the treated unit just wastes degrees of freedom and can produce misleadingly small pre-treatment fit statistics that don’t generalize.
• A sufficiently long, complete pre-treatment panel. The pre-treatment period is what the weights are fit on — a short or gappy pre-period gives the optimization too little information to distinguish a genuinely good match from an accidental one.

The weights w assigned to each donor unit are typically constrained to be non-negative and to sum to 1 — a convex combination. This constraint is deliberate, not incidental: it rules out extrapolation. A regression-based comparison could assign a donor unit a weight above 1 or below 0 to force a better pre-treatment fit, but that means projecting outside the range of the observed donor data, which is exactly the kind of unsupported extrapolation the method is designed to avoid. The weights are chosen to minimize the distance between the treated unit’s pre-treatment predictors (and, in most implementations, its pre-treatment outcome trajectory itself) and the weighted average of the same variables across donor units.

Judging pre-treatment fit before trusting the post-treatment gap

A synthetic control is only as credible as its pre-treatment fit. Before looking at the post-treatment period at all, check:

• Visual fit. Plot the treated unit’s actual outcome against the synthetic control’s outcome for the entire pre-treatment period. The two lines should track closely — not just at the endpoints, but through the whole series.
• Root mean squared prediction error (RMSPE) over the pre-treatment period, as a single summary number. A large pre-treatment RMSPE means the synthetic control never actually resembled the treated unit, and no inference drawn from the post-treatment gap is trustworthy — the study should stop here, not proceed to interpret a post-treatment divergence as a treatment effect.
• Predictor balance. The table comparing the treated unit’s pre-treatment predictor values to the synthetic control’s should show them close together, not just the outcome trend. A synthetic control that matches the outcome path by coincidence while diverging sharply on the predictors that are supposed to explain it is a warning sign, not a success.

A common mistake is including too many predictors relative to the length of the pre-treatment panel, which lets the optimizer find weights that fit the pre-period almost perfectly by chasing noise rather than signal — a form of overfitting that produces an excellent-looking pre-treatment plot and an unreliable post-treatment estimate. Favor a small number of predictors with real theoretical justification over an exhaustive list.

Validating the estimate: placebo tests instead of a standard error

With exactly one treated unit, the usual sampling-based logic behind a standard error and a p-value doesn’t apply in the conventional way — there’s no repeated-sampling variation across treated units to estimate. Synthetic control substitutes placebo tests, sometimes called permutation or randomization inference, to judge whether the estimated gap is unusually large or could plausibly have arisen by chance.

• Placebo-in-space. Apply the identical synthetic control procedure to every donor unit in turn, treating each as if it, rather than the actual treated unit, had received the treatment. This produces a distribution of “placebo effects” — gaps between each untreated unit and its own synthetic control. If the real treated unit’s gap is not clearly larger than most of these placebo gaps, the result is not distinguishable from noise, regardless of how large it looks in absolute terms. Donor units whose own pre-treatment fit was poor are typically excluded from this placebo distribution (a bad synthetic control produces a large, uninformative placebo gap by construction) or the comparison is restricted to placebo runs with pre-treatment fit at least as good as the real unit’s.
• Placebo-in-time. Re-run the procedure with the treatment date moved earlier, into a period when no actual intervention occurred. A “significant” gap appearing at this fake treatment date suggests the method is picking up something other than a genuine treatment effect — pre-existing divergence, an unrelated shock, or an artifact of the fitting procedure.

Reporting a synthetic control result without at least a placebo-in-space test is the single most common way this method is used incorrectly — the pre/post gap on its own says nothing about whether that gap is unusual.

Synthetic control versus difference-in-differences

Both are quasi-experimental tools for a single (or small number of) treated unit and untreated comparators, and they solve overlapping but distinct problems:

• Use difference-in-differences when a comparison group’s parallel-trends assumption is defensible on its own terms — typically several comparison units, or a large group, where you’re willing to argue absent treatment they’d have moved together with the treated unit. See evidencing parallel trends and the staggered-adoption problem for what that assumption actually requires and where the standard two-way fixed-effects estimator breaks under staggered treatment timing.
• Use synthetic control when there is exactly one treated unit and no single comparator (or defensible unweighted group) resembles it closely enough pre-treatment — the data-driven weighting is doing the work that an assumed parallel trend would otherwise have to do, and it comes with an explicit, checkable pre-treatment fit diagnostic that DiD’s parallel-trends assumption does not.
• Synthetic control generally needs a longer pre-treatment panel than DiD, because the weights are estimated from that period alone — a short pre-period leaves too little information to distinguish a genuine match from an accidental one.
• Both methods are vulnerable to a treated unit that is genuinely unlike anything available in comparison data; synthetic control makes that vulnerability visible (a poor pre-treatment fit), where an unweighted DiD comparison group can obscure it inside an averaged trend.

For a broader map of how these methods relate to other approaches to isolating a causal estimate from observational data, see causal analysis: a guide to causal inference and, for a design that addresses selection on observed characteristics rather than trends over time, propensity score matching: how it works and what it cannot fix. For the more general vocabulary distinguishing this whole family of designs from a randomized experiment, see experimental vs. quasi-experimental design.

A different “synthetic control”: don’t confuse this with clinical-trial external control arms

The term “synthetic control” is also used in clinical trial methodology for a different technique: an external or synthetic control arm built from historical trial data, registries, or real-world data to substitute for (or supplement) a concurrently randomized control group, most often in single-arm oncology or rare-disease trials. That is a real, distinct methodology with its own regulatory literature — see digital twins in clinical trials: synthetic control arms and FDA’s evolving posture for that topic. It shares a name and a general “build a comparison group without randomizing one” logic with the method on this page, but the mechanics are different: the clinical-trial version substitutes patient-level historical data for a concurrent randomized arm, while the method described here constructs a weighted composite of untreated aggregate units (states, countries, firms) to estimate the effect of an intervention on the one unit that received it. If your search brought you here looking for clinical-trial external control arms, that guide is the right page.

Limitations and extensions

• Interpolation bias. If no combination of donor units can plausibly resemble the treated unit, the method will still produce weights and a pre-treatment fit statistic — a poor one, if the diagnostics above are actually checked, but a researcher who skips them can be misled into trusting an estimate built from a bad match.
• Sensitivity to donor pool composition. Adding or removing a donor unit can change the weights and, sometimes, the substantive conclusion. Reporting which donors were considered and why, and checking robustness to reasonable alternative donor pools, is standard practice, not an optional extra.
• No formal p-value in the classical sense. Placebo-based inference is the accepted substitute, but it is not the same guarantee a standard error from a random sample provides, and should be described accordingly rather than reported as if it were a conventional significance test.
• Single (or few) treated units. The original method is built for one treated unit at a time. For settings with multiple treated units and more general panel structures, look at the extensions that grew out of it: the generalized synthetic control method (which combines the approach with an interactive fixed-effects model) and synthetic difference-in-differences (which blends synthetic-control-style weighting with the DiD two-way fixed-effects framework). An augmented synthetic control method, using ridge regression to correct residual pre-treatment imbalance, addresses cases where no convex combination achieves a good fit outright.

What to report

• The full donor pool considered, and the criteria used to include or exclude candidate units.
• The predictors used to construct the weights, and the justification for that specific set.
• The pre-treatment fit: the visual comparison, the RMSPE, and the predictor-balance table.
• The post-treatment gap, with the placebo-in-space distribution (and placebo-in-time result, where the pre-treatment panel is long enough to support it) shown alongside it — not the gap alone.
• A robustness check varying the donor pool or predictor set, showing the conclusion doesn’t hinge on one specific specification.

Frequently asked questions

How is the synthetic control different from just averaging several comparison units?

An unweighted average treats every donor unit as an equally good match, which is rarely true. Synthetic control lets the pre-treatment data determine the weights, so donor units that actually resemble the treated unit’s pre-treatment path contribute more than units that don’t — and the fit is a checkable diagnostic rather than an assumption.

Can weights be negative, to improve the pre-treatment fit further?

Not in the standard formulation. Constraining weights to be non-negative and sum to 1 (a convex combination) is what prevents the method from extrapolating outside the range of the observed donor data — allowing negative weights would let a poor-fitting donor pool appear to fit well through unsupported extrapolation rather than genuine resemblance.

What sample size or panel length does this method need?

There’s no fixed minimum, but the pre-treatment period needs to be long enough, and the donor pool rich enough, to distinguish a real match from an accidental one — both of the original applications (California tobacco control, German reunification) use pre-treatment panels running well over a decade.

Does a large post-treatment gap always mean the treatment worked?

No. A large gap only supports a causal claim once pre-treatment fit is confirmed to be good and the placebo tests show the gap is unusual relative to what the same procedure produces on untreated units. A large gap built on a poor pre-treatment fit is evidence the synthetic control was never a valid counterfactual, not evidence of a treatment effect.

Is this the same thing as a “synthetic control arm” in a clinical trial?

No — related name, different method and literature. See the section above distinguishing this aggregate-unit weighting method from clinical-trial external/synthetic control arms built from patient-level historical data.

Sources

• Abadie, A. and Gardeazabal, J. (2003). “The Economic Costs of Conflict: A Case Study of the Basque Country” — the original synthetic control application.
• Abadie, A., Diamond, A., and Hainmueller, J. (2010). “Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California’s Tobacco Control Program” — the California Proposition 99 study and the formalization of the donor-pool weighting procedure.
• Abadie, A., Diamond, A., and Hainmueller, J. (2015). “Comparative Politics and the Synthetic Control Method” — the German reunification application and a fuller treatment of the method for a social-science audience.